<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Angelska on Medium]]></title>
        <description><![CDATA[Stories by Angelska on Medium]]></description>
        <link>https://medium.com/@angelska327?source=rss-3a644af46fc3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*nvLk1V68NMcw3m3IPLzx6g@2x.jpeg</url>
            <title>Stories by Angelska on Medium</title>
            <link>https://medium.com/@angelska327?source=rss-3a644af46fc3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 13:48:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@angelska327/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Future of Financial Cybersecurity]]></title>
            <link>https://medium.com/@angelska327/the-future-of-financial-cybersecurity-62fe150df267?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/62fe150df267</guid>
            <category><![CDATA[quantum-threat]]></category>
            <category><![CDATA[artex]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[pqc]]></category>
            <category><![CDATA[cryptographic-agility]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 10:46:43 GMT</pubDate>
            <atom:updated>2026-11-18T11:39:40.317Z</atom:updated>
            <content:encoded><![CDATA[<h3>How should financial institutions prepare for AI-driven cyberattacks and quantum computing threats?</h3><h3>Executive Summary</h3><p>Attackers increasingly use AI to automate reconnaissance, identify vulnerabilities and accelerate multistage attacks. Compromised identities, excessive permissions and exposed applications can give attackers access to payment systems, customer information and critical infrastructure.</p><p>This article examines ransomware, account takeover, offensive security, fraud detection, incident response and operational resilience. It includes MITRE’s Fight Fraud Framework (F3), the six-level Proactive Fraud Prevention Pyramid and 17 detection opportunities based on the FBI’s November 2026 ARTEX advisory.</p><p>Additional sections cover autonomous AI agent security, third-party dependencies and post-quantum cryptography (PQC), including NIST standards and financial-sector migration planning.</p><h3>Part 1. Financial Cyber Threats and Attack Progression</h3><h3>1.1 Ransomware and Initial Access</h3><p>Ransomware attacks frequently begin through phishing, stolen credentials, exposed remote access services or exploitable applications.</p><p>After initial access, attackers may obtain privileged credentials, move laterally, disable security tools and steal sensitive data. Some groups use data theft and extortion without deploying encryption.</p><p>For financial institutions, ransomware can interrupt payment processing, customer authentication, trading operations and access to financial records.</p><p><strong>Priority controls:</strong></p><ul><li>Require phishing-resistant MFA for privileged and remote access.</li><li>Restrict administrative permissions and segment critical networks.</li><li>Monitor suspicious remote administration, security tool tampering and unusual data transfers.</li><li>Maintain protected backups and test restoration procedures.</li><li>Revoke compromised credentials and investigate persistence before restoring services.</li></ul><h3>1.2 Application Vulnerabilities and Attack Paths</h3><p>Financial institutions operate web applications, APIs, cloud services, legacy systems and third-party integrations.</p><p>An exposed vulnerability becomes more dangerous when combined with compromised credentials, excessive permissions or weak network segmentation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/992/1*3k9qBaYxqoic05AlC9Tw9g.png" /><figcaption><strong>Figure 1. Attack chain path</strong>: Exposed application leading to service-account compromise, privilege escalation, internal network access, and payment infrastructure.</figcaption></figure><p>Vulnerability assessments should establish whether:</p><ol><li>The vulnerability is reachable and exploitable.</li><li>The compromised system can access sensitive services.</li><li>Additional weaknesses allow further movement.</li><li>Exploitation could enable unauthorized transactions, modification of payment instructions or account balances, or disruption of operations.</li><li>An attacker could obtain additional identities or permissions.</li></ol><p>Vulnerability prioritization should consider exploit probability (EPSS), severity (CVSS), system reachability and business criticality. Vulnerability scanning should also be integrated into CI/CD pipelines to identify weaknesses before deployment.</p><p>Remediation should eliminate the demonstrated attack path. Patching the initial vulnerability may be insufficient if stolen credentials or unauthorized access remain.</p><p>Security teams should revoke compromised credentials, remove unauthorized accounts and persistence mechanisms, restrict excessive permissions and verify through retesting that the attack path is no longer exploitable.</p><h3>1.3 Account Takeover (ATO) and Financial Fraud</h3><p>Account takeover occurs when an attacker gains unauthorized access to a customer or employee account.</p><p>A common sequence is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/684/1*IeZ1hU8ymD7eHFUZrDyWIw.png" /><figcaption><strong>Figure 2. Attack chain path:</strong> Credential compromise leading to successful login, account changes, new beneficiary addition, and unauthorized transfer.</figcaption></figure><p>A successful login alone may appear legitimate. Risk increases when it is followed by changes to recovery information, authentication settings or payment instructions.</p><p><strong>Detection and prevention:</strong></p><ul><li>Correlate authentication activity, MFA failures and unfamiliar devices.</li><li>Monitor password resets, account recovery events and beneficiary changes.</li><li>Flag unusual transaction amounts and payment destinations.</li><li>Require additional verification when suspicious account changes precede high-value transfers.</li></ul><h3>1.4 Social Engineering, Impersonation and Deepfake Fraud</h3><p>Attackers impersonate customers, executives, employees and suppliers to steal credentials or authorize fraudulent payments.</p><p>AI-generated messages, deepfake videos, synthetic voices and compromised email accounts can make fraudulent requests appear legitimate.</p><p><strong>Preventive measures:</strong></p><ul><li>Verify payment instructions through a trusted, separate communication channel.</li><li>Require two-person approval for high-value transfers and beneficiary changes.</li><li>Monitor unusual transaction amounts, destinations and account activity.</li><li>Require additional identity checks for suspicious account recovery requests.</li><li>Never rely on voice or video alone to approve financial transactions.</li><li>Train employees to recognize deepfakes and suspicious requests (e.g., through tabletop exercises).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/661/1*WkEYDU-FWn21E9kfO4iSfA.png" /><figcaption><strong>Figure 3.</strong> AI-generated deepfakes remain the most common form of synthetic media in late 2026, easily created from a single photo or video.</figcaption></figure><h3>1.5 Financial Fraud Methods</h3><p>Financial institutions face several types of fraud involving stolen identities, manipulated payments and unauthorized account activity.</p><ul><li><strong>Business Email Compromise (BEC):</strong> Attackers impersonate trusted business contacts to redirect payments.</li><li><strong>Synthetic Identity Fraud:</strong> Criminals combine real and fake personal information to create fraudulent identities.</li><li><strong>New Account Fraud:</strong> Criminals open accounts using stolen or fabricated identities.</li><li><strong>Authorized Push Payment (APP) Fraud:</strong> Victims are tricked into authorizing fraudulent transfers.</li><li><strong>Mule Account Fraud:</strong> Criminals use accounts to receive and transfer stolen funds.</li><li><strong>Card Fraud:</strong> Stolen card information is used for unauthorized purchases.</li><li><strong>Loan and Credit Fraud:</strong> Criminals use stolen identities or false information to obtain credit.</li><li><strong>Check Fraud:</strong> Criminals forge, alter or steal checks.</li></ul><p><strong>Detection and prevention:</strong> Monitor unusual account openings, identity verification failures, beneficiary changes, transaction patterns and rapid movement of funds between accounts. Apply additional verification to high-risk transactions and investigate accounts receiving suspicious transfers.</p><h3>Part 2. Attack Surface, Identity and Preventive Security</h3><h3>2.1 Asset Visibility and Shadow Technology</h3><p>Legacy applications, unmanaged SaaS platforms, partner portals and unauthorized AI services may operate outside established security monitoring.</p><p>A five-phase approach to shadow technology management includes:</p><ol><li><strong>Discover:</strong> Identify applications, services and integrations.</li><li><strong>Assess:</strong> Examine exposure, permissions and access to sensitive data.</li><li><strong>Govern:</strong> Assign ownership and establish security requirements.</li><li><strong>Enable:</strong> Provide approved alternatives where needed.</li><li><strong>Sustain:</strong> Monitor applications and integrations continuously.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/869/1*r64UVxsPq-ZDeiF444TPmw.png" /><figcaption><strong>Figure 4.</strong> Shadow technology management. Source: Synthesized from common shadow IT/AI management practices (Adaptive Security, Corma, BlackFog, Microsoft Defender for Cloud Apps guidance, 2026).</figcaption></figure><p>Discovery sources include SSO records, OAuth grants, cloud audit logs, network activity and procurement records.</p><p>Asset discovery and attack detection require different capabilities. Finding an unknown application does not establish whether suspicious activity against it can be detected.</p><p>An application without usable authentication, access or network logs creates a monitoring gap. Unsupported and end-of-life applications should be identified and removed or isolated when security patches are no longer available.</p><h3>2.2 Human and Non-Human Identity Security</h3><p>Financial institutions rely on employees, contractors, service accounts, workload identities and automated systems.</p><p>Non-human identities may operate continuously, access multiple systems and retain permissions long after their original purpose changes.</p><p>Controls should include:</p><ul><li>Defined ownership for each identity.</li><li>Least-privilege access.</li><li>Short-lived credentials where supported.</li><li>Secure storage and rotation of secrets.</li><li>Periodic permission reviews.</li><li>Monitoring of privileged role changes and credential creation.</li></ul><p>Security teams should investigate unexpected OAuth grants, new application credentials and service accounts accessing systems outside their normal responsibilities.</p><p>AI agents may use non-human identities, but their ability to select tools and initiate actions introduces additional authorization risks.</p><h3>2.3 Cloud Security and Network Segmentation</h3><p>Cloud security assessments should examine publicly exposed services, IAM permissions, storage access, network rules and audit logging.</p><p>Critical payment systems should be isolated from ordinary user environments and unnecessary administrative access.</p><p>Monitor for:</p><ul><li>Unexpected privileged role assignments.</li><li>New public IP addresses.</li><li>Unapproved virtual machine deployments.</li><li>Changes to storage permissions.</li><li>Administrative activity from unfamiliar networks.</li></ul><p>Correlate suspicious changes with the initiating identity, source location and approved change records.</p><h3>2.4 Mobile Applications, APIs and Business Logic</h3><p>Mobile banking applications rely on APIs connecting customer interfaces to financial services.</p><p>Broken authorization and insecure session management can expose customer accounts or permit unauthorized operations. Authorization must be enforced server-side for every sensitive request.</p><p>Penetration testing should examine whether legitimate application functions can be combined to bypass financial controls, including:</p><ul><li>Changing beneficiaries without appropriate authorization.</li><li>Manipulating transaction limits.</li><li>Reusing expired sessions.</li><li>Bypassing approval workflows.</li><li>Accessing another customer’s records.</li></ul><p>API monitoring should identify repeated authorization failures, rapid endpoint enumeration and unusual request sequences.</p><h3>Part 3. Fraud Prevention and Detection Engineering</h3><h3>3.1 The Proactive Fraud Prevention Pyramid</h3><p>The six-level Proactive Fraud Prevention Pyramid organizes fraud prevention capabilities from coordination and data collection to behavioral and predictive analysis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/1*Ja4sIJBj9U5QVr5hRrFV_g.png" /><figcaption><strong>Figure 5</strong>. Proactive Fraud Prevention Pyramid: Synthesized framework for proactive fraud prevention integrating cybersecurity and fraud operations (common in financial services risk management practices).</figcaption></figure><ol><li><strong>Cyber / Fraud Relationship:</strong> Establish shared responsibilities and coordination between cybersecurity and fraud teams.</li><li><strong>Consumer Journey:</strong> Map customer interactions across authentication, account management and financial transactions.</li><li><strong>Data:</strong> Collect and normalize identity, device, application and transaction records.</li><li><strong>Tactical:</strong> Detect known fraud indicators and suspicious activity patterns.</li><li><strong>Behavioral:</strong> Identify deviations from established customer behavior.</li><li><strong>Predictive:</strong> Apply machine learning, adversarial testing and gap analysis to identify emerging fraud risks.</li></ol><p>Each level builds on the capabilities established below it. Reliable data and coordination provide the foundation for effective behavioral analysis and predictive fraud detection.</p><h3>3.2 MITRE Fight Fraud Framework (F3)</h3><p>MITRE F3 organizes fraud-related attacker behavior into tactics and techniques. Detection engineers can use these behaviors to identify which events should be monitored.</p><p>Consider an account takeover followed by a fraudulent transfer.</p><p><strong>Attack sequence:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/674/1*tUyTJSKH2mJG8TY-RobGZg.png" /><figcaption><strong>Figure 2. </strong>Used earlier. Attack chain path: Credential compromise leading to successful login, account changes, new beneficiary addition, and unauthorized transfer.</figcaption></figure><p><strong>Required telemetry:</strong></p><ul><li>Authentication logs.</li><li>Device identifiers.</li><li>Account modification records.</li><li>Beneficiary changes.</li><li>Transaction records.</li></ul><p><strong>Example detection:</strong></p><p>Generate a high-priority alert when a newly authenticated device changes account recovery information and initiates a high-value payment to a newly added beneficiary within a short period.</p><p>The time window and transaction threshold must be based on legitimate customer behavior.</p><p>Five questions guide detection development:</p><ol><li>What malicious behavior should be detected?</li><li>Which logs and fields expose that behavior?</li><li>What logic identifies it?</li><li>How will the logic be tested against malicious and legitimate activity?</li><li>What investigation or response should follow?</li></ol><h3>3.3 Detection Rules and Fraud Risk Scoring</h3><p>Detection rules identify specific suspicious behaviors or combinations of events. Risk scoring evaluates multiple signals to estimate the likelihood of fraud.</p><p>Rules can generate alerts when predefined conditions are met, while risk scoring considers the severity, frequency and context of suspicious activity.</p><p>Financial institutions can use risk scores to determine appropriate responses:</p><ul><li><strong>Low risk:</strong> Continue monitoring.</li><li><strong>Medium risk:</strong> Require additional authentication or analyst review.</li><li><strong>High risk:</strong> Temporarily hold the transaction and initiate a fraud investigation.</li></ul><p>Detection rules and risk thresholds should be tested against legitimate customer activity to reduce false positives and unnecessary payment disruptions. Regular tuning helps maintain detection accuracy as fraud patterns change.</p><h3>3.4 ARTEX: 17 Attack Signals and Detection Opportunities</h3><p>The FBI’s November 13, 2026 advisory describes ARTEX, an AI-driven offensive framework reportedly used to breach South Korean financial institutions, exposing personal information belonging to approximately 67,000 individuals.</p><p>The following 17 vendor-neutral detection proposals identify suspicious activity<strong> that may indicate an AI-assisted cyberattack against a financial institution.</strong></p><h4>1. Access to Forgotten Applications</h4><p><strong>Signal:</strong> Someone accesses an old application that employees rarely use or monitor.</p><p><strong>Detection:</strong> Alert when an unfamiliar IP address accesses an application with very few regular users.</p><p><strong>Telemetry:</strong> Application logs, login records and network activity.</p><h4>2. Automated Reconnaissance</h4><p><strong>Signal:</strong> Automated tools rapidly scan websites or APIs to discover weaknesses.</p><p><strong>Detection:</strong> More than 50 different URLs or 20 HTTP 404 errors from one IP address within five minutes may indicate automated scanning.</p><p><strong>Telemetry:</strong> Web server, API and WAF logs.</p><h4>3. Distributed Login Attacks</h4><p><strong>Signal:</strong> Multiple IP addresses repeatedly attempt to log into the same account.</p><p><strong>Detection:</strong> More than 10 failed logins from over five IP addresses within 10 minutes.</p><p><strong>Telemetry:</strong> Login records, IP addresses and timestamps.</p><h4>4. Possible Account Compromise</h4><p><strong>Signal:</strong> Someone successfully logs in after several failed attempts.</p><p><strong>Detection:</strong> Five or more failed logins followed by a successful login within 10 minutes, especially from a new device.</p><p><strong>Telemetry:</strong> Login history and device records.</p><h4>5. Unexpected Connections to AI Services</h4><p><strong>Signal:</strong> A system that should not use AI connects to external AI services.</p><p><strong>Detection:</strong> Identify connections to AI APIs from systems without permission to use them.</p><p><strong>Telemetry:</strong> DNS, proxy and firewall logs.</p><p><strong>Note:</strong> AI connections are not automatically malicious. Verify whether they are authorized.</p><h4>6. Data Exfiltration</h4><p><strong>Signal:</strong> A system sends unusually large amounts of data outside the organization.</p><p><strong>Detection:</strong> Outbound data transfers exceed three times the system’s normal hourly volume and go to a new destination.</p><p><strong>Telemetry:</strong> Network traffic, proxy logs and data loss prevention records.</p><h4>7. Cloud Permission Testing</h4><p><strong>Signal:</strong> Someone repeatedly checks what cloud resources and permissions they can access.</p><p><strong>Detection:</strong> More than 20 permission checks or resource access tests within five minutes.</p><p><strong>Telemetry:</strong> Cloud activity and IAM logs.</p><h4>8. Unauthorized Privilege Changes</h4><p><strong>Signal:</strong> Someone creates privileged accounts, changes permissions or grants access to applications.</p><p><strong>Detection:</strong> Alert on privileged changes that were not approved.</p><p><strong>Telemetry:</strong> Identity management logs and cloud audit records.</p><h4>9. Suspicious Administrator Access</h4><p><strong>Signal:</strong> Someone accesses cloud management tools or administrator services unexpectedly.</p><p><strong>Detection:</strong> Alert when unfamiliar users or IP addresses access administrative tools without authorization.</p><p><strong>Telemetry:</strong> Administrator login records and cloud activity logs.</p><h4>10. Web Shell Deployment</h4><p><strong>Signal:</strong> An attacker places a malicious file on a web server to execute commands remotely.</p><p><strong>Detection:</strong> Alert when unauthorized processes create or modify PHP, JSP or ASPX files in web application directories.</p><p><strong>Telemetry:</strong> File monitoring, endpoint logs and deployment records.</p><h4>11. Unauthorized Persistence</h4><p><strong>Signal:</strong> An attacker creates a way to regain access after the initial intrusion.</p><p><strong>Detection:</strong> Identify unexpected SSH keys, new privileged accounts or unauthorized credential changes.</p><p><strong>Telemetry:</strong> Operating system logs and account management records.</p><h4>12. Lateral Movement</h4><p><strong>Signal:</strong> An attacker moves from one compromised computer to other systems.</p><p><strong>Detection:</strong> Identify unusual remote access using WinRM or NTLM, or unauthorized directory replication requests.</p><p><strong>Telemetry:</strong> Windows security logs, endpoint activity and directory service logs.</p><h4>13. Security Tool Reconnaissance</h4><p><strong>Signal:</strong> Someone investigates which security tools are installed.</p><p><strong>Detection:</strong> Alert when unauthorized users or suspicious processes inspect antivirus or EDR configurations.</p><p><strong>Telemetry:</strong> Endpoint activity and command execution logs.</p><h4>14. Unauthorized Cloud Resource Creation</h4><p><strong>Signal:</strong> Someone creates cloud servers, public IP addresses or privileged resources without approval.</p><p><strong>Detection:</strong> Alert when cloud resources are created unexpectedly, especially with elevated permissions or public internet access.</p><p><strong>Telemetry:</strong> Cloud activity, IAM logs and configuration changes.</p><h4>15. Command-and-Control Communication</h4><p><strong>Signal:</strong> A compromised system communicates with an attacker-controlled server.</p><p><strong>Detection:</strong> Identify suspicious recurring connections, unauthorized network tunnels or communication with known malicious infrastructure.</p><p><strong>Telemetry:</strong> DNS, network, proxy and endpoint logs.</p><h4>16. Unusual Data Collection</h4><p><strong>Signal:</strong> Someone accesses far more database records or cloud files than normal.</p><p><strong>Detection:</strong> An account reads more than 10 times its usual hourly data volume.</p><p><strong>Telemetry:</strong> Database logs, cloud storage records and account activity.</p><h4>17. Phishing and Impersonation</h4><p><strong>Signal:</strong> Emails impersonate banks, employees or trusted organizations to steal information or request payments.</p><p><strong>Detection:</strong> Identify lookalike domains, failed email authentication and suspicious requests for credentials or money.</p><p><strong>Telemetry:</strong> Email security logs and message analysis.</p><h3>Correlating ARTEX-Related Activity</h3><p>Individual alerts may appear harmless, but a sequence of related events can indicate a larger attack. These behaviors are not unique to ARTEX.</p><h3>3.5 Detection Validation and Performance</h3><p>Before deployment, security teams should test detection rules for:</p><ul><li><strong>Coverage:</strong> Which attacks are detected or missed?</li><li><strong>Accuracy:</strong> Does legitimate activity trigger false alerts?</li><li><strong>Latency:</strong> How quickly does an alert appear?</li><li><strong>Investigation:</strong> Does the alert provide enough evidence to take action?</li></ul><p>Testing should include simulated attacks and legitimate activity. Detection rules should be retested when configurations change.</p><h3>3.6 Threat Hunting</h3><p>Threat hunting proactively searches an organization’s internal network and systems for malicious activity that existing detection rules may miss.</p><p>Security teams use findings from confirmed incidents and attack simulations to identify detection gaps and develop new rules.</p><h3>Part 4. Future Security Testing in Finance</h3><p>AI-driven penetration testing can accelerate vulnerability discovery and continuous attack simulation. Detection engineering should use these results to identify monitoring gaps and improve attack detection.</p><p>Autonomous testing requires strict authorization, monitoring and human oversight.</p><p>PQC and cryptographic agility testing should verify that payment systems, authentication and interbank communications remain secure during cryptographic transitions.</p><h3>Part 5. Next-Generation Incident Response and Recovery</h3><p>As AI accelerates attacks, incident response must support real-time detection and automated containment.</p><ul><li><strong>API observability:</strong> Monitoring should identify shadow endpoints and abnormal token usage.</li><li><strong>Token lineage:</strong> Tracking should identify compromised sessions, token issuance and unauthorized token reuse.</li><li><strong>Behavioral containment:</strong> Controls should block unusual bulk database access and downloads.</li><li><strong>Automated response:</strong> SOAR workflows should support revocation of compromised credentials and certificates, termination of unauthorized sessions and blocking of malicious connections.</li><li><strong>Operational resilience:</strong> Critical financial services should have defined and tested recovery timelines.</li></ul><h3>Part 6. Third-Party Risk and Operational Resilience</h3><h3>6.1 Third-Party Security Requirements</h3><p>Financial institutions should strengthen supplier security through:</p><ul><li>Risk-based patching deadlines.</li><li>AI-assisted vulnerability assessments.</li><li>Software Bill of Materials (SBOM) visibility and verification of software provenance and artifact integrity before deployment.</li><li><strong>Revocation of unused application access, accounts and integrations!</strong></li><li>Tested recovery procedures for critical vendors.</li></ul><p>Security requirements should also extend to subsidiaries and their technology providers.</p><h3>6.2 Operational Resilience and Compliance</h3><p>Institutions should identify concentration risks and validate alternative arrangements for critical financial services.</p><p>The minimum set of financial services required to maintain operations during a major cyber incident should be defined, and their dependencies should be tested to verify restoration within agreed timelines.</p><p>The EU Digital Operational Resilience Act (DORA) and New York’s 23 NYCRR Part 500 establish cybersecurity and operational requirements for covered financial institutions.</p><p>As cyberattacks become faster and more automated, institutions should test incident reporting procedures, simulate disruptions to payment services, verify recovery times and assess whether critical third-party providers can meet their security and recovery obligations.</p><h3>Part 7. AI Security and Security Operations</h3><h3>7.1 Autonomous AI Agent Security</h3><p>AI agents can retrieve information, call tools and execute actions across financial applications.</p><p>Attackers may exploit these capabilities through prompt injection, excessive permissions, memory poisoning and unauthorized tool execution.</p><p>For example, malicious instructions embedded in a document or tool response could cause an agent to disclose customer information or initiate an unauthorized transaction.</p><p><strong>Persistent memory creates additional risk because manipulated information may influence future agent decisions.</strong></p><p>Financial institutions should implement:</p><ul><li>Unique identities for agents and service accounts.</li><li>Least-privilege permissions and restrictions on tool execution.</li><li>Authorization checks before transactions or changes to customer accounts.</li><li>Separation of external content from trusted instructions.</li><li>Memory integrity checks and provenance tracking.</li><li>Human approval for high-impact financial operations.</li></ul><p>Audit logs should link each action to the initiating user, agent identity, delegated permissions, tools invoked and affected resources.</p><h3>7.2 AI-Assisted Security Analysis</h3><p>AI can accelerate investigations by correlating authentication events, transaction activity and security logs.</p><p>AI-generated conclusions must be verified against original records before triggering account restrictions or other consequential actions.</p><p>Financial institutions should evaluate AI systems for unsupported conclusions, missed attack indicators and incorrect recommendations before allowing automated security decisions.</p><p>The NIST AI Risk Management Framework provides guidance for governing, mapping, measuring and managing these risks.</p><h3>Part 8. Emerging Technologies and Future Readiness</h3><h3>8.1 Post-Quantum Cryptography and Financial Industry Migration</h3><p>Future quantum computers could compromise widely used public-key cryptographic algorithms, exposing financial institutions to two major risks:</p><ul><li><strong>Harvest Now, Decrypt Later (HNDL):</strong> Attackers collect encrypted financial data today for potential decryption using future quantum computers.</li><li><strong>Digital signature compromise:</strong> Quantum attacks could undermine the authenticity of signed transactions, financial instructions and communications.</li></ul><p>In August 2024, NIST finalized three post-quantum cryptography standards:</p><ul><li><strong>FIPS 203 (ML-KEM):</strong> Quantum-resistant key encapsulation.</li><li><strong>FIPS 204 (ML-DSA):</strong> Quantum-resistant digital signatures.</li><li><strong>FIPS 205 (SLH-DSA):</strong> Stateless hash-based digital signatures.</li></ul><p>Financial institutions should identify and document vulnerable cryptographic dependencies across payment systems, authentication, digital certificates and interbank communications to prioritize migration <strong>before quantum threats become practical.</strong></p><h3>Financial-Sector PQC Migration</h3><p>In January 2026, the G7 Cyber Expert Group published a financial-sector PQC migration roadmap covering preparation, risk assessment, implementation and validation.</p><p>The roadmap identifies approximately 2035 as a potential transition target, with critical systems potentially prioritized for 2030–2032. These are planning benchmarks, not universal regulatory deadlines.</p><p>NIST IR 8547, finalized in October 2026, <strong>identifies quantum-vulnerable algorithms and their intended quantum-resistant replacements.</strong></p><h3>Cryptographic Dependencies and Migration Priorities</h3><p>A Cryptographic Bill of Materials (CBOM) can help financial institutions document vulnerable algorithms, certificates, security protocols, hardware security modules and third-party cryptographic dependencies.</p><p>Migration priorities should reflect:</p><ul><li>Exposure of payment infrastructure and interbank communications.</li><li>Sensitivity and required confidentiality period of financial records.</li><li>Dependencies on third-party cryptographic services.</li><li>Compatibility with existing authentication and signing infrastructure.</li></ul><p>Migration testing should verify interoperability, performance and resistance to unintended fallback to vulnerable algorithms.</p><p><strong>Cryptographic agility</strong> allows institutions to replace algorithms and protocols without redesigning entire applications.</p><h3>8.2 Digital Assets and Blockchain Security</h3><p>Digital asset platforms face risks involving private-key compromise, unauthorized signing, smart-contract vulnerabilities and wallet delegation.</p><p>Security assessments should focus on:</p><ul><li>Private-key custody and signing authority.</li><li>Separation of privileged responsibilities.</li><li>Transaction approval and delegation controls.</li><li>Emergency revocation of compromised permissions.</li></ul><p>Ethereum’s EIP-7702 introduces account delegation capabilities that require careful authorization controls. Compromised signing authority may allow attackers to establish malicious delegations.</p><p>Wallet security should include verification of delegation destinations and monitoring of EIP-7702 Type-4 transactions.</p><p>PQC migration presents additional challenges for blockchain systems because adopting quantum-resistant signatures <strong>may require network-wide protocol upgrades. </strong>If the blockchain network does not support these upgrades, financial institutions cannot independently complete the migration for affected transactions.</p><h3>8.3 Future Security Architecture</h3><p>Future financial security architectures should prioritize the ability to replace compromised components without interrupting unrelated financial services.</p><p>This requires clearly defined trust boundaries between applications, identity infrastructure, payment systems and external integrations.</p><p>Architecture reviews should evaluate whether critical services can continue operating when individual components or dependencies are isolated.</p><h3>Conclusion</h3><p>AI-driven offensive systems can accelerate reconnaissance, exploitation and privilege escalation. The FBI’s ARTEX advisory highlights the importance of detecting complete attack sequences. The 17 proposed detections address reconnaissance, authentication compromise, lateral movement, persistence and data theft.</p><p>Financial institutions must strengthen detection engineering, correlate cybersecurity and fraud activity, validate attack paths and improve incident response capabilities.</p><p>Preparing for post-quantum cryptography (PQC) also requires identifying vulnerable cryptographic dependencies, developing cryptographic agility and ensuring compromised keys can be revoked and replaced without disrupting critical services.</p><p>Financial institutions have significant work ahead. As offensive AI capabilities advance and quantum threats approach, security architectures must support faster detection, effective containment and continuous adaptation.</p><h3>Glossary</h3><p><strong>ARTEX (Autonomous Penetration Testing Framework):</strong> An AI-driven offensive framework that coordinates reconnaissance, exploitation and multistage attack activity.</p><p><strong>ASPX (Active Server Pages Extended)</strong> is a file extension used by ASP.NET web applications on Microsoft servers.</p><p><strong>Adversary Emulation:</strong> Authorized security testing that reproduces realistic attacker techniques to evaluate defensive controls.</p><p><strong>Behavioral Analytics:</strong> Analysis of activity patterns to identify unusual or suspicious behavior.</p><p><strong>CBOM (Cryptographic Bill of Materials):</strong> An inventory of cryptographic algorithms, certificates, protocols and dependencies used across systems.</p><p><strong>Cryptographic Agility:</strong> The ability to replace or update cryptographic algorithms and mechanisms without major system redesign.</p><p><strong>Detection Coverage:</strong> The extent to which security monitoring can identify specified malicious behaviors.</p><p><strong>Detection Engineering:</strong> The process of developing, testing and maintaining security detections based on observable attacker activity.</p><p><strong>DORA (Digital Operational Resilience Act):</strong> European Union regulation establishing digital operational resilience requirements for financial entities.</p><p><strong>F3 (Fight Fraud Framework):</strong> MITRE’s framework for classifying fraud actor tactics, techniques and procedures.</p><p><strong>FIPS (Federal Information Processing Standards):</strong> U.S. government standards that define security requirements for protecting information, including encryption and digital signatures.</p><p><strong>HNDL (Harvest Now, Decrypt Later):</strong> Collecting encrypted information today with the intention of decrypting it using future quantum capabilities.</p><p><strong>JSP (JavaServer Pages)</strong> is a technology used to create dynamic web pages in Java applications.</p><p><strong>ML-DSA (Module-Lattice-Based Digital Signature Algorithm):</strong> A NIST-standardized<strong> quantum-resistant </strong>digital signature algorithm.</p><p><strong>ML-KEM (Module-Lattice-Based Key-Encapsulation Mechanism):</strong> A NIST-standardized<strong> quantum-resistant </strong>mechanism for establishing shared cryptographic secrets.</p><p><strong>NTLM (New Technology LAN Manager)</strong> is a Microsoft authentication protocol used to verify a user’s identity in Windows networks.</p><p><strong>Pass-the-Hash:</strong> An attack technique that uses a captured password hash for authentication without requiring the original plaintext password.</p><p><strong>Persistence:</strong> Techniques attackers use to maintain access after the initial compromise.</p><p><strong>PQC (Post-Quantum Cryptography):</strong> Cryptographic algorithms designed to resist attacks from quantum computers.</p><p><strong>SLH-DSA (Stateless Hash-Based Digital Signature Algorithm):</strong> A NIST-standardized quantum-resistant digital signature algorithm based on hash functions.</p><p><strong>Trust Boundary:</strong> A point where data or operations cross between systems or components with different security assumptions or levels of trust.</p><p><strong>WinRM (Windows Remote Management):</strong> A Windows service that enables remote administration and can be abused for lateral movement.</p><h3>References and Supporting Frameworks</h3><ol><li>FBI. <em>Threat Overview of ARTEX Autonomous Penetration Testing Framework.</em> FLASH-20261113–001, November 13, 2026.</li><li>MITRE. <em>Fight Fraud Framework (F3).</em></li><li>MITRE. <em>ATT&amp;CK Framework and Cyber Analytics Repository (CAR).</em></li><li>NIST. <em>AI Risk Management Framework (AI RMF 1.0).</em></li><li>NIST. <em>FIPS 203, FIPS 204 and FIPS 205.</em> August 2024.</li><li>NIST. <em>IR 8547: Transition to Post-Quantum Cryptography Standards.</em> October 2026.</li><li>G7 Cyber Expert Group. <em>Financial-Sector Post-Quantum Cryptography Migration Roadmap.</em> January 2026.</li><li>European Union. <em>Digital Operational Resilience Act (DORA).</em></li><li>New York State Department of Financial Services. <em>23 NYCRR Part 500.</em></li></ol><p><em>Dear Reader,</em></p><p><em>Thank you for your time. If you enjoyed reading my article, please share it with others.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9hrz1Wfk7d-cB1r5Ib8aIA@2x.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62fe150df267" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Binance Security Incident Analysis: Attack Path Reconstruction and Detection Engineering]]></title>
            <link>https://medium.com/@angelska327/binance-security-incident-analysis-malicious-code-pyenv-trojanization-and-execution-beefcb84ed61?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/beefcb84ed61</guid>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[binance]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[tradertraitor]]></category>
            <category><![CDATA[lazarus-group]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 05:10:34 GMT</pubDate>
            <atom:updated>2026-11-17T07:43:21.019Z</atom:updated>
            <content:encoded><![CDATA[<h3>1. Executive Summary</h3><p>On October 29, 2026, Binance experienced a multi-stage security incident. According to Binance’s November 3 incident report, approximately $6.34 billion in digital assets were affected, of which approximately $2.85 billion was frozen or intercepted.</p><p>The intrusion began with a fraudulent recruitment process. An engineer downloaded a malicious coding assignment onto a work computer. The attackers replaced a legitimate pyenv component, used an internal AI tool to reach other engineering environments, obtained a container deployment token and compromised wallet-signing operations.</p><p>The attackers subsequently abused Ethereum’s EIP-7702 delegation mechanism to authorize malicious smart contract execution and transfer cryptocurrency from a Binance hot wallet.</p><p><strong>Key finding:</strong> The reported attack crossed multiple security boundaries, from a developer workstation to internal AI tooling, deployment infrastructure and blockchain signing operations.</p><p>This investigation identifies four opportunities for detection engineering:</p><ol><li>Detect unexpected replacement of trusted developer executables.</li><li>Monitor unauthorized cross-user access through internal AI tools.</li><li>Identify suspicious deployment-token usage and unauthorized container deployments.</li><li>Alert on unexpected EIP-7702 wallet delegations.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LatXMxSBAdKIHnbO5d3ftw.png" /><figcaption><strong>Figure 1.</strong> <strong>EIP-7702 delegation</strong>: Etherscan record of the successful EIP-7702 authorization transaction involving Binance 14 on October 29, 2026. The transaction included 50 authorizations and was confirmed in Ethereum block 26,300,753 at 20:20:59 UTC, approximately three minutes before the main asset sweep. Source: Etherscan (2026).</figcaption></figure><h3>2. Detection Engineering: Four Priority Detections</h3><h3>Detection 1: Suspicious pyenv Executable Replacement</h3><p><strong>Attack behavior</strong></p><p>The attackers replaced a legitimate pyenv component on the engineer&#39;s Mac. The replacement file was 70 bytes larger than the original.</p><p>The reported sequence included executable deletion, renaming, symbolic link modification, shell execution and communication with an attacker-controlled server.</p><p><strong>What to detect</strong></p><p>Monitor unexpected modifications to trusted developer executables, particularly pyenv.</p><p>Correlate executable replacement with subsequent shell execution and outbound network activity.</p><p><strong>Required telemetry</strong></p><ul><li>File creation, deletion and rename events</li><li>Symbolic link changes</li><li>Process creation and parent-child relationships</li><li>Outbound network connections</li></ul><p><strong>Recommended alert</strong></p><p>Generate a high-severity alert when an unexpected executable replacement is followed by suspicious shell execution and network communication.</p><p>The 70-byte file-size difference is an investigative indicator, not a reliable standalone detection condition.</p><h3>Detection 2: Unauthorized Cross-User Access Through Internal AI Tools</h3><p><strong>Attack behavior</strong></p><p>The attackers reportedly used an internal AI tool to access other engineers’ systems.</p><p>They subsequently discovered a container deployment token on another engineer’s computer.</p><p><strong>What to detect</strong></p><p>Monitor internal AI tools for privileged actions involving other users’ environments.</p><p>Identify tool calls that exceed the initiating user’s authorized permissions.</p><p><strong>Required telemetry</strong></p><ul><li>AI tool audit logs</li><li>Initiating user and service identities</li><li>Delegated permissions</li><li>Target system access records</li><li>Privileged tool-call activity</li></ul><p><strong>Recommended alert</strong></p><p>Generate an alert when an internal AI tool attempts or performs unauthorized cross-user access or privileged operations. The exact mechanism used to abuse Binance’s internal AI tool remains unknown.</p><h3>Detection 3: Suspicious Deployment-Token Usage</h3><p><strong>Attack behavior</strong></p><p>The attackers obtained a container deployment token and reportedly used compromised engineering access to deploy a malicious container.</p><p><strong>What to detect</strong></p><p>Identify deployment credentials used from unexpected workstations or outside approved CI/CD workflows.</p><p>Correlate unusual token usage with unauthorized container deployments.</p><p><strong>Required telemetry</strong></p><ul><li>Deployment authentication logs</li><li>Token usage records</li><li>Source workstation information</li><li>CI/CD pipeline events</li><li>Container deployment logs</li></ul><p><strong>Recommended alert</strong></p><p>Generate a high-severity alert when deployment credentials are used outside approved workflows to modify sensitive infrastructure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0nhQpzc0br88akzglZohYQ.png" /><figcaption><strong>Figure 2. Main Asset Sweep:</strong> Etherscan transaction record showing the transfer of 167,464.25 ETH from Binance Hot Wallet 14 during the October 29, 2026 exploit. The transaction was confirmed in Ethereum block 26,300,770 at 20:24:23 UTC. Source: Etherscan (2026).</figcaption></figure><h3>Detection 4: Unexpected EIP-7702 Wallet Delegation</h3><p><strong>Attack behavior</strong></p><p>The attackers reportedly submitted approximately 50 authorization attempts and established a malicious EIP-7702 delegation.</p><p>The successful delegation allowed an attacker-controlled smart contract to execute wallet operations.</p><p><strong>What to detect</strong></p><p>Monitor Ethereum Type-4 transactions involving protected wallets.</p><p>Identify delegations to unapproved smart contracts and unusual authorization nonce patterns.</p><p><strong>Required telemetry</strong></p><ul><li>Type-4 transaction records</li><li>EIP-7702 authorization data</li><li>Delegated contract addresses</li><li>Wallet nonce activity</li><li>Asset transfer records</li></ul><p><strong>Recommended alert</strong></p><p>Generate a critical alert when a protected wallet delegates execution to an unauthorized smart contract.</p><p>Monitoring should be combined with pre-signing authorization controls because an on-chain alert may arrive after a malicious delegation has already taken effect.</p><h3>Detection Coverage Summary</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/741/1*dWLZHprZt6Qzx23Y1KLnyA.png" /><figcaption>Table 1. These are proposed detection priorities. Severity should be adjusted to each organization’s environment and operational risk.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XslQVcG8XBuWCzIhTPCs6Q.png" /><figcaption><strong>Figure 3. Reported Attacker Address</strong>: Etherscan identifies Ethereum address 0x7C8601d4645b93cea45F71e118A729332e6B5CA0 as the reported Binance Hot Wallet exploiter. The address is labeled “Exploit” and “Heist,” with transaction history showing initial funding of 0.1 ETH through Tornado Cash. Source: Etherscan (2026).</figcaption></figure><h3>3. Scope and Limitations</h3><p>This investigation examines the Binance incident using Binance’s official disclosure, published security research and publicly accessible Ethereum blockchain records.</p><p>The original malware, endpoint logs, internal AI tool audit logs, deployment records, container logs and network captures were unavailable for independent analysis.</p><p>Reported activity is distinguished from analytical inference.</p><p>The detection engineering recommendations are conceptual and require implementation and testing before production use.</p><h3>4. Offensive Technical Analysis: Reconstructing the Attack Path</h3><h3>4.1 Initial Access: Malicious Coding Assignment</h3><p>The attackers approached an engineer through a legitimate employment platform while impersonating recruitment personnel.</p><p>After multiple interviews, the engineer was instructed to download a coding assignment.</p><p>The repository contained malicious code that executed during setup.</p><p>The original repository was not publicly available for analysis. Therefore, the exact payload location remains unknown.</p><h3>4.2 Trojanized pyenv Executable</h3><p>Binance disclosed the following file-operation sequence:</p><pre>CREATE  /opt/homebrew/Cellar/pyenv/2.8.0/libexec/pyenv_<br>MODIFY  pyenv_ → 3,414 bytes<br>DELETE  /opt/homebrew/Cellar/pyenv/2.8.0/libexec/pyenv (Original: 3,344 bytes)<br>RENAME  pyenv_ → pyenv (Replacement: 3,414 bytes)<br>DELETE  /opt/homebrew/bin/pyenv<br>RECREATE SYMBOLIC LINK (ln)<br>EXECUTE SHELL (zsh)</pre><p>The modified file was 3,414 bytes, compared with 3,344 bytes for the legitimate version.</p><p>This represents an increase of 70 bytes, approximately 2.09%.</p><p>Binance reported that the modified tool was used to execute commands and communicate with an attacker-controlled server.</p><p>Without the original executable, it is not possible to determine exactly what malicious instructions were introduced.</p><p>The reported shell execution does not independently establish persistence across system restarts.</p><h3>4.3 Malware Investigation Methodology</h3><p>If the original malicious repository and modified executable become available, I would apply static and dynamic analysis.</p><p><strong>Static analysis</strong></p><p>Use shasum -a 256, file and strings to examine file characteristics, hashes and embedded content.</p><p>Use diff -u and cmp -l to compare the legitimate and modified pyenv components.</p><p>Review setup scripts and dependencies for unexpected commands or external connections.</p><p><strong>Dynamic analysis</strong></p><p>Execute the suspicious component in an isolated macOS environment.</p><p>Monitor process creation, file modifications, DNS requests and network connections.</p><p>Tools such as tcpdump and fs_usage could help collect relevant telemetry.</p><p>The objective would be to identify the exact execution mechanism, determine whether the modification enabled repeated execution and establish how the malware communicated externally.</p><h3>4.4 Internal AI Tool and Credential Compromise</h3><p>On October 20, attackers reportedly leveraged the initially compromised workstation and an internal AI tool to access additional engineering environments.</p><p>They discovered a container deployment token on another engineer’s computer.</p><p>The reported path was:</p><p>Compromised Workstation → Internal AI Tool → Other Engineering Environments → Deployment Token → Container Infrastructure.</p><p>The available evidence does not establish how the attackers accessed other engineers’ systems or whether the internal AI tool helped locate the deployment token.</p><h3>4.5 Malicious Container Deployment</h3><p>After obtaining the deployment token, the attackers reportedly used compromised engineering access to deploy a malicious container.</p><p>The attack subsequently reached wallet-signing operations.</p><p>The complete internal deployment sequence and authorization decisions were not available for independent examination.</p><h3>4.6 Ethereum EIP-7702 Abuse</h3><p>EIP-7702 allows an Externally Owned Account (EOA) to delegate execution behavior to smart contract code.</p><p>It introduces Type-4 transactions, identified by 0x04.</p><p>The delegation designator is a 23-byte value:</p><pre>0xef0100 + 20-byte contract address</pre><p>Because the attackers reportedly compromised wallet-signing operations, they could generate valid Type-4 authorizations.</p><p>Binance reported approximately 50 authorization attempts, with one successful malicious delegation.</p><p>Based on the reported mechanics, the repeated attempts appear consistent with a nonce-manipulation strategy.</p><p>The exact attacker decision-making process cannot be independently established from the available evidence.</p><h3>4.7 Blockchain Transaction Reconstruction</h3><p>The successful delegation was recorded in Ethereum block 26,300,753 on October 29, 2026, at 20:20:59 UTC.</p><p>The subsequent asset sweep occurred in block 26,300,770 at 20:24:23 UTC.</p><p>The reported transfers included approximately:</p><ul><li>167,464 ETH</li><li>1.615 billion USDT</li></ul><p>The affected wallet was Binance 14:</p><p>0x28C6c06298d514Db089934071355E5743bf21d60</p><h3>4.8 MPC Signing and Emergency Revocation</h3><p>Binance used Multi-Party Computation (MPC) for the affected hot wallet.</p><p>MPC distributes signing authority among multiple participants. No single participant holds the complete private key.</p><p>According to the incident reporting, this complicated the emergency response because clearing the malicious delegation required a specialized signing procedure.</p><p>Suspending ordinary withdrawals did not automatically remove the active on-chain delegation.</p><p>Binance reportedly cleared the delegation by submitting a new EIP-7702 authorization to the zero address at approximately 00:10 UTC on October 30.</p><h3>4.9 Attack Novelty and Comparison With Previous Campaigns</h3><p>The initial access technique was not new. In 2022, the FBI, CISA and U.S. Treasury documented TraderTraitor campaigns involving fraudulent recruitment offers and malicious applications targeting cryptocurrency employees.</p><p>Microsoft’s March 2026 research on Contagious Interview described similar attacks involving fake technical interviews and malicious coding assignments.</p><p>The Binance incident combined these established techniques with reported internal AI tool access, deployment credential compromise and EIP-7702 abuse. The available evidence does not establish that these individual techniques were new.</p><h3>5. Detailed Detection Logic and Validation</h3><p>This section explains how detection engineers could implement and validate the proposed logic.</p><h3>5.1 Endpoint Detection: pyenv Replacement</h3><p><strong>Detection hypothesis</strong></p><p>An attacker may replace a legitimate developer executable and use it to execute malicious commands within a trusted workflow.</p><p><strong>Required telemetry</strong></p><p>Endpoint file events, process execution logs, parent-child process relationships and outbound network connections.</p><p><strong>Proposed detection sequence:</strong></p><ol><li>Identify unexpected replacement or deletion of the legitimate pyenv executable.</li><li>Correlate the replacement with associated rename operations or symbolic link changes.</li><li>Identify subsequent shell execution associated with the modified executable.</li><li>Detect unexpected outbound network connections following execution.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0m7Vfm8VhIleYLw_BZw17Q.png" /><figcaption><strong>Figure 4.</strong> <strong>Reported Attacker-Controlled Smart Contract:</strong> Etherscan identifies smart contract 0x63e842B054Bd624a465454D533C33F573e9c6fAe as associated with the October 29, 2026 Binance hot wallet exploit. The contract is labeled “Binance Hot Wallet Exploiter 24” and flagged for suspicious activity reported by HashDit. Source: Etherscan (2026).</figcaption></figure><h4><strong>False positives</strong></h4><p>Legitimate Homebrew upgrades or developer tool maintenance may produce file modifications and shell execution.</p><p>Detection logic should account for approved updates and expected administrative behavior.</p><h4><strong>Validation</strong></h4><p>Simulate executable replacement, shell execution and outbound network activity in a controlled macOS environment.</p><p>Verify that the detection identifies suspicious activity while avoiding unnecessary alerts during legitimate software updates.</p><h3>5.2 Internal AI Tool Detection</h3><p><strong>Detection hypothesis</strong></p><p>An attacker controlling a developer workstation may abuse an internal AI tool to perform privileged actions across user or system boundaries.</p><p><strong>Required telemetry</strong></p><p>AI tool audit logs, initiating identities, delegated permissions, target resources and privileged tool-call records.</p><p><strong>Proposed detection logic</strong></p><ol><li>Identify privileged AI tool calls involving another user’s engineering environment.</li><li>Determine whether the initiating identity is authorized for the target.</li><li>Identify unusual access to sensitive resources.</li><li>Correlate repeated cross-user activity with endpoint compromise indicators.</li></ol><p><strong>Alert condition</strong></p><p>Generate an alert when an AI tool attempts or performs a privileged action outside the initiating user’s approved authorization scope.</p><p><strong>Validation</strong></p><p>Test authorized and unauthorized cross-user tool calls in a controlled environment.</p><p>Verify that unauthorized tool calls generate detection alerts and that legitimate cross-user activity does not produce unnecessary alerts. Separately, confirm that authorization controls block prohibited actions.</p><h3>5.3 Deployment Token and Container Detection</h3><p><strong>Detection hypothesis</strong></p><p>An attacker may use stolen deployment credentials to introduce malicious containers into sensitive infrastructure.</p><p><strong>Required telemetry</strong></p><p>Token authentication records, source workstation information, CI/CD events, container deployment logs and service identity records.</p><p><strong>Proposed detection logic</strong></p><ol><li>Identify deployment-token usage from unexpected workstations.</li><li>Detect deployments outside approved CI/CD workflows.</li><li>Correlate deployments with unusual authentication activity.</li><li>Identify unauthorized changes to sensitive services.</li></ol><p><strong>Alert condition</strong></p><p>Generate a high-severity alert when deployment credentials are used outside approved workflows to modify sensitive infrastructure.</p><p><strong>Validation</strong></p><p>Simulate unauthorized token usage and container deployment attempts.</p><p>Verify that unauthorized token usage and deployment attempts generate alerts. Separately, test whether deployment authorization controls reject unauthorized requests.</p><h3>5.4 Blockchain Detection: EIP-7702</h3><p><strong>Detection hypothesis</strong></p><p>An attacker who compromises wallet-signing operations may establish malicious delegation through EIP-7702.</p><p><strong>Required telemetry</strong></p><p>Type-4 transactions, EIP-7702 authorization records, delegated contract addresses, wallet nonce changes and asset transfer events.</p><p><strong>Proposed detection logic</strong></p><ol><li>Monitor EIP-7702 authorizations involving protected wallets.</li><li>Identify delegations to contracts outside an approved allowlist.</li><li>Detect unusual authorization nonce patterns.</li><li>Correlate delegation changes with subsequent asset transfers.</li></ol><p><strong>Alert condition</strong></p><p>Generate a critical alert when a protected wallet delegates execution to an unauthorized contract.</p><p><strong>Validation</strong></p><p>Test monitoring against representative Type-4 transactions, including legitimate delegations and simulated unauthorized authorizations.</p><p>Confirm that approved activity does not generate unnecessary alerts.</p><h3>5.5 Detection Engineering Limitations</h3><p>These detection hypotheses are based on publicly reported behavior and available blockchain evidence.</p><p>They have not been independently validated against Binance’s original incident telemetry.</p><p>Before production deployment, each detection would require:</p><ul><li>Mapping to actual event schemas</li><li>Validation using representative malicious activity</li><li>Testing against legitimate administrative behavior</li><li>False-positive analysis</li><li>Alert severity tuning</li><li>Documentation of response procedures</li><li>Assessment of detection coverage</li><li>Identification of false negatives</li><li>Measurement of detection latency</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1pZ8QgIXm6DUIodjnnfJyA.png" /><figcaption><strong>Figure 5.</strong> <strong>Reported Attacker-Associated Wallet</strong>: Etherscan identifies Ethereum address 0x5227f603C1eE34CE3C06eFe6477003BDfCCB2f8c as associated with the October 2026 Binance hot wallet exploit. The address holds approximately $2.09 billion in tokens and is labeled “Binance Hot Wallet Exploiter 2.” Source: Etherscan (2026).</figcaption></figure><h3>6. Preventive Security and Related Risks</h3><h3>6.1 Developer Workstation Security</h3><p>Execute unfamiliar coding assignments in isolated environments.</p><p>Monitor trusted developer executables for unauthorized modifications.</p><p>Evaluate whether developer workstations have unnecessary access to sensitive production resources.</p><h3>6.2 Internal AI Tool Authorization</h3><p>Enforce the permissions of the initiating user for every privileged AI tool action.</p><p>Log delegated actions and access to sensitive resources.</p><p>Investigate whether AI tools can access other users’ environments through excessive permissions or shared service credentials.</p><h3>6.3 Deployment Credential Protection</h3><p>Use short-lived deployment tokens with limited permissions.</p><p>Separate production deployment authority from ordinary developer workstation access.</p><p>Investigate whether sensitive credentials are exposed through local files, environment variables or other accessible resources.</p><h3>6.4 Production Infrastructure Protection</h3><p>Require approved identities and workflows for container deployments.</p><p>Restrict direct production deployment access from developer workstations.</p><p>Monitor unexpected changes to wallet-related services.</p><h3>6.5 Blockchain Signing Security</h3><p>Require strict authorization checks before signing EIP-7702 delegations.</p><p>Maintain approved delegation destinations and independent approval procedures.</p><p>Test emergency revocation processes for wallets protected by MPC.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H6wH5zfzf94m5MDV9WIVtw.png" /><figcaption><strong>Figure 6.</strong> Etherscan overview of Binance 14, the <strong>victim hot wallet </strong>(0x28C6c06298d514Db089934071355E5743bf21d60), showing a remaining balance of approximately 0.808 ETH ($1,466.40) and token holdings exceeding $125,000 at the time of capture, following the October 29, 2026 exploit. Source: Etherscan (2026).</figcaption></figure><h3>7. MITRE ATT&amp;CK Mapping</h3><p>The following mappings are proposed based on reported behavior. They have not been independently validated against original endpoint logs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/745/1*eKI_p7htSp0pTSp-C_CafQ.png" /><figcaption>Table 2. EIP-7702 delegation abuse is analyzed separately because these ATT&amp;CK techniques do not specifically describe the blockchain delegation mechanism.</figcaption></figure><h3>8. FBI Attribution to North Korea’s TraderTraitor</h3><p>On October 31, 2026, the FBI officially attributed the Binance attack to North Korean cyber actors known as TraderTraitor.</p><p>The attack involved tactics associated with this group, including fake recruitment, malicious coding assignments and compromised developer environments.</p><p>The FBI also reported that TraderTraitor actors were converting stolen assets primarily into ETH and warned that additional laundering could occur through decentralized finance and cryptocurrency swapping services.</p><p><strong>Detection engineering relevance:</strong> Security teams should monitor execution of untrusted coding assignments, unexpected modifications to developer tools and cryptocurrency transactions involving addresses identified in FBI advisories.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MRcQe1qMILsNOHxENTatNQ.png" /><figcaption><strong>Figure 7.</strong> Etherscan identifies Ethereum address 0x7C8601d4645b93cea45F71e118A729332e6B5CA0 as the reported <strong>Binance Hot Wallet exploiter.</strong> The address is flagged for involvement in the October 2026 exploit, with an initial funding source listed as Tornado Cash (0.1 ETH). Source: Etherscan (2026).</figcaption></figure><h3>9. Unanswered Questions</h3><ol><li>Why was an external coding assignment executed on a company workstation? Was this permitted under Binance’s security policies?</li><li>Which endpoint security controls were active, and why did they not prevent the malicious pyenv replacement?</li><li>What permissions allowed the internal AI tool to access other engineers’ systems?</li><li>How did the attackers discover the deployment token, and did the internal AI tool help locate it?</li><li>Which security control could have interrupted the attack before it reached wallet-signing operations?</li></ol><h3>10. Conclusion</h3><p>The Binance incident demonstrates how a compromised developer workstation can become an entry point into sensitive production infrastructure.</p><p>The main defensive lesson is to enforce authorization controls at each trust boundary and validate detections against realistic attack scenarios.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zD7_SqyP1tweaC3obEsLzA.png" /><figcaption><strong>Figure 8. Major Asset Sweep:</strong> Etherscan transaction record showing the transfer of 167,464.25 ETH from Binance Hot Wallet 14 during the October 29, 2026 exploit. The transaction was confirmed in Ethereum block 26,300,770 at 20:24:23 UTC. Source: Etherscan (2026).</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zcpODIs-n0UG0Lb1l8Fj8A.png" /><figcaption><strong>Figure 9.</strong> Etherscan identifies Ethereum address 0x5227f603C1eE34CE3C06eFe6477003BDfCCB2f8c as <strong>“Binance Hot Wallet Exploiter 2,”</strong> associated with the October 29, 2026 Binance exploit. The address holds approximately $2.09 billion across 38 tokens and lists Binance 14 as its funding source. Source: Etherscan (2026).</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Plms6KUc-jdj0aT23k59Yw.png" /><figcaption><strong>Figure 10.</strong> <strong>Binance 18:</strong> Etherscan shows that Binance 18 was originally funded by Binance 14 and retained approximately 1.03 ETH and over $858 in tokens. This illustrates that compromising Binance 14 through EIP-7702 did not automatically grant access to Binance 18. Source: Etherscan (2026).</figcaption></figure><h3>11. Glossary</h3><p>EIP-7702 (Ethereum Improvement Proposal 7702): Allows Ethereum wallets to delegate execution to smart contracts.</p><p>EOA (Externally Owned Account): An Ethereum account controlled by cryptographic keys.</p><p>ETH (Ether): Ethereum’s cryptocurrency.</p><p>Etherscan: A website for viewing Ethereum transactions.</p><p>MITRE ATT&amp;CK: A framework describing how attackers operate.</p><p>MPC (Multi-Party Computation): A method of distributing signing authority among multiple parties.</p><p>Nonce: A number used to validate transaction order and prevent reuse.</p><p>On-chain: Activity recorded on a blockchain.</p><p>pyenv: A tool for managing Python versions.</p><p>Smart Contract: A program that runs on a blockchain.</p><p>Sweeper Contract: A program used to transfer cryptocurrency from a wallet.</p><p>TraderTraitor: A North Korean hacking campaign involving fake recruitment.</p><p>Trust Boundary: A separation between systems with different permissions.</p><p>Type-4 (0x04): An Ethereum transaction type supporting EIP-7702 authorizations.</p><p>USDT (Tether USD): A cryptocurrency designed to track the US dollar.</p><p>XRP: A cryptocurrency used on the XRP Ledger.</p><p>Zero Address: An Ethereum address containing only zeros, used to clear EIP-7702 delegation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/321/1*K2zjl7Q_yZ3Fx7OaWKgPHw.png" /><figcaption>Source: Getty Images (UK).</figcaption></figure><h3>12. References</h3><ol><li>Binance. (2026, November 3). <em>29 October 2026 Binance security incident: Initial findings and response update.</em> <a href="https://www.binance.com/en/blog/community/3229153443759566763">https://www.binance.com/en/blog/community/3229153443759566763</a></li><li>Hacken. (n.d.). <em>Binance hack analysis.</em> <a href="https://hacken.io/insights/binance-hack/">https://hacken.io/insights/binance-hack/</a></li><li>Palo Alto Networks. (n.d.). <em>How a developer endpoint became the path to Binance’s $6.3B heist.</em> <a href="https://www.paloaltonetworks.com/blog/security-operations/how-a-developer-endpoint-became-the-path-to-binances-6-3b-heist/">https://www.paloaltonetworks.com/blog/security-operations/how-a-developer-endpoint-became-the-path-to-binances-6-3b-heist/</a></li><li>Ethereum Improvement Proposals. (2024). <em>EIP-7702: Set code for EOAs.</em> <a href="https://eips.ethereum.org/EIPS/eip-7702">https://eips.ethereum.org/EIPS/eip-7702</a></li><li>MITRE. (n.d.). <em>MITRE ATT&amp;CK Framework.</em> <a href="https://attack.mitre.org/">https://attack.mitre.org/</a></li><li>Etherscan. (2026). <em>Ethereum blockchain explorer: Binance 14 wallet and associated transactions.</em> <a href="https://etherscan.io/address/0x28C6c06298d514Db089934071355E5743bf21d60">https://etherscan.io/address/0x28C6c06298d514Db089934071355E5743bf21d60</a></li><li>FBI. (2026, October 31). <em>North Korea Responsible for $6 Billion+ Binance Attack. </em><a href="https://www.fbi.gov/investigate/cyber/alerts/2026/north-korea-responsible-for-6-billion-binance-attack">https://www.fbi.gov/investigate/cyber/alerts/2026/north-korea-responsible-for-6-billion-binance-attack</a></li><li>Cybersecurity and Infrastructure Security Agency. (2022, April 18). <em>North Korean state-sponsored APT targets blockchain companies (AA22–108A).</em> <a href="https://www.cisa.gov/sites/default/files/publications/AA22-108A-TraderTraitor-North_Korea_APT_Targets_Blockchain_Companies.pdf">https://www.cisa.gov/sites/default/files/publications/AA22-108A-TraderTraitor-North_Korea_APT_Targets_Blockchain_Companies.pdf</a></li><li>Microsoft Security. (2026, March 11). <em>Contagious Interview: Malware delivered through fake developer job interviews.</em> <a href="https://www.microsoft.com/en-us/security/blog/2026/03/11/contagious-interview-malware-delivered-through-fake-developer-job-interviews/">https://www.microsoft.com/en-us/security/blog/2026/03/11/contagious-interview-malware-delivered-through-fake-developer-job-interviews/</a></li></ol><p>Thank you to my new mentor, Mike Krygier, for assigning me this task. I had fun working on it.</p><p>Dear Reader,</p><p>Thank you for your time. If you liked this post, please share it with others.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/484/1*8n1P2DhFWiMdfBP6vsFWAw@2x.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=beefcb84ed61" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Exploit PHP2Shell (CVE-2026–109618 & 109619): From Out-of-Memory to Remote Code Execution]]></title>
            <link>https://medium.com/@angelska327/exploit-php2shell-cve-2026-109618-109619-from-out-of-memory-to-remote-code-execution-ad6d66b3871d?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad6d66b3871d</guid>
            <category><![CDATA[php2shell-vulnerabilities]]></category>
            <category><![CDATA[php]]></category>
            <category><![CDATA[php2shell]]></category>
            <category><![CDATA[exploit]]></category>
            <category><![CDATA[kali-linux]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Sat, 31 Oct 2026 03:48:58 GMT</pubDate>
            <atom:updated>2026-10-31T05:56:55.514Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>I recently was challenged by my new mentor to find or build a working PHP exploit, analyze it, and document the findings without releasing the weaponized payload. I chose to look into the “PHP2Shell” vulnerabilities (CVE-2026–109618 and CVE-2026–109619) that were recently disclosed. While both vulnerabilities share the same root cause — PHP failing to clean up memory during early request processing — my successful lab reproduction specifically leveraged the <strong>Out-of-Memory corruption path (CVE-2026–109618)</strong>. These are critical memory corruption bugs in PHP that allow an attacker to run commands on a web server.</p><p>While millions of automated scanners have tried to exploit this in the wild, confirmed, reliable Remote Code Execution (RCE) has been very difficult to achieve.</p><p><strong>This article documents how I successfully </strong>reproduced the exploit in a safe, isolated lab environment on my Kali Linux machine, and breaks down exactly how the vulnerability works under the hood.</p><h3>Summary of the Attack Chain</h3><p>To achieve Remote Code Execution, the exploit relies on a four-step chain combined with precise memory math:</p><ul><li><strong>Step 0 (Pin):</strong> Lock 149 workers with fake traffic. This forces the server to process the entire attack on the 1 remaining worker, keeping the memory state stable.</li><li><strong>Step 1 (A1):</strong> Send a massive file upload to cause the Out-of-Memory crash. Because the server panics, it skips cleanup, creating the “dangling pointer.”</li><li><strong>Step 2 (B1):</strong> Trigger an out-of-bounds memory read. This steals the secret memory addresses (like the libphp base address).</li><li><strong>Step 3 (A2 + B2):</strong> Send the final payload to overwrite pDestructor with zif_system, forcing the server to execute the id command.</li></ul><p>The exploit relies on offsets (like libphp+0x65EB60). The exploit uses these math formulas to calculate exactly where zif_system lives inside the server&#39;s memory based on the leaked libphp address.</p><h3>The Goal &amp; The Lab Setup</h3><p>My goal was to get the server to run the id command and return uid=33(www-data). To prove successful Remote Code Execution (RCE) without causing damage to the target, I used this built-in Linux command because it safely returns the current user context. It just asks the computer: <em>&quot;Who am I right now?&quot;</em> If the server answers this for us, it proves we successfully tricked it into running our instructions. Because the output is www-data, we know we achieved RCE as the low-level web server user and only have control over the website files.</p><p>I used a local Docker container running PHP 8.4.26 with Apache (mod_php). First, I acquired a public proof-of-concept (PoC) toolkit to my Kali machine to simulate the attack:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/428/1*JWOOykUCr40uaYHcdtRJnw.png" /><figcaption><strong><em>Figure 1:</em></strong><em> Initializing the local Docker lab. The script successfully starts the vulnerable PHP 8.4.26 target, which will be the victim of the exploit.</em></figcaption></figure><p>The reason this specific configuration is targeted is that official security bulletins explicitly identify Apache mod_php as the highest-risk deployment method.</p><p>In this mode, PHP runs as an Apache module, and the vulnerability triggers during the interpreter’s memory management phase before the application code even begins to execute. The fix versions are PHP 8.4.27 and higher. My lab intentionally used the unpatched 8.4.26 to verify the exploit’s effectiveness.</p><h3>Step 1: Breaking the Server (The OOM Bailout)</h3><p>To hack a server, you have to break its normal flow. PHP has a feature called session.upload_progress which runs very early in a request, <em>before</em> normal request initialization.</p><ol><li>The exploit script sends a specially crafted file upload that forces the PHP server to run out of memory (OOM).</li><li>When PHP runs out of memory, it uses a mechanism called a longjmp to jump out of the request early and abort it.</li><li><strong>The Bug:</strong> Because it jumped out early, it skips the cleanup phase. The session ID pointer (PS(id)) is left pointing to a memory address that has already been freed. It is a &quot;dangling pointer.&quot;</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/563/1*eaUeVUIgFVxEZ8QqcSv9Iw.png" /><figcaption><strong><em>Figure 2: </em></strong>When I ran the exploit, the terminal output confirmed the exploit triggered this exact Out-of-Memory condition (A1: END-cleanup send_cookie OOM).</figcaption></figure><p>To bypass the server’s worker pool and guarantee the memory state was preserved, the exploit used a technique called ‘worker pinning.’ As seen in the terminal output (pinned 150/150 workers (149 holders)), the script held 149 connections open to distract the server. This forced the server to process the attack on the single remaining worker, making the exploit reliable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/564/1*y33h-IxjbE43rwKhLO2pTQ.png" /><figcaption><strong>Figure 3:</strong> Because this exploit relies on winning a race condition (timing the memory allocation perfectly), it does not always work on the first try. In my lab, the first attempt resulted in a heap leak failed error because the timing was off.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/557/1*uxTqo6slL9l5QbobxbGmHw.png" /><figcaption><strong>Figure 4:</strong> On the next attempt, the memory layout aligned perfectly, and the exploit successfully executed the command.</figcaption></figure><h3>Step 2: Stealing Memory Addresses (The Info Leak)</h3><p>Modern operating systems use <strong>ASLR (Address Space Layout Randomization),</strong> which shuffles memory addresses so hackers can’t guess where things are. To bypass this, I needed to leak addresses from the server.</p><ol><li>The exploit sends a second request. The server tries to use that “dangling” session ID pointer.</li><li>Because the memory was freed and then reallocated with fake data crafted by the exploit, the server ends up reading out-of-bounds (OOB).</li><li>The server takes this leaked memory address and accidentally sends it back to me in the Set-Cookie HTTP header.</li><li>With this leak, the exploit now knows exactly where the Zend memory manager and the libphp library are located in memory.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/571/1*wsYPAPOvZPpWfH6g6Su6Yw.png" /><figcaption><strong>Figure 5:</strong>The successful memory leak.</figcaption></figure><h3>Step 3: Taking Control (The Exploit Payload)</h3><p>Now that we have the memory addresses, we can attack.</p><ol><li>The exploit sends a final crafted request to overwrite a specific piece of memory called pDestructor.</li><li>Normally, pDestructor points to a function that cleans up memory. The exploit overwrites this pointer so it points to zif_system instead (a built-in PHP function that executes shell commands).</li><li>When the PHP server finishes processing the request, it tries to “clean up” by calling pDestructor.</li><li>But instead of cleaning up, it executes my command: id.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/579/1*_M4AsgJ5HSmjHYSnEASnYw.png" /><figcaption><strong>Figure 6:</strong> The successful exploit chain- the terminal shows the calculation of zif_system (the command execution function) followed immediately by the [4/4] RCE OK and the uid=33 proof.</figcaption></figure><p>At this exact moment in the script, the exploit uses the leaked libphp address to calculate where zif_system lives. It then silently overwrites the pDestructor pointer with this address. We can&#39;t see pDestructor on the screen because it happens inside the server&#39;s RAM, but the immediate uid=33 output proves the overwrite was successful.</p><h3><strong>Verifying the Sibling Vulnerability (CVE-2026–109619)</strong></h3><p>While my primary lab reproduction used the Out-of-Memory path (CVE-2026–109618), I also verified the second PHP2Shell vulnerability, CVE-2026–109619 (stale global state), using a dedicated differential lab.</p><p>This path relies on a Producer/Consumer relationship:</p><p>the attacker first triggers an OOM crash (the Producer) to leave a stale session pointer in memory, and then sends a second request (the Consumer) to that same worker to consume the stale state before normal initialization can clean it up.</p><p>The exploit operates in three distinct stages, all visible in the terminal output below:</p><ol><li><strong>Heap Leak:</strong> The exploit crashes the worker to leak the Zend Memory Manager chunk base.</li><li><strong>Library Leak:</strong> It triggers a second crash to leak the exact memory address of libphp.</li><li><strong>Control Hijack:</strong> Using the leaked offsets, it calculates the address of zif_system and overwrites the memory pointer.</li></ol><p>Although the worker crashes with a Segmentation Fault during the final stage, the payload executes successfully before the process dies, as proven by the uid=33(www-data) output. Furthermore, the differential lab confirms the patch works: the exact same attack against PHP 8.4.27 fails with a heap leak failed error.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/697/1*TNOk173rcaT6hKHEPQYOQA.png" /><figcaption><strong>Figure 7: </strong>The execution of CVE-2026–109619. The terminal shows the three stages: leaking the Zend heap chunk, leaking the libphp base address, and calculating zif_system. The final proof confirms code execution as uid=33(www-data) on the vulnerable PHP 8.4.26.</figcaption></figure><p>The screenshot below shows the patched target. The exploit attempts the exact same attack, but because the patch resets the internal state before the second request can consume it, the memory leak fails completely. The lab concludes with a confirmed differential result.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/694/1*2VJa72_pa6cBjTrcbOxnow.png" /><figcaption><strong><em>Figure 8:</em></strong><em> The negative control. The exact same exploit is run against the patched PHP 8.4.27. The server successfully blocks the attack (</em><em>heap leak failed), resulting in the final confirmation: vulnerable code executes, patched code does not.</em></figcaption></figure><h3>How do we stop this?</h3><p>The official PHP release notes say <strong>“There are no complete mitigations, updating is the only safe choice”.</strong></p><p>The vulnerability is fixed in PHP versions 8.2.35, 8.3.36, 8.4.27, and 8.5.12. The patch ensures that both during early Out-of-Memory errors and fatal crashes, PHP properly resets the internal state and cleans up the session pointers.</p><p>So, if you are running a PHP server, <strong>updating to the latest version is the only permanent fix</strong>.</p><h3>Glossary</h3><p><strong>ASLR (Address Space Layout Randomization):</strong> A security technology that randomizes memory locations to block attacks. The exploit bypasses it by leaking the exact addresses.</p><p><strong>Dangling Pointer:</strong> A memory pointer that references an address that has already been deleted or freed. The OOM crash causes the session pointer to dangle, allowing access to leftover memory.</p><p><strong>Differential Testing:</strong> A testing method that compares how different versions of software behave under the same conditions. In this lab, it compares vulnerable PHP 8.4.26 with patched PHP 8.4.27.</p><p><strong>Heap:</strong> A region of memory where programs dynamically allocate and free data during execution.</p><p><strong>Heap Leak:</strong> The unintended exposure of information from heap memory. In this exploit, leaked addresses help locate important memory structures.</p><p><strong>libphp:</strong> The core software library that executes PHP. Leaking its memory address gives the exploit the exact location of the PHP code.</p><p><strong>longjmp:</strong> A C programming function that jumps back to a previously saved execution point, bypassing the normal sequence of function returns.</p><p><strong>Memory Corruption:</strong> A condition where a program incorrectly modifies or accesses memory, potentially causing crashes or allowing attackers to manipulate execution.</p><p><strong>Memory Offset:</strong> The distance between two memory addresses. Attackers can use a known base address and an offset to calculate another address.</p><p><strong>Negative Control:</strong> A test designed to demonstrate that an expected effect does not occur under specific conditions. Here, the patched PHP version serves as the negative control.</p><p><strong>OOM (Out-of-Memory):</strong> A server error condition where a program uses up all its allowed memory and crashes. The exploit triggers this to interrupt the server’s normal cleanup process.</p><p><strong>OOM Bailout:</strong> An emergency exit from program execution when memory runs out. In this exploit, the bailout skips normal cleanup and leaves invalid pointers behind.</p><p><strong>OOB (Out-of-Bounds) Read:</strong> A memory access flaw where a program reads data outside its assigned boundaries. The exploit uses the dangling pointer to trigger this and steal the hidden memory addresses.</p><p><strong>pDestructor:</strong> A built-in function pointer used to clean up and free memory. The exploit overwrites this pointer to hijack the server’s execution flow.</p><p><strong>PoC (Proof of Concept):</strong> A demonstration that shows whether a vulnerability can be triggered or exploited under specific conditions.</p><p><strong>Producer/Consumer:</strong> A relationship where one operation creates a state that another operation later uses. In this exploit, the Producer leaves stale memory state and the Consumer accesses it.</p><p><strong>Race Condition:</strong> A situation where the outcome depends on the timing or order of operations. Exploits may require precise timing to succeed.</p><p><strong>RCE (Remote Code Execution):</strong> A vulnerability that allows an attacker to execute commands on a remote system.</p><p><strong>Segmentation Fault (Segfault):</strong> A crash that occurs when a program attempts to access memory in an invalid or unauthorized way.</p><p><strong>Stale Global State:</strong> Internal program data that remains accessible after it should have been reset or invalidated.</p><p><strong>Use-After-Free (UAF):</strong> A vulnerability that occurs when a program accesses memory after it has been freed.</p><p><strong>Worker:</strong> An individual server process that handles incoming web traffic. The exploit needs the exact same worker to handle both requests so the memory stays corrupted.</p><p><strong>Worker Pinning:</strong> A technique that holds 149 connections open to occupy 149 workers. This forces the server to use the one remaining worker for the attack, keeping the memory state stable.</p><p><strong>Zend Memory Manager:</strong> PHP’s internal system for allocating, tracking and freeing memory during execution.</p><p><strong>zif_system:</strong> A built-in PHP function that executes shell commands. The exploit redirects pDestructor here so the server runs the id command.</p><h3>References</h3><p>PHP Group. (2026). <em>PHP 8.4.27, 8.3.36, 8.2.35, and 8.5.12 released</em> [Security advisory]. https://www.php.net/archive</p><p>VulnCheck. (2026). <em>CVE-2026–109618 and CVE-2026–109619: PHP2Shell memory corruption</em> [Threat intelligence report].</p><p>Wiz Security Research. (2026). <em>Understanding the PHP2Shell vulnerability chain: OOM bailout to RCE</em> [Technical blog post].</p><p>Sophos. (2026). <em>PHP2Shell: Why reliable exploitation remains elusive in the wild</em> [Security analysis].</p><p>PoC[undisclosed]</p><h3>Dear Reader,</h3><p>Thank you for your time. If you enjoyed my article share it with other.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/963/1*sepzatcjQB4h_cRY0mJLKw.png" /><figcaption>Me and Mocha.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad6d66b3871d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Threat Modeling 101— BSides Toronto]]></title>
            <link>https://medium.com/@angelska327/threat-modeling-101-bsides-toronto-519784c8ba46?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/519784c8ba46</guid>
            <category><![CDATA[toronto]]></category>
            <category><![CDATA[data-modeling]]></category>
            <category><![CDATA[bsides]]></category>
            <category><![CDATA[threat-modeling]]></category>
            <category><![CDATA[canada]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 04:25:29 GMT</pubDate>
            <atom:updated>2026-10-12T04:27:40.826Z</atom:updated>
            <content:encoded><![CDATA[<p>On October 3, I sat in a workshop at BSides Toronto taught by Yuk Fai Chan, co-leader of the OWASP Toronto chapter. He had volunteered to run Threat Modelling 101. I arrived excited to build on what I had previously learned about threat modeling.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xKGk1NYrrk70lDeM_2mg-g.jpeg" /><figcaption>TM 101, BSides Toronto.</figcaption></figure><h3>1. Before You Hunt, Do Threat Modeling</h3><p>The first distinction the instructor drew was between two activities that share a word and almost nothing else.</p><p><strong>Threat modeling</strong> is proactive. You examine a design, proposed or existing, and ask what can go wrong. You do it early, while the shape of the system can still be changed without great cost. The work is mostly thinking, drawing, and arguing about assumptions.</p><p><strong>Threat hunting </strong>is operational. You examine data from systems that are already running, including logs, endpoint telemetry and network traffic. The goal is to find signs of suspicious or malicious activity that existing security tools and alerts may have missed.</p><p>Threat modeling asks, “What could go wrong, and how can we prevent it?” Threat hunting asks, “What suspicious or malicious activity is occurring in our environment”? Both are necessary. The earlier you find a design weakness, the cheaper it usually is to fix.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vrdSSQ92ugVMqOx_EBG2Rw.jpeg" /><figcaption>Thank you to my anonymous donor!</figcaption></figure><h3>2. First Exercise: Protect a House</h3><p>Imagine you are leaving home for a week. What do you want to protect? My team responded, jewelry, passports, property documents, keys and electronics. What could happen while you are gone? A thief picks the lock. Someone breaks a window. Heavy rain floods the basement. A family member who should not have a key lets themselves in.</p><p>Then we worked on the defenses: stronger locks, reinforced doors, alarms, glass-break sensors, a sump pump, insurance. Our task was simple on paper. Name three or four assets. List several threats for each. Propose two or three ways to reduce each risk. Share with the group.</p><p>We all made assumptions: the lock works, the basement stays dry and the person with the spare key can be trusted. Software carries the same kind of assumptions, except they hide inside an API call, a database query or an authorization check that no one has looked at in months. The house made those assumptions visible before we drew a diagram.</p><h3>3. Four Questions That Make the Threat Modeling Look Easy</h3><p>The workshop used Adam Shostack’s four questions as the spine of the day:</p><p>1. What are we working on?</p><p>2. What can go wrong?</p><p>3. What are we going to do about it?</p><p>4. Did we do a good job?</p><p>Before you can name attacks you have to understand the system.</p><ol><li>What is in scope?</li><li>Which assets matter?</li><li>Who uses it?</li><li>What enters and leaves?</li><li>Which external services does it depend on?</li><li>Where does the application stop trusting an input, or where do the permissions under which a request runs change? This helps us identify where to place trust boundaries.</li></ol><p><strong>A trust boundary </strong>is a point in a system where data or requests move between areas with different levels of trust or permissions. In a threat model, trust boundaries are placed between components with different permissions or security requirements, such as a user’s browser and a website’s server, or between a normal user role and an administrative one. Crossing a trust boundary should prompt questions about authentication, authorization, input validation and logging.</p><p>Trust boundaries matter because they<strong> identify where security risks can arise</strong> when data or requests move between parts of a system with different levels of trust.</p><p>For example, a website’s server cannot assume that information coming from a user’s browser is safe or that the user has permission to perform the requested action</p><p>Threat modeling is not a one-time worksheet. Whenever the team adds or changes a feature, component, user role, dependency or data flow, the threat model needs to be revisited.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Szf6yG634gBWStwcLENlMw.jpeg" /><figcaption>Scope Slide - TM 101, BSides Toronto.</figcaption></figure><h3>4. How to Draw a DFD (Data Flow Diagram)</h3><p>A data flow diagram maps how information travels. It usually contains four kinds of elements:</p><ul><li>External entities: sources or destinations of data, such as a customer or a payment processor.</li><li>Processes: anything that handles or transforms data, such as login or checkout.</li><li>Data stores: databases, files, caches.</li><li>Data flows: the arrows that show what moves between them.</li></ul><p>Trust boundaries are also added to the diagram, usually as dashed lines, to show where data crosses between areas with different levels of trust or permissions.</p><p>The data flow diagram shows what information actually travels, and under what assumptions. Looking at all three together gives a clearer picture than any one of them alone.</p><p>Each arrow raises a few questions: What data is moving between these components? Who controls it? Is it validated?</p><h3>5. Who Might Attack, and Why?</h3><p>Before choosing specific attacks we talked about threat actors. Knowing who might try something changes what counts as a reasonable concern.</p><p>• Nation-state groups pursue intelligence or disruption.</p><p>• Cybercriminals seek money, customer data, or leverage for extortion.</p><p>• Hacktivists act from ideological motives.</p><p>• Insiders already hold some legitimate access.</p><p>• Thrill-seekers may simply want to see whether they can get in.</p><p>• The workshop also mentioned terrorist groups.</p><p>For our workshop, we used EcomSite as an example. Potential threat actors included competitors, financial criminals, spammers, malicious insiders and curious attackers. A spammer harvesting email addresses poses a different threat from an attacker trying to steal an entire customer database.</p><h3>6. STRIDE: Six Ways to Ask What Can Go Wrong</h3><p>STRIDE is a mnemonic that helps you look systematically instead of relying on whatever attack comes to mind first.</p><p><strong>Spoofing. </strong>Pretending to be a legitimate user or service. Example: using stolen credentials to impersonate a customer. Ask how identities and sessions are verified.</p><p><strong>Tampering. </strong>Changing data or behavior without permission. Example: altering a submitted quantity, product price, or inventory record. Ask what the server validates and who can write to the record.</p><p><strong>Repudiation. </strong>Disputing an action when the system lacks reliable evidence. Example: a seller denies changing a price. Ask whether meaningful actions are logged and whether those records can be trusted.</p><p><strong>Information disclosure. </strong>Exposing sensitive information to someone who is not authorized to access it. Example: a website accidentally reveals customers’ personal information, such as names, addresses or payment details. Ask whether sensitive data is properly protected and accessible only to authorized users.</p><p><strong>Denial of service. </strong>Making the application unavailable or exhausting resources. Example: repeated expensive search requests slow the store for everyone. Ask about resource limits and resilience.</p><p><strong>Elevation of privilege. </strong>Obtaining capabilities beyond those assigned. Example: a seller gains access to another seller’s management actions, or a customer reaches an admin function. Ask whether role and ownership checks are enforced.</p><p>Other attack techniques discussed during the workshop included SQL injection, insecure direct object references (IDOR), man-in-the-middle attacks, expensive regular expressions that can cause denial of service and even physical tailgating. These are specific ways an attack might happen, while STRIDE helps classify the types of security threats they can create. A single attack technique may fall under more than one STRIDE category, depending on its impact.</p><p>However, STRIDE does not cover every aspect of an attack or its business impact. For an e-commerce application like EcomSite, potential threats include credential stuffing, payment fraud, coupon abuse, fake reviews, inventory hoarding and automated scraping. Many of these can be mapped to STRIDE categories, but the framework alone may not reveal how legitimate features can be misused. For example, a customer might exploit discount combinations, abuse refund policies or manipulate product rankings without bypassing access controls. Identifying these risks requires examining the application’s business logic, how its features are intended to work and how attackers could exploit them for financial gain or other purposes.</p><h3>7. Attack Trees: Start with the Attacker’s Goal</h3><p>Next, we looked at attack trees. The root represents the attacker’s main objective, the branches show possible routes to achieve it and the leaves represent specific actions.</p><p>We used a safe full of money as our example. The main goal was to steal the money, but achieving it required subgoals. These included locating the safe, observing when the owner leaves and finding a way to access it.</p><p>The same thinking applies to EcomSite. Suppose the objective is to read another customer’s order. Possible routes to investigate include changing the order ID in a request, guessing other customers’ order IDs or finding another interface that exposes the same information.</p><h3>8. The EcomSite Architecture</h3><p>The fictional store has customers who search and buy, sellers who manage products and fulfillment, and administrators who oversee accounts and reporting. The architecture shown in the workshop was organized in layers:</p><p>• Client layer: a web browser.</p><p>• Presentation layer: Jinja2 templates and a REST JSON API.</p><p>• Flask application layer: authentication, customer, seller, admin, and API blueprints.</p><p>• Business logic: authentication helpers, stock and order operations, and PyOTP-based multifactor authentication.</p><p>• Data layer: SQLAlchemy ORM connected to PostgreSQL.</p><p>An important distinction is the difference between a rendered web page and a REST API. If a security control exists in one route, does the other route enforce it too? The diagram showed both interfaces, but it did not establish how every authentication or authorization check was implemented.</p><p>The data model added more detail. The USERS table stores account fields such as email, password hash, role, active status, and MFA settings. PRODUCTS links items to sellers. CART_ITEMS connects customers with products and quantities. ORDERS records purchases and shipping details. ORDER_ITEMS records the items, quantities, and unit prices associated with each purchase. Roles include customer, seller, admin, and admin_viewer. That raises a simple question: is an admin_viewer truly read-only everywhere?</p><p>Order records also preserve snapshots of details such as shipping addresses, product names, and prices so a later profile or catalog change does not rewrite purchase history. That helps record integrity, but it also raises retention and privacy questions.</p><h3>9. Level 1 DFD: The Whole Store on One Page</h3><p>The Level 1 diagram placed the customer, seller, and administrator around the main application processes, with a database underneath. Its major processes were account and authentication management, catalog and search, order and cart processing, and administration and reporting.</p><p>This overview is useful for scope. It shows who interacts with which part of the system. But it is too broad to answer detailed questions about how a checkout works or when MFA is verified. That is why we moved to Level 2 diagrams, which break each major process into smaller steps.</p><h3>10. Level 2: Account and Authentication</h3><p>The account diagram included user registration, profile management, security and MFA setup, primary authentication, and MFA verification. It also showed interactions with stored user data, including credential and MFA-related information.</p><p>Our worksheet included a brute-force login scenario: an external attacker repeatedly guesses passwords. We discussed mitigations such as rate limits or lockouts, CAPTCHA, and MFA. Each has trade-offs. Aggressive account lockouts can themselves be abused to deny legitimate users access. MFA needs to be enforced correctly, and the secret needs protection.</p><p>Other questions raised included: Can a user change another person’s profile? Can an attacker bypass MFA through a different route? Does changing security settings require appropriate re-verification? These are potential security risks to investigate, not confirmed vulnerabilities.</p><h3>11. Level 2: Catalog and Search</h3><p>The search diagram showed how many steps are involved in processing a search request. A customer submits a search query, price range or seller filter. The request passes through query parameter parsing, a visibility and seller filter, search logic that reads product records, then pagination and result formatting before the response goes back.</p><p>That visibility filter matters. If it fails, could unpublished or restricted products appear in search results? Could expensive queries exhaust resources? Could untrusted product descriptions create a cross-site scripting problem when displayed?</p><p>Search functionality can introduce security risks beyond those associated with login. Even without requiring a password, it can expose sensitive data or consume significant system resources if appropriate controls are missing.</p><h3>12. Level 2: Carts, Checkout, and Orders</h3><p>The order diagram was the most detailed. A customer adds or removes items from the cart. During checkout, the system checks product visibility and stock availability. When the order is finalized, it creates an order and its associated items, updates inventory and clears the cart. The customer receives an order ID. Sellers can later update fulfillment information, and the system combines these updates to determine the overall order status.</p><p>One of the two threats documented on our workshop worksheet involved a customer changing the order ID in a URL to access someone else’s order. This could lead to information disclosure if the application does not verify that the customer is authorized to access that specific order. A possible mitigation is to check ownership or other appropriate access permissions on every request, even when order IDs are difficult to guess.</p><p>Another potential concern is inventory consistency. What happens if two customers try to buy the last available item at nearly the same time? The diagram included an atomic stock decrement to prevent incorrect inventory updates, but failed operations and database transaction boundaries also need to be considered. An operation that works correctly on its own may behave differently when multiple requests arrive simultaneously.</p><p>Other areas worth investigating include whether the server independently recalculates product prices instead of trusting values submitted by the customer, and whether order creation, inventory updates and cart clearing are handled consistently. If one of these operations fails, the system needs to prevent incomplete orders, incorrect stock counts or other inconsistent results.</p><h3>13. Level 2: Seller Functions</h3><p>The seller diagram covered product lifecycle management, sales and stock analytics, order fulfillment and order-status aggregation. Sellers can create or edit product listings, update inventory, remove products from sale, review low-stock information and manage order fulfillment.</p><p>Sellers have permission to manage their own products. When a seller submits a request to edit a listing, the server needs to check whether that product belongs to them. This check is especially important when the request contains a product ID that could be changed to reference another seller’s listing.</p><p>This illustrates the difference between authentication and authorization. Authentication verifies who the user is, while authorization determines what that user is allowed to access or modify. Ownership and permission checks are important for product edits, unlisting, inventory changes and fulfillment updates. The diagram showed how these operations work, but did not confirm whether the necessary security checks were implemented.</p><h3>14. Level 2: Administration and Reporting</h3><p>The admin diagram covered system-wide analytics, user account management, order oversight and detailed reporting. Some functions allow administrators to access information across multiple users and orders, while others allow them to change account roles or account status.</p><p>Because these functions involve access to large amounts of data and sensitive settings, mistakes can have serious consequences. Can a user with the admin_viewer role export detailed reports? Could those reports expose customer addresses or email addresses unnecessarily? Are changes to user roles logged? Do sensitive administrative actions require appropriate authorization or approval?</p><p>The workshop diagram showed a role check for user management, but the permissions for individual reporting functions were unclear. These details would need to be confirmed with the application team before drawing conclusions about potential security weaknesses.</p><h3>15. From a Diagram to a Threat Register</h3><p>The final exercise involved creating a threat register. For each component or data flow, we identified potential threats, threat actors, STRIDE categories, affected assets, impacts and possible mitigations.</p><p>Our worksheet was mostly empty, with only two scenarios started: brute-force login and changing an order ID to access another customer’s order. Developing a useful threat model takes time, and each threat needs enough detail to investigate.</p><p>A threat description should explain who might act, what they could do, which asset is affected and the potential consequences. For example, “SQL injection” is a general label, while “An unauthenticated attacker manipulates API input to access customer records” describes a specific scenario worth investigating. Neither confirms an actual vulnerability.</p><h3>16. How Do We Decide What Matters Most?</h3><p>How cybersecurity teams prioritize threats after identifying them during threat modeling?</p><h3>16. How Do We Decide What Matters Most?</h3><p>How do cybersecurity teams prioritize threats after identifying them during threat modeling?</p><p>Threats vary in likelihood and impact.</p><p>A <strong>risk matrix</strong> helps teams prioritize threats by comparing how likely they are to occur and how much damage they could cause. Each factor is rated, usually as <strong>low, medium or high</strong>, and their combination determines the overall risk level. For example, a threat that is highly likely to occur and could expose sensitive customer data would receive a high priority.</p><p>Some risk matrices also use numerical scores. A common formula is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/390/1*I4Ie6NddlP3Zp-yOWVMsCg.png" /><figcaption>A Risk Score common formula</figcaption></figure><p>with each factor rated from 1 to 5. For example, a likelihood of 4 and an impact of 5 produce a risk score of 20 out of 25, which could be classified as Critical depending on the organization’s scoring criteria.</p><p>Another approach is <strong>DREAD</strong>, which stands for Damage potential, Reproducibility, Exploitability, Affected users and Discoverability. For example, when evaluating a potential SQL injection vulnerability, we can consider how much data could be exposed, how easily the attack could be repeated, how difficult it would be to exploit, how many users might be affected and how easily the vulnerability could be discovered.</p><p>Both approaches help teams decide which threats to address first. However, risk ratings are estimates and should be supported by evidence, existing security controls and potential business impact.</p><h3>17. Four Responses to Risk</h3><p>• Mitigate: reduce likelihood or impact with controls such as parameterized database queries, authorization checks, MFA, rate limits, or transaction safeguards.</p><p>• Avoid: remove the risky functionality or activity entirely.</p><p>• Transfer: shift some financial consequences, for example through insurance, without assuming the technical risk disappears.</p><p>• Accept: document a conscious decision to live with a risk based on its severity, context, and cost of further treatment.</p><p>A control needs to address the actual scenario. Audit logs may help reconstruct an unauthorized change, but logging does not prevent SQL injection. Encryption can protect data in some situations, but it does not replace object-level authorization. OWASP guidance, including its Cheat Sheet Series and Proactive Controls, can help when selecting specific defenses.</p><h3>18. What the Diagrams Did Not Show</h3><p>Revisiting the material, some questions remained unanswered. The order tables contained totals, but the diagrams did not clearly show a payment processor, payment token or webhook. If EcomSite used these in a real deployment, they would introduce additional dependencies, data flows and trust boundaries. The diagrams also lacked details about audit-log storage and protection.</p><p>The model included MFA-related fields but did not explain how MFA secrets are stored securely or rotated. These missing details should be documented as open questions rather than assumed vulnerabilities.</p><p>Threat modeling identifies potential threats and assumptions. Confirming vulnerabilities requires further investigation through design reviews, code inspection, configuration checks or testing.</p><h3>19. What I Would Add to the Worksheet Next</h3><p>If I continued the exercise I would investigate five more scenarios: simultaneous checkouts affecting stock integrity; a seller changing another seller’s product; an admin_viewer accessing unnecessarily detailed reports; API routes that enforce weaker checks than equivalent web routes; and unauthorized access to stored MFA secrets or backups.</p><p>For each I would first ask the system owner how the relevant control is actually implemented. I would also add the missing payment and logging flows if they are within scope. Only then could I evaluate the risk with confidence.</p><h3>20. My Reflections After the Workshop</h3><p>The workshop also made me think about threat modeling beyond traditional web applications.</p><p>In agentic AI systems, information can come from retrieved documents, tool outputs, other agents and persistent memory. These sources may contain untrusted content that influences an agent’s decisions and actions. Where should we draw trust boundaries when information moves between these components? How can we distinguish legitimate instructions from instructions embedded in untrusted data? And what happens when an agent treats that data as something it should obey?</p><p>Thank you to Yuk Fai Chan for volunteering his time and delivering a threat modeling workshop through practical and fun exercises.</p><p><em>Dear Reader, thank you for your time.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AVjESur7YbcFoxODqNh9IQ.jpeg" /><figcaption>Toronto, Canada</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4_0vfp1P_xhFiErBtmE4Mg.jpeg" /><figcaption>On my way to BSides Toronto.</figcaption></figure><p>Thank you to my anonymous donor who gifted me tickets to SecTor and BSides Toronto.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=519784c8ba46" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Who Gets to Decide? How Helen Oakley’s SecTor Keynote Made Agentic AI Security Concrete.]]></title>
            <link>https://medium.com/@angelska327/who-gets-to-decide-how-helen-oakleys-sector-keynote-made-agentic-ai-security-concrete-5d622de71894?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/5d622de71894</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[canada]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[blackhat]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Sat, 10 Oct 2026 03:31:25 GMT</pubDate>
            <atom:updated>2026-10-10T03:38:43.733Z</atom:updated>
            <content:encoded><![CDATA[<p>On Thursday, October 8, I attended Helen Oakley’s keynote at SecTor in Toronto, “Who Gets to Decide? Security in the Agentic Enterprise.”</p><p>Seeing Helen Oakley deliver a technical keynote reminded me how powerful female representation in cybersecurity can be.</p><p>Her confidence, expertise and presence on stage sent a clear message to other women: “Yes, you too can be here. You belong.” A woman delivering a keynote can open doors for other women in cybersecurity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sukdyqzx1Cl4GR07Mxh1ug.jpeg" /><figcaption>With Helen Oakley at SecTor 2026.</figcaption></figure><p>I saw someone doing work I care about and explaining it with clarity and passion.</p><p>She used Alice’s story to show how an AI workflow can exceed the authority it was given.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gPXvfoO5asf8Es69CWVGyA.jpeg" /><figcaption>Helen Oakley, Keynote at SecTor 2026</figcaption></figure><h3>Twenty Minutes of Access</h3><p>It was a regular morning. A support request.</p><p>Alice works in tech support. She asks an AI agent to investigate a customer issue using an approved support path. The access is supposed to last 20 minutes.</p><p>By the end of the workflow, an admin console is reachable from the public internet.</p><p>No hacker needed to contact the agent. The starting point was someone trying to do their job.</p><p>The agent creates a plan. <strong>The plan asks for a connectivity change. </strong>That request reaches a service operating under an identity that can change network rules. The service carries out the change. Nothing checks whether it stays within the scope of Alice’s request.</p><p>Alice asked for temporary support access. The workflow changed who could reach the service.</p><p>Somewhere between the request, the plan and the execution, permission changed meaning. <strong>The service’s permissions stood in for task-specific authority. </strong>The service checked its own permissions, but not whether the change was allowed for Alice’s task. The resulting change exceeded the approved support scope.</p><p>This is what losing the original task limits can look like. As a request moves between components, its constraints can be dropped or left unenforced. Alice’s request for limited access becomes public exposure downstream. The action carried out in production exceeds what she requested.</p><p>This is how quickly we can lose control.<strong> A generated plan can become a production change before anyone checks whether that change was authorized.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*JZlxZDAKIUl68rjw22ChpA.jpeg" /><figcaption>Arsenal 4, SecTor, OWASP AIBOM (AI Bill of Materials) Generator.</figcaption></figure><h3>When Authority Gets Lost in the Handoff</h3><p>Agentic AI is moving fast. <strong>A</strong>gents can plan, delegate and use tools across connected systems. Following what happens to authority as a task moves through those systems is another challenge entirely.</p><p>A service might have permission to change network rules. That does not mean every task reaching that service is authorized to use its full power.</p><p>“Capability is not authority”, Helen said.</p><p>The executor needs to check whether this particular task authorizes the proposed change, including its target, reach and duration.</p><p><strong>Every handoff needs to preserve the original limits. </strong>In Alice’s case, that means approved support access, for this task, for 20 minutes. Those limits need to reach the service making the change in a form it can actually enforce.</p><p>If the proposed action would make the console public, <strong>the workflow should stop.</strong> Broader access needs a new authorization decision and a fresh look at the risk that change introduces. If the scope changes, the threat model needs to be revisited too.</p><p>These risks connect to categories OWASP identifies in its Top 10 for Agentic Applications: Tool Misuse and Exploitation and Identity and Privilege Abuse. No single component guarantees that the full workflow respects its limits. The relevant checks need to hold at every boundary.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*38RENgWlkEfH2Azwi5GgzA.jpeg" /><figcaption>Keynote: Helen Oakley at SecTor 2026, Toronto, Canada</figcaption></figure><p>How do we establish what actually changed? Missing audit evidence leaves gaps in our understanding. <strong>The workflow needs to record the actions taken and their outcomes, and those records must be protected from unauthorized alteration or deletion.</strong></p><p>This is why getting the foundations right from the start matters. Once agents connect to production systems, their mistakes can affect real infrastructure, private data and people.</p><p>A development change can affect a production database if the environment allows that reach. Recovery matters. Restoring deleted data does not fix the authorization failure that allowed its deletion. Separation between development and production helps constrain what a development workflow can affect.</p><p>We already have pieces of the foundation: identity, authorization, least privilege, scoped access and a record of who was authorized to do what. We also need a record of what happened during the workflow. <strong>The work is to make those controls hold across the full path, including the point where a tool executes a change.</strong></p><p>Someone has to own the consequences.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n67LX4MlMVvki0Sb3TipnA.jpeg" /><figcaption>Helen Oakley’s Presentation</figcaption></figure><p>Helen’s proposed operating model brings business and service owners, engineers and security teams into a shared responsibility.</p><p>One takeaway is that ownership needs to be clear before the workflow runs: who approves its scope, who implements the controls and who responds when something goes wrong?</p><p>Her 30-, 60- and 90-day approach offers a practical starting point.</p><ul><li><strong>Assess: In the first 30 days,</strong> choose a consequential workflow, such as a financial process. Understand what it can change, where authority passes between components and who owns the outcome. Threat model the workflow. Identify and test boundaries where permissions or scope could change. Define what behavior should be allowed and what must be blocked.</li><li><strong>Enforce: By 60 days,</strong> bring one useful workflow under control. Implement its limits. Approved support access must become a grant that the executing service can verify. A note in a ticket cannot enforce that restriction on its own. Test both authorized actions and requests that exceed the approved scope, then examine the results.</li><li><strong>Expand: By 90 days,</strong> earn the expansion. Evaluate whether the evidence supports extending the workflow’s reach or deploying it elsewhere. Can it block out-of-scope actions? Can access be revoked? Has recovery been tested? Include third-party systems in this assessment. Understand the permissions, dependencies and data flows they introduce. A successful demo alone is not sufficient evidence.</li></ul><h4><strong>Three principles to carry forward:</strong></h4><ol><li>Know the reach.</li><li>Enforce the limits.</li><li>Check the outcome.</li></ol><p>The main question is this: <strong>What is the agent authorized to cause?</strong></p><p>It focuses on the consequences of an agent’s actions, distinguishes technical capability from authorization, questions whether task limits survive delegation and handoffs and establishes the need for clear ownership and accountability.</p><p>Helen’s topic is particularly relevant in 2026 because AI agents can exceed their authorized scope, creating risks that organizations must address as they deploy and govern these systems.</p><p>Her impact goes beyond the technical lessons she shared. That is what I mean when I say one woman can open doors for others. Seeing her on stage can encourage other women to submit a talk, share their research or speak up about a technical problem they understand.</p><p><em>Dear reader, thanks for your time.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Wy2jsPqIC_Ss3dwlQ6ZByA.jpeg" /><figcaption>SecTor 2026, Toronto, Canada.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AVjESur7YbcFoxODqNh9IQ.jpeg" /><figcaption>Downtown Toronto, Ontario, Canada.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5d622de71894" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What you are building is valid]]></title>
            <link>https://medium.com/@angelska327/what-you-are-building-is-valid-35205b2b97f2?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/35205b2b97f2</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[trauma]]></category>
            <category><![CDATA[mentorship]]></category>
            <category><![CDATA[seniors]]></category>
            <category><![CDATA[authority]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 01:08:53 GMT</pubDate>
            <atom:updated>2026-09-02T01:57:39.791Z</atom:updated>
            <content:encoded><![CDATA[<h3>You do not need anyone’s permission to keep going.</h3><p>This matters especially when you are early in cybersecurity.</p><p>Cybersecurity can make you feel small quickly. You enter a room where someone has 20 years of experience, another has found CVEs and someone else has spoken at Black Hat.</p><p>Suddenly, your three years of learning can feel insignificant.</p><h4>But being new does not mean you cannot identify a gap and fill it.</h4><p>You can build something useful and have an idea someone more experienced did not have.</p><p>Experience deserves respect. It does not require obedience. Cybersecurity is technical, but you bring your psychology into it: how you learned to handle criticism, authority, rejection and how you learned to trust yourself.</p><p>Someone who grew up seeking approval may treat a senior engineer’s opinion like an instruction. Someone who learned to do everything alone may struggle to ask for mentorship or collaborate. These patterns influence how you learn, build relationships and develop your career.</p><p>Building a cybersecurity career can also rebuild your relationship with yourself.</p><h3>Short list of my key points when you’re early in cybersecurity:</h3><h3>1. The gatekeepers don’t own the perimeter anymore</h3><p>You do not need a hiring manager or industry veteran to validate your skills before proving what you can do. Open-source tools, home labs, bug bounties and public repositories have lowered the barrier to entry -build and create evidence before someone hires you.</p><p>Build. Document. Contribute. Publish.</p><h3>2. Failure is information, not identity</h3><p>A failed attempt does not define your potential. A mistake means you investigate what happened and determine what to improve.</p><p>Cybersecurity already teaches you to investigate failure. Apply the same principle to yourself.</p><p>What failed and why? What needs to change?</p><h3>3. Separate feedback from old wounds</h3><p>Technical criticism can sometimes activate something much older.</p><p>Someone questions your work and suddenly the thought becomes:</p><p>“Maybe I do not belong here, I am too inexperienced.”</p><p>If you learned early that criticism meant something bigger, your reaction can become bigger too. Instead of evaluating the criticism, you start evaluating yourself. That is when you need to separate the present feedback from the older emotional wound.</p><p>Ask yourself: Is the criticism actually valid? That question brings you back to the work.</p><h3>4. Someone else’s confidence is not evidence that your judgment is wrong</h3><p>If you grew up with a parent who was controlling, loud and certain, you may have learned to read confidence as authority and authority as something you defer to.</p><p>So when someone radiates certainty, even if their argument is weak or wrong, you may start shrinking before you have evaluated whether they actually know what they are talking about.</p><p>This matters in cybersecurity because you will meet people who speak with enormous confidence.</p><p>They may have senior titles and years of experience, speak in conclusions, and sound like the discussion is already over.</p><h4>But someone speaking in conclusions is using a style, not presenting proof.</h4><p>Confidence and correctness are not the same thing. Someone can sound certain and be wrong.</p><p>A confident tone, title or seniority does not replace evidence.</p><p>You can listen to what someone is saying without automatically treating the confidence behind it as proof that their judgment is better than yours.</p><h3>5. Authority is not absolute truth</h3><p>Mentors, CISOs, researchers and senior engineers can be wrong. You can respect someone’s experience and still question their conclusion.</p><p>Cybersecurity itself depends on questioning assumptions. That is one of the strongest lessons cybersecurity can teach you. You are constantly asking:</p><p><strong>Can we trust it, can we verify the underlying data that supports this person’s claim, is there a single point of failure if this individual makes a mistake, how do we audit this person’s actions the same way we log a system, what happens if this person’s credentials or trust are compromised, and are we trusting this decision based on authority rather than objective evidence?</strong></p><p>That way of thinking should not disappear when the assumption comes from a person instead of a system.</p><p>Someone having more experience doesn&#39;t necessarily mean their opinion deserves serious consideration. And it does not mean automatic obedience. Respecting expertise and thinking independently are not opposites.</p><h3>6. Being open to feedback does not mean accepting every suggestion</h3><p>Cybersecurity depends on review, criticism and challenge. You will meet people who test your assumptions and challenge your idea if there is a real weakness in it. But listening does not require surrendering your judgment.</p><p>Being open to feedback means you consider it, but you don’t have to follow it. Investigate whether that could improve your work. If not, say</p><p>“I understand the suggestion, but I am keeping my approach.”</p><p>That response can come from someone who is genuinely open to feedback. Do not redirect because someone challenged your approach.</p><h3>7. Evidence is your anchor</h3><p>Cybersecurity already teaches this:</p><p>Test. Collect. Verify.</p><p>If someone says your idea is weak, investigate it.</p><p>If someone says your project is unnecessary, look at whether it actually helps people and what it produces. Answer it with evidence.</p><h3>8. You do not have to convince everyone</h3><p>Sometimes people understand exactly what you are saying and simply disagree. Don’t keep producing better explanations until everyone is satisfied and approves of what you are doing.</p><h4>That game can never end.</h4><p>The urgency to explain can become its own trap.</p><p>You explain once. They disagree. You explain again. They still disagree and they keep questioning you. You add more context. You try another angle.</p><p>Eventually, you are no longer improving the work. You are trying to produce an explanation that will finally make someone give you permission to continue. Stop.</p><p>People can understand you perfectly and still disagree. Explain only when necessary. Then return to the work.</p><h3>9. Be visible</h3><p>Cybersecurity has established communities, strong personalities and people who have worked together for years. Entering those spaces, contributing and becoming visible can feel uncomfortable.</p><p>Sometimes other people may also be uncomfortable with you and your growth. Let them.</p><h4>There is room for fresh blood: new contributors, researchers, speakers and leads.</h4><p>Someone else already being established does not mean you have to remain small. Growth changes how you participate.</p><p>You speak more, question something, stop automatically agreeing and start your own initiative. You may become visible. That is growth.</p><h3>10. Seniority can create blind spots a beginner may catch</h3><p>Experience gives people pattern recognition, but it can also create rigid mental models. People get used to existing systems and ways of doing things.</p><p>A newcomer does not have all those assumptions yet. That can actually be an advantage. Unfamiliarity can reveal what familiarity hides.</p><p>The gap you can see from the doorway may be the one the room stopped noticing, especially when it comes to rigid corporate workflows.</p><p>One of my mentors said, cybersecurity “dinosaurs”, -experienced senior professionals, need fresh blood - early-career people who bring new perspectives and question things that everyone else has simply gotten used to.</p><h3>11. Technical arrogance is not technical competence</h3><p>In cybersecurity, you will meet people who use jargon aggressively. People who question in a way that feels more like a test than a discussion.</p><p>People who make simple concepts sound overly complicated. People who use condescension as if it proves expertise. Do not confuse hostility with intelligence.</p><p>A person making you feel small may be defending their ego.</p><p>The strongest practitioners often do something very different. They are patient and explain complex ideas clearly. They make room for you and your questions.</p><h4>They can challenge your thinking without humiliating you.</h4><p>They can tell you that you are wrong without making the entire interaction about your worth.</p><h3>12. You have the right to be a beginner</h3><p>You are allowed to misunderstand something, correct yourself and learn publicly. You do not have to prove that you know everything before you are allowed to participate.</p><p>If you decide that you are only allowed to contribute once nobody can correct you, you may wait forever.</p><h3>13. Being easy to work with can become self-erasure</h3><p>It can feel safer to become the person who never creates friction. Agree with the senior person.</p><h4>Do not say no. Do not challenge the direction or the rigid workflow.</h4><p>That makes you easy to work with and it also slowly removes your own judgment from the room.</p><p>Cybersecurity needs people willing to ask uncomfortable questions.</p><p>Should we trust this? Why do we trust this? Is this control enough? Why has nobody fixed this? Why does nobody tell them this?</p><p>If you spend your entire career trying to be easy, you become very good at disappearing. Being collaborative does not require that. You can be kind without becoming small. You can disagree without becoming difficult.</p><h3>14. Clarity can feel like coldness at first</h3><p>If you are used to over-explaining, direct communication can feel uncomfortable.</p><p>“I disagree with that assessment. I understand your position but I am keeping my approach.”</p><p>Those sentences can feel much harsher than they are.</p><p>You may feel pressure to add another paragraph. Explain your motives. Soften. Make sure nobody feels uncomfortable. That is a fawn response.</p><h4>Clarity is not disrespect.</h4><p>Cybersecurity often requires clear communication.</p><p>A risk either exists or it does not. A control either addresses the problem or it does not. A decision has to be made. Communicate clearly without managing everyone’s reaction to the clarity.</p><h3>15. Mentorship should strengthen your judgment, not replace it</h3><p>Mentorship matters. You can appreciate your mentor’s time.</p><h4>But basic professional respect should not create permanent indebtedness.</h4><p>Someone listening to you does not buy ownership over your direction. Respect can exist without dependence.</p><p>Healthy professional relationships should help you grow into stronger judgment. They should not make you afraid to disagree with the person who helped you.</p><h3>16. The body can keep the old timeline longer than the mind</h3><p>You can intellectually understand that you are allowed to speak and still feel your body react when you do. You can know that you have a right to disagree and still feel guilty afterward. You can know that your contribution has value and still feel yourself shrinking when someone more senior challenges it.</p><h4>You can react to a confident tone as if it were authority.</h4><p>This is unnecessary. The mind can understand something before the body fully catches up.</p><p>The feeling is information.</p><h3>17. Cybersecurity can teach you how to practice adulthood</h3><p>Cybersecurity will teach you much more than cybersecurity.</p><p>It teaches you how to distinguish evidence from opinion.</p><p>How to think through uncertainty, fail and try again, receive criticism, handle rejection, ask difficult questions, learn from people without automatically obeying them, hear confidence without mistaking it for proof, set a boundary and stop explaining, participate before you feel completely ready and keep going when the person you wanted to learn from rejects you or does not respond.</p><p>It may hurt at first. It hurt me.</p><p>And if you were not fully allowed to grow or make decisions for yourself when you were younger, your career may become one of the places where you finally practice those things as an adult.</p><p>Respect experience without surrendering your judgment, accept mentorship without replacing your own thinking, be corrected without disappearing and becoming shaky.</p><p>Fail without making failure your identity, be new without treating yourself as insignificant, build, lead and contribute before somebody decides you are ready.</p><h4>What you are building is valid.</h4><h4>You do not need anyone’s permission to keep going.</h4><h4>You do not owe anything to anyone.</h4><p>Dear early in cybersecurity,</p><p>Thanks for reading.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BMdbnuOjV1HjX7dWIkxKeQ.jpeg" /><figcaption>IEEE Project at Stanford</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qcKybdf4PTgf7M_y7Y-5wQ.jpeg" /><figcaption>IEEE Project at Stanford</figcaption></figure><p>If you want to join OWASP and you do not know how to start or which project to choose, please message me on LinkedIn.</p><p>Angelika</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=35205b2b97f2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Persistent Memory as a Security Boundary in Agentic AI — UPDATED 09/09/26]]></title>
            <link>https://medium.com/@angelska327/persistent-memory-as-a-security-boundary-in-agentic-ai-f1ef2dcc05f2?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/f1ef2dcc05f2</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Thu, 18 Jun 2026 23:43:11 GMT</pubDate>
            <atom:updated>2026-09-10T19:19:41.663Z</atom:updated>
            <content:encoded><![CDATA[<p>UPDATED 09/09/2026</p><h3>Abstract</h3><p>Persistent memory, dynamic context assembly and autonomous tool execution change the security model of agentic AI. These architectures enable temporal decoupling between compromise and consequence: a poisoned memory write or corrupted tool schema may influence actions across subsequent sessions, days or weeks. Agentic memory must therefore be managed as security-relevant state, with explicit controls governing its influence on decisions and actions.</p><p>This section provides practical, secure-by-design guidance across three architectural pillars:</p><p><strong>First</strong>, it addresses memory integrity, isolation and lifecycle management through authorized write paths, input validation, cryptographic tamper detection, tenant partitioning and versioned recovery to contain self-reinforcing contamination.</p><p><strong>Second</strong>, it establishes context trust boundaries and assembly integrity through retrieval-time defenses, provenance-preserving assembly and separation of trusted instructions from untrusted content across users, sessions and agent domains.</p><p><strong>Third</strong>, it defines mediated (validated), least-privilege tool execution through independently enforced authorization at the point of action, validated tool definitions, sandboxing and short-lived, narrowly scoped credentials to constrain side effects. Across these pillars, tamper-resistant audit logging and incident-response mechanisms support detection, containment and recovery, while recognizing that cryptographic integrity alone does not establish content trustworthiness.</p><p>Ultimately, the goal is to help architects design agentic systems that preserve state integrity, isolation, trustworthiness, auditability and recoverability throughout long-lived deployments. Broader identity management and inter-agent protocols are covered elsewhere; this section focuses on the operational security boundaries of memory, context and tool execution.</p><p>New Key Components (KCs)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/1*c0jgrwcEg367U84ws6brtw.png" /><figcaption>Key Components</figcaption></figure><h3>Memory and State Hygiene</h3><p>Persistent memory introduces a security boundary in agentic AI. These systems may retain instructions, retrieved context, reasoning artifacts, summaries and behavioral patterns across sessions. Memory may also be shared among users, tools and collaborating agents. Compromised memory can silently influence later reasoning, orchestration and tool execution long after the original write.</p><p>Treat memory as security-relevant state, with explicit controls governing its influence on decisions and actions. Integrity verification, provenance tracking, access controls, isolation and lifecycle management help prevent or contain memory poisoning, delayed activation of malicious instructions, replay attacks, cross-session contamination and unauthorized state changes. Integrity checks alone do not establish that stored content is trustworthy.</p><p>Baseline session caching, memory flushing and host-level runtime triggers must follow the environment criteria specified in Version 1.0, Section 9.3 (Memory and State Hygiene).</p><h3>Memory Type Classification and KC4 Mapping</h3><p>Different memory classes expose distinct attack surfaces and require corresponding security controls. The following table maps common agent memory types to the KC4 Memory Module taxonomy. Persistence, sharing scope and the authority granted to retrieved content determine how compromised memory can influence subsequent actions. Controls should address both the initial compromise and its propagation into derived summaries, shared context and downstream tool execution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/627/1*HWQWqevejCEZ6TM5YtXnZQ.png" /><figcaption>Memory Type Classification and KC4 Mapping</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/621/1*5Hy-3TlBDV93W_kzhjYsrw.png" /><figcaption>Memory Type Classification and KC4 Mapping</figcaption></figure><h3>Secure Memory Write Path</h3><p>Agentic runtimes should enforce a controlled, cryptographically validated write path for all persistent memory operations. Accept updates only from authorized execution flows and bind each request to its intended operation, target memory scope and execution context. Block direct writes from untrusted tools and external inputs. Their content may be stored only through an authorized workflow that validates it before storage.</p><p>Use HMAC-protected envelopes or digital signatures to bind the payload to authenticated request metadata. Include the requesting identity, user or tenant scope, session or workflow identifier, timestamp, nonce or unique transaction identifier and key identifier. A policy-enforced memory gateway should verify this binding, enforce authorization and reject stale or replayed requests. Service attestation may provide additional evidence about the requesting workload but does not replace these checks.</p><p>Cryptographic validation confirms integrity and authenticity relative to the key used. It does not establish authorization, user intent or content trustworthiness. Poisoned content submitted through a legitimate workflow can still pass integrity checks. Combine cryptographic verification with independent authorization, provenance checks and content validation before storage.</p><h4>Architectural Implementation Considerations</h4><ul><li>Enforce the write gateway at the storage boundary to prevent bypass.</li><li>Bind each update to the requesting identity, target memory scope and authorized execution context.</li><li>Validate request integrity, payload schemas, data types and size limits before committing changes.</li><li>Enforce freshness checks and prevent replay through nonce tracking or idempotent transaction handling.</li><li>Use version checks or atomic conditional writes to prevent stale updates from overwriting newer state.</li><li>Maintain tamper-resistant audit logs for memory creation, modification, rollback and deletion.</li><li>Support quarantine and recovery to verified state, including affected summaries, embeddings and caches.</li></ul><h4>Key Management Requirements</h4><p>Keep private signing keys and HMAC secrets <strong>outside the agent’s accessible runtime in a separate trust domain</strong>, such as a Key Management Service (KMS), Hardware Security Module (HSM) or isolated cryptographic service. Restrict key use by workload identity and permitted operation. Define procedures for key generation, rotation and revocation.</p><p>This separation reduces key-extraction risk but does not prevent a compromised runtime from misusing an authorized signing service. The signing service or memory gateway must independently enforce authorization and scope restrictions. Public signature-verification keys may reside with verifiers if their integrity and trust configuration are protected.</p><p>This control primarily mitigates T1 (Memory Poisoning) by blocking unauthorized state changes and detecting tampering. Combined with least-privilege execution and tool governance, it also reduces T3 (Privilege Compromise) risks. Because authorized workflows can still introduce poisoned content, write-path controls must operate alongside retrieval-time defenses and memory lifecycle controls.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/926/1*2AgOzAd_CuxMt9X4xNCdIA.png" /><figcaption>Secure State Mutation Verification &amp; Commit Flow</figcaption></figure><h3>Memory Integrity, Provenance and Semantic Trust</h3><p>Cryptographic integrity checks detect changes relative to an authenticated record. They do not establish content trustworthiness. A signed entry can still contain poisoned, misleading or malicious content introduced through a legitimate write path.</p><p>Every memory record should carry tamper-resistant provenance identifying its origin, writing entity, source trust classification and originating execution context. Identify content from external tools, retrieved documents, user submissions or prior-session context. Preserve this metadata throughout the record’s lifecycle and record changes to trust assessments as auditable updates.</p><p>Provenance must govern retrieval behavior. Apply access restrictions, context isolation, quarantine or reduced retrieval priority according to source trust and task scope. Suspected malicious content should be excluded or quarantined rather than merely ranked lower. Retrieved content must not gain authority to override instructions or authorize privileged actions.</p><p>Preserve source links and trust classifications through summarization, embedding, transformation and compaction. Derived records must not receive a higher trust classification solely because they were transformed. For records combining sources with different trust levels, preserve those distinctions or apply the most restrictive applicable classification.</p><p>Revalidate retrieved content for delayed malicious instructions triggered by a later prompt, tool call, user action or execution condition. Acceptance during an earlier session does not establish safety in the current context.</p><p>Use semantic checks to detect gradual poisoning that makes malicious content more likely to appear in future retrievals. Flag unexpected associations with privileged operations, abnormal clusters, repeated retrieval of unrelated content and results outside the task’s scope. Investigate flagged records using their provenance and transformation history. Semantic similarity alone does not establish trust.</p><p>These controls reduce risk but do not guarantee safety. Combine them with secure write paths, vector-memory protections and pre-storage validation.</p><h3>Long-Term Memory Lifecycle Controls</h3><p>Persistent memory requires security controls across creation, retrieval, summarization, archival, rollback, expiration and deletion. These controls should limit the persistence of untrusted content, support recovery after compromise and reduce long-term behavioral manipulation.</p><p>Recommended controls include:</p><ul><li><strong>Retention and expiration:</strong> Set time-to-live (TTL) limits and periodically remove memory that is no longer required.</li><li><strong>Secure deletion and archival:</strong> Restrict access to archived records and propagate deletions to affected summaries, embeddings and caches.</li><li><strong>Provenance tracking:</strong> Preserve source links and trust classifications through summarization and transformation.</li><li><strong>Rollback and recovery:</strong> Quarantine compromised records and verify recovery points before restoring them.</li><li><strong>Pre-storage validation:</strong> Check content, schemas and provenance before storage. Apply sanitization without treating it as proof of safety.</li><li><strong>Replay protection:</strong> Detect stale requests and prevent duplicate execution of memory updates.</li><li><strong>Mutation monitoring:</strong> Flag unexpected updates, unauthorized retention extensions and abnormal propagation across agents.</li></ul><p>Account for delayed activation of persistent malicious content. Hidden instructions in records or summaries may survive across sessions and activate after a particular prompt, user interaction, tool call or execution condition. Poisoned entries may influence later reasoning or spread compromised state to collaborating agents. Hallucinated content may also persist and propagate without malicious intent.</p><p>Revalidate memory when retrieved because harmful content may activate before expiration. When compromise is detected, quarantine the source and affected derivatives and preserve evidence for investigation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/691/1*16T-d1At1cHaPKzTaX2J6w.png" /><figcaption>Memory Lifecycle Security Matrix</figcaption></figure><h3>Sensitive Data and Privilege-Bearing Memory</h3><p>Treat credentials, private cryptographic keys, session tokens, capability-bearing file-system handles, administrative capabilities and infrastructure control channels as privilege-bearing assets. Keep them outside model-visible context and ordinary agent memory stores. Mediate their use through controlled interfaces that independently enforce authorization and least privilege.</p><p>Restrict access to authorized components and limit its scope and duration to what an approved task requires. Task approval does not replace authorization checks for individual privileged operations.</p><p>Recommended controls include:</p><ul><li><strong>Secret isolation:</strong> Exclude credentials, tokens and private key material from prompts, memory stores, summaries, embeddings and general-purpose logs.</li><li><strong>Least privilege:</strong> Limit tools and runtime components to the resources and operations required for the task.</li><li><strong>Just-in-time access:</strong> Grant access when needed and revoke it when the operation or task ends.</li><li><strong>Ephemeral credentials:</strong> Use narrowly scoped credentials with automatic expiration and revocation support.</li><li><strong>Audit logging:</strong> Record privileged operations and authorization outcomes without exposing secrets.</li><li><strong>Execution-time validation:</strong> Verify authorization, target resources and permitted effects before file-system, infrastructure or configuration-management actions execute.</li></ul><p>Apply access restrictions, data minimization and retention limits to sensitive data even when it does not grant privileges.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/1*PN1N-7sjY7nWbEzpd3hXMQ.png" /><figcaption>High-Privilege Memory Vector Matrix</figcaption></figure><p>Compromise of an agent must not automatically grant access to privileged memory, administrative capabilities, infrastructure controls or long-lived credentials. Enforce authorization outside the reasoning layer to contain failures and prevent unauthorized privileged actions.</p><h3>RAG and Vector-Memory Design Controls</h3><p>RAG systems, vector databases and shared memory stores can expose information across access boundaries and propagate poisoned content into later reasoning and tool use. Risks include cross-session contamination, cross-agent leakage, unauthorized sharing and misuse of privileged information.</p><p>Attackers may gradually inject content designed to rank highly for targeted queries or create misleading semantic associations. When retrieved, this content can distort decisions, reinforce false information and contaminate downstream agents that use the same store.</p><p>Implement the following controls:</p><ul><li><strong>Scope isolation:</strong> Separate memory by user, tenant, session, agent and application. Permit sharing only through explicit access policies.</li><li><strong>Record ownership:</strong> Bind ownership, tenant identifiers and access-policy metadata to every persistent record.</li><li><strong>Retrieval enforcement:</strong> Apply access restrictions using authenticated caller context before records enter model-visible context.</li><li><strong>Least-privilege sharing:</strong> Limit each agent’s access to the records and operations required for its task.</li><li><strong>Poisoning detection:</strong> Monitor anomalous clusters, unexpected semantic associations and repeated retrieval of unrelated content.</li><li><strong>Containment and recovery:</strong> Quarantine compromised records and affected derivatives. Support removal and rollback to verified versions.</li><li><strong>Index maintenance:</strong> Review source content for staleness, compromise and uncertain provenance. Update affected indexes after corrections or deletions and verify that excluded records are no longer retrievable. Re-indexing alone does not detect or remove poisoning.</li></ul><h3>Context Integrity</h3><p>Context integrity preserves the origin and authority of system instructions, user requests, conversation history, retrieved memory, tool outputs and agent messages during prompt assembly. Content must not gain instructional authority merely because it appears in the active context.</p><p>Implement the following controls:</p><ul><li><strong>Instruction separation:</strong> Keep governing instructions distinct from external content. Distinguish the user’s request from instructions embedded in supplied documents, retrieved records or tool responses.</li><li><strong>Structured inputs:</strong> Validate tool-output schemas, types and size limits before including results in context.</li><li><strong>Protected control fields:</strong> Prevent untrusted data from directly changing system prompts, policies or orchestration parameters. Independently validate and authorize proposed plan changes before execution.</li><li><strong>Content validation:</strong> Check provenance, access scope and task relevance before assembly. Apply sanitization where appropriate and isolate or exclude suspicious content.</li><li><strong>Manipulation detection:</strong> Monitor prompt-injection indicators, instruction conflicts and unexpected changes to protected fields or context sources.</li><li><strong>Assembly records:</strong> Log source identifiers, trust classifications and assembly decisions to support investigation without unnecessarily retaining sensitive content.</li></ul><p>Formatting and sanitization cannot guarantee that a model will ignore malicious instructions. Connect assembly checks to runtime monitoring and enforce action permissions independently of model output.</p><h3>Tool Execution Control</h3><p>Mediate tool execution through controls that validate requests and enforce authorization before side effects occur. A manipulated invocation can extend prompt injection or memory poisoning into external systems, sensitive resources and persistent state.</p><p>Implement the following controls:</p><ul><li><strong>Least privilege:</strong> Restrict each tool and integration to the operations and resources required for its function.</li><li><strong>Tool-definition integrity:</strong> Verify tool definitions and schemas against approved versions and control subsequent changes.</li><li><strong>Input and output validation:</strong> Check arguments, target resources and operation constraints before execution. Validate returned content before using it in context or memory.</li><li><strong>Execution-time authorization:</strong> Verify permission for each requested operation and target. Require additional approval where policy specifies.</li><li><strong>Credential containment:</strong> Use just-in-time access and short-lived, narrowly scoped credentials. Keep secrets outside model-visible context.</li><li><strong>Side-effect containment:</strong> Apply sandboxing and limits on file access, network destinations, execution duration and transaction scope.</li><li><strong>Audit and monitoring:</strong> Record requesting identities, authorization decisions, invocations, state changes and outcomes. Investigate abnormal usage, excessive permissions and unexpected execution behavior.</li></ul><p>Tool-mediated memory writes, state changes and context updates must meet the same validation, authorization, provenance and audit requirements as direct operations. Connect these records to runtime monitoring and incident response to support detection, investigation and containment.</p><h3>Additional Safeguards for Memory, Context and Tool Security</h3><p>The following additions extend the existing guidance with explicit controls for retrieval authorization, derived memory, recovery and tool-definition integrity. They are proposed implementation requirements informed by OWASP guidance, NIST publications and recent research. These safeguards extend the preceding controls to retrieval authorization, derived state, recovery and tool-definition integrity.</p><h3>Secure Memory Read Path</h3><p>A valid signature does not authorize retrieval into every future context. Before releasing memory into prompt assembly, verify current permissions, ownership, permitted sharing scope, expiration status and provenance.</p><ul><li>Resolve user, tenant, session and agent identifiers from authenticated runtime context.</li><li>Apply access filters during retrieval and verify authorization before returning records.</li><li>Exclude expired records and block or quarantine records whose provenance or trust classification does not satisfy task policy.</li><li>Recheck permissions for cached results so revoked access cannot survive through previously accepted state.</li><li>Preserve retrieved content as data. Successful validation does not authorize it to override governing instructions.</li><li>Log admission, down-ranking and blocking decisions with record identifiers and policy reasons.</li></ul><p>These controls extend the retrieval filtering, access inheritance and cache protections described in the <a href="https://cheatsheetseries.owasp.org/cheatsheets/RAG_Security_Cheat_Sheet.html">OWASP RAG Security Cheat Sheet</a>.</p><h3>Explicit Isolation Rules</h3><p>Document sharing policies across five axes: user, session, agent, tenant and tool domain. For each boundary, specify permitted sharing, prohibited transfers, enforcement points and failure handling.</p><p>Enforce these rules during writes, retrieval, caching and tool execution. Allow cross-boundary access only through explicit authorization. Missing or inconsistent scope must result in denied access rather than an unrestricted fallback.</p><p>Assign ownership and trust classifications through trusted runtime components. A record’s claims about its own origin or authority must not determine its access rights. This applies OWASP’s deny-by-default and per-request authorization principles to agent memory. <a href="https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html">OWASP Authorization Cheat Sheet</a></p><h3>Summarization and Compaction as Memory Mutations</h3><p>Summarization, embedding and compaction create derived state and must pass the validated write path.</p><ul><li>Preserve source record identifiers, versions and provenance.</li><li>Retain distinctions between source trust classifications. Where this is not possible, apply the most restrictive applicable handling rules.</li><li>Prevent transformation alone from increasing trust or instructional authority.</li><li>Record the transformation identity, inputs, outputs and validation outcome.</li><li>Propagate corrections, quarantines and deletions to affected derivatives.</li></ul><p>These requirements operationalize provenance tracking across transformations. NIST identifies content origin, modification history and source metadata as inputs to traceability and risk management. <a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf">NIST AI 600–1, Appendix A.1.6</a></p><h3>Verified Recovery and Forensics</h3><p>Recovery must restore an assessed recovery point and support reconstruction of the contamination path. A signed snapshot proves neither that its content is safe nor that it predates compromise.</p><ul><li>Maintain tamper-resistant, append-only audit records for creation, modification, retrieval, transformation, rollback and deletion.</li><li>Preserve versioned snapshots and sufficient lineage to restore or regenerate derived summaries and embeddings.</li><li>Quarantine suspected records without destroying investigation evidence.</li><li>Assess recovery points for contamination before restoration.</li><li>Regenerate affected derivatives from verified sources.</li><li>Correlate record and version identifiers with session, agent, tool and transaction identifiers.</li><li>Verify that compromised records and stale derivatives cannot re-enter active retrieval.</li></ul><p>Exercise recovery procedures before incidents and update them after failures. This adapts NIST’s recovery planning and testing guidance to persistent agent state. <a href="https://csrc.nist.gov/pubs/sp/800/184/final">NIST SP 800–184</a></p><h3>Content Excluded from Ordinary Agent Memory</h3><p>Keep credentials, private keys, session tokens, raw secrets and capability-bearing handles outside ordinary agent memory. Store them in protected services and expose only authorized operations using narrowly scoped access with expiration and revocation support. <a href="https://cheatsheetseries.owasp.org/cheatsheets/Secrets_Management_Cheat_Sheet.html">OWASP Secrets Management Cheat Sheet</a></p><p>Store sensitive personal data only when its purpose, classification, access restrictions and retention requirements permit it.</p><p>Tool descriptions and agent messages may be retained as reference data where policy allows. Retrieval must not convert them into executable definitions or governing instructions.</p><h3>Tool-Schema Integrity</h3><p>Keep the authoritative tool catalog outside agent-writable memory. Protect definitions through signatures or an equivalently authenticated configuration mechanism.</p><ul><li>Version definitions and bind them to tool identities, endpoints, input schemas and permission scopes.</li><li>Verify integrity, approval status and current validity before use.</li><li>Reject substituted, revoked or unauthorized older definitions.</li><li>Route changes through controlled configuration management with independent authorization.</li><li>Allow agents to propose definitions without granting authority to approve or activate privileged tools.</li></ul><p>Tool annotations are descriptive hints and may misrepresent actual behavior. Enforcement must remain in authorization, sandboxing and resource controls. The MCP project explicitly distinguishes annotations from security guarantees. <a href="https://blog.modelcontextprotocol.io/posts/2026-03-16-tool-annotations/">MCP: Tool Annotations as Risk Vocabulary</a></p><h3>Execution Binding and Failure Handling</h3><p>Bind authorization to the exact tool version, arguments, target and relevant state version. Revalidate when these change before execution. An earlier approval must not authorize a materially different action.</p><p>Define handling for timeouts and uncertain outcomes. Record transaction status and use idempotency controls to prevent duplicate side effects. Reconcile an uncertain result before retrying a consequential operation. Apply compensating actions where supported.</p><p>If required authorization or integrity verification is unavailable, block the protected operation. Do not silently bypass the check. This follows OWASP’s fail-closed approach for retrieval pipelines and its guidance on parameter-bound approval for high-impact agent actions. <a href="https://cheatsheetseries.owasp.org/cheatsheets/RAG_Security_Cheat_Sheet.html">OWASP RAG Security</a>, <a href="https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html">OWASP AI Agent Security</a></p><h3>Detection Signals and Acceptance Tests</h3><p>Monitor for:</p><ul><li>Replayed nonces or duplicate transaction identifiers with conflicting payloads.</li><li>Retrieval attempts outside authorized sharing scopes.</li><li>Unexpected semantic associations with privileged operations.</li><li>Derived records with elevated trust or missing lineage.</li><li>Tool-mediated writes that bypass the validated gateway.</li><li>Credentials or privileged handles usable beyond their authorized lifetime.</li></ul><p>Maintain tests for delayed instructions in summaries, unauthorized cross-tenant retrieval, provenance loss during compaction, revoked access to cached records and tool definitions supplied through memory. Include verification-service failures and tool timeouts with uncertain outcomes.</p><p>For each test, define the expected denial, quarantine or recovery action and verify the audit evidence. Repeat relevant tests after changes to models, prompts, tools, retrieval or policy. OWASP recommends repeatable abuse cases and retained validation evidence for these changes. <a href="https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html">OWASP AI Agent Security Cheat Sheet</a></p><h3>Residual Risk</h3><p>Authorized writes may still contain plausible malicious content. Cryptographic verification does not establish intent, factual accuracy or semantic safety.</p><p>Apply independent policy checks before stored content can influence high-impact actions. Require human approval where policy specifies and bind that approval to the proposed action and scope. Approval must not grant blanket trust to a memory record.</p><p>OWASP’s May 2026 analysis of persistent prompt injection reinforces why retained memory and configuration require security boundaries beyond initial input acceptance. <a href="https://genai.owasp.org/2026/05/13/memory-is-a-feature-it-is-also-an-attack-surface/">OWASP: Memory Is a Feature. It Is Also an Attack Surface</a></p><h3>Recent Research Supporting These Additions</h3><p>Two recent preprints provide useful supporting evidence. They should be cited as research rather than established standards:</p><ul><li><strong>When Agents Remember Too Much</strong>, submitted July 6, 2026, examines separate memory-injection and later activation phases. Its proposed defense combines a memory-saving policy with retrieval screening, directly supporting controls on both paths. <a href="https://arxiv.org/abs/2607.06595">Torres et al., 2026</a></li><li><strong>MemSentry</strong>, submitted September 8, 2026, proposes memory-write decisions based on source trust, semantic risk, dependencies and access risk. Its evaluation uses synthetic scenarios, so its results should not be presented as demonstrated production effectiveness. <a href="https://arxiv.org/abs/2609.08747">Roy and Basu, 2026</a></li></ul><h3>References:</h3><p>OWASP GenAI Security Project. (2026). <em>OWASP Top 10 for Agentic Applications 2026</em>. Agentic Security Initiative, Version 2026.</p><p>OWASP GenAI Security Project. (2025). <em>OWASP Top 10 for LLM Applications 2025</em>. Version 2025.</p><p>OWASP GenAI Security Project. (2025). <em>LLM01: Prompt Injection</em>. OWASP Top 10 for LLM Applications.</p><p>OWASP GenAI Security Project. (2025). <em>LLM06: Excessive Agency</em>. OWASP Top 10 for LLM Applications.</p><p>OWASP GenAI Security Project. (2025). <em>LLM08: Vector and Embedding Weaknesses</em>. OWASP Top 10 for LLM Applications.</p><p>Schneider, C. (2026, February 26). <em>From Session Attacks to Persistent Compromise: Memory Poisoning in AI Agents</em>.</p><h3>Additional References:</h3><p><strong>OWASP Cheat Sheet Series.</strong> <a href="https://cheatsheetseries.owasp.org/cheatsheets/RAG_Security_Cheat_Sheet.html">RAG Security Cheat Sheet</a>.</p><p><strong>OWASP Cheat Sheet Series.</strong> <a href="https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html">Authorization Cheat Sheet</a>.</p><p><strong>OWASP Cheat Sheet Series.</strong> <a href="https://cheatsheetseries.owasp.org/cheatsheets/Secrets_Management_Cheat_Sheet.html">Secrets Management Cheat Sheet</a>.</p><p><strong>OWASP Cheat Sheet Series.</strong> <a href="https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html">AI Agent Security Cheat Sheet</a>.</p><p><strong>Autio et al. (2024).</strong> <a href="https://doi.org/10.6028/NIST.AI.600-1">Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile</a>. NIST AI 600–1. See Appendix A.1.6 for content provenance.</p><p><strong>Bartock et al. (2016).</strong> <a href="https://doi.org/10.6028/NIST.SP.800-184">Guide for Cybersecurity Event Recovery</a>. NIST SP 800–184.</p><p><strong>Model Context Protocol. (2026, March 16).</strong> <a href="https://blog.modelcontextprotocol.io/posts/2026-03-16-tool-annotations/">Tool Annotations as Risk Vocabulary: What Hints Can and Can’t Do</a>.</p><p><strong>Habler, I. (2026, May 13).</strong> <a href="https://genai.owasp.org/2026/05/13/memory-is-a-feature-it-is-also-an-attack-surface/">Memory Is a Feature. It Is Also an Attack Surface</a>. OWASP Gen AI Security Project.</p><p><strong>Torres, G., Shrestha, S. and Misra, S. (2026, July 6).</strong> <a href="https://arxiv.org/abs/2607.06595">When Agents Remember Too Much: Memory Poisoning Attacks on Large Language Model Agents</a>. arXiv preprint.</p><p><strong>Roy, A. and Basu, K. (2026, September 8).</strong> <a href="https://arxiv.org/abs/2609.08747">MemSentry: A Framework for Detecting Persistent Memory Poisoning in Agentic AI</a>. arXiv preprint.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f1ef2dcc05f2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Threat Modeling with Adam Shostack]]></title>
            <link>https://medium.com/@angelska327/threat-modeling-with-adam-shostack-200eec3b2693?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/200eec3b2693</guid>
            <category><![CDATA[stride]]></category>
            <category><![CDATA[threat-modeling]]></category>
            <category><![CDATA[model]]></category>
            <category><![CDATA[assumptions]]></category>
            <category><![CDATA[cyber-kill-chain]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Wed, 31 Dec 2025 03:19:55 GMT</pubDate>
            <atom:updated>2025-12-31T05:04:28.928Z</atom:updated>
            <content:encoded><![CDATA[<p>At the beginning of December 2025, I had the opportunity to participate in Adam Shostack’s <strong>two-day threat modeling workshop</strong>, working alongside Rick Toledo and David Lauro as part of the same team.</p><p>Adam Shostack is a recognized authority in threat modeling, a long-time security practitioner and instructor, a Black Hat board member, and the author of <em>Threat Modeling: Designing for Security</em>, a widely used reference in modern security engineering.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qhEFkAM75qzzngwu9_j1pQ.jpeg" /><figcaption>My awesome team and our instructor, Adam.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*M3BspOTSWb2YLXwmoAB2KQ.png" /><figcaption>The four core questions defined by Adam Shostack</figcaption></figure><p>According to Adam Shostack’s<strong> threat modeling framework</strong>, the first question to answer is:</p><h3><strong>1. What are we working on?</strong></h3><p>This question matters because threat modeling is only as effective as the <strong>accuracy </strong>of the system being analyzed. Without a clear and shared understanding of system components, data flows, and trust boundaries, any discussion of threats becomes speculative rather than actionable.</p><h3>What my team and I did:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iIi-xQPMHpru1RB6WpxkDA.jpeg" /><figcaption>First time drawing DFD</figcaption></figure><p>During the workshop, this question was answered by collaboratively constructing a <strong>Data Flow Diagram (DFD)</strong> for the Grab-a-Bike Bikes-as-a-Service (BaaS) platform. The objective of this exercise was to establish a shared, explicit understanding of how the system actually works. The diagram translated a narrative system description into concrete components, data flows, and <strong>trust boundaries</strong> that the entire group could see, discuss, and challenge.</p><p>The resulting DFD organized the system into three primary domains: the <strong>user-facing mobile application</strong>, the <strong>Bikes-as-a-Service Cloud Service (BaaSCS)</strong>, and the <strong>in-bike computer</strong>. These domains were deliberately separated by trust boundaries to indicate where control, responsibility, and security assumptions change. Identifying these boundaries early was essential, because many meaningful threats emerge when data or execution crosses from<strong> one trust context to another.</strong></p><p>A key assumption made explicit during modeling was that the application runs on <strong>user-controlled devices</strong> and therefore exists <strong>outside the system’s trust boundary</strong>. As a result, all data originating from the app, including user identifiers, bike identifiers, images, and location requests, must be treated as <strong>untrusted input</strong> until validated by backend services. This assumption prevents implicit trust in<strong> client-side</strong> <strong>logic </strong>and directly shapes later threat analysis.</p><p>At the center of the diagram, the <strong>BaaS Cloud Service (BaaSCS)</strong> functions as the system’s primary authority. Exposed through an API gateway and implemented using serverless components, it handles user registration, payment setup, bike discovery, unlock authorization, billing, and state management. A critical assumption documented during the exercise was that <strong>all authorization decisions are centralized in the cloud service</strong>. The bike itself does not independently determine whether it may unlock; it only executes authenticated commands received from BaaSCS. This assumption establishes the cloud service as a high-value trust anchor and clarifies where control truly resides.</p><p>We explicitly assumed that the bike can unlock only after it receives a valid command from the cloud. The bike sends <strong>telemetry </strong>such as location, lock status, and ride state to the cloud, while the cloud sends control commands back to the bike. Making these directions explicit clarified that the bike has the ability to unlock, but the cloud has the authority to decide when unlocking is allowed, highlighting the<strong> trust boundary</strong> between the physical device and cloud-controlled <strong>logic</strong>.</p><p>The diagram also included external dependencies such as payment processors, mapping services, GPS infrastructure, and app distribution platforms. These were modeled as external entities outside the system’s trust boundary. We assumed that these services behave correctly according to their contracts, while acknowledging that their availability and integrity <strong>are not</strong> under direct control of the BaaS system. This assumption intentionally constrained the scope of analysis while making dependency risk explicit rather than implicit.</p><p><em>Terms:</em></p><p><strong>DFD (Data Flow Diagram):</strong> a simple diagram that shows <strong>how data moves through a system</strong>, including users, processes, data stores, and external services.<br><strong>Why it matters:</strong> you can’t identify real security risks unless you clearly understand how the system actually works.</p><p><strong>Trust boundaries:</strong> lines in a system where <strong>trust, control, or security assumptions change</strong> (for example, you fully control your backend, but you do <strong>not</strong> control a user’s phone, browser, or the internet).<br><strong>Why they matter:</strong> most serious threats happen <strong>at or across trust boundaries</strong>, where untrusted data or actions enter more trusted parts of the system.</p><p>In Adam Shostack’s <strong>threat modeling framework</strong>, the second question to be addressed is:</p><h3>2. What can go wrong?</h3><p>This question matters because understanding a system is not enough on its own. Once the architecture, data flows, trust boundaries, and initial assumptions are defined, teams must deliberately explore how that system could be abused. Without this step, security analysis remains <strong>descriptive </strong>rather than analytical.</p><p>Shostack emphasizes that threat modeling is about <strong>anticipating misuse</strong> and adversarial behavior while the system is still <strong>malleable</strong>. Asking<strong> <em>what can go wrong</em></strong> forces teams to reason from an attacker’s perspective and to test whether the assumptions documented earlier can be violated in realistic ways.</p><p>Importantly, this question shifts the focus from <strong>accidental failure to intentional abuse</strong>. Systems rarely fail only because of bugs; they fail because<strong> design choices interact with attacker incentives</strong>.</p><h4><em>Threat modeling makes those interactions explicit.</em></h4><p>STRIDE <strong>purpose </strong>is to ensure disciplined coverage across common threat categories so that teams do not focus only on familiar, recent, or sensational attack types.</p><p>In Shostack’s methodology, STRIDE acts as a thinking aid that prompts systematic questioning of each element in the Data Flow Diagram, grounded in the system’s actual architecture.</p><h3>What my team and I did:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jyJiFFInB_2d_WyCAjY66w.jpeg" /><figcaption>STRIDE done by us</figcaption></figure><p>After completing the Data Flow Diagram and agreeing on system boundaries and assumptions, we applied the <strong>STRIDE </strong>framework directly to the Grab-a-Bike Bikes-as-a-Service (BaaS) architecture. Each STRIDE category was examined in the context of specific components, data flows, and trust boundaries captured on the whiteboard. The goal was to surface realistic abuse cases that followed naturally from the system design.</p><h4>STRIDE</h4><p>S For <strong>spoofing</strong>, the team explored how attackers could <strong>impersonate</strong> legitimate users or system components. Examples included phishing users to obtain credentials, placing fake QR codes on physical bikes to redirect users to malicious endpoints, and socially engineering cloud administrators to gain privileged access. These scenarios directly challenged assumptions about identity verification, trust in physical artifacts, and administrative access controls.</p><p>T For <strong>tampering</strong>, the discussion focused on unauthorized modification of system state or device behavior. Identified risks included overriding the bike lock mechanism, falsifying GPS data, spoofing bike or user identifiers, and manipulating Bluetooth communications between the bike and external interfaces. These threats tested assumptions about the integrity of embedded systems and the trustworthiness of communication channels, reinforcing the importance of clearly defining which components are authoritative for state changes.</p><p><strong>R Repudiation</strong> threats addressed situations where actions could be denied or falsely attributed. My team examined scenarios such as fraudulent ride activity, disputed credit card charges, and false damage reports submitted through the mobile application. These discussions highlighted assumptions around logging, auditability, and non-repudiation, making it clear that insufficient evidence trails create operational and financial risk even in the absence of technical compromise.</p><p>I For <strong>information disclosure</strong>, the analysis focused on unauthorized access to sensitive data, including credit card information, personally identifiable information (PII), and backend APIs. These threats were tied directly to assumptions about access control, data minimization, and secure transport.</p><p><strong>D Denial of service</strong> threats examined how system availability could be disrupted across physical and network layers. Examples included cellular signal jamming, excessive API requests, hardware overload, and dependency outages. These scenarios challenged assumptions about network reliability, infrastructure capacity, and third-party service resilience.</p><p>E Finally, <strong>elevation of privilege</strong> threats explored how attackers might gain higher levels of access than intended. The team identified risks involving compromised cloud administrator accounts, local privilege escalation on the bike’s embedded Linux system, and abuse of command-and-control pathways. These threats emphasized that high-privilege paths represent concentrated risk and must be protected disproportionately.</p><h4>ASSUMPTIONS</h4><p>Key assumptions made explicit during this phase included:</p><ul><li>The user is authentic once authentication succeeds</li><li>Third-party vendors behave securely according to their contracts</li><li>App stores distribute legitimate versions of the application</li><li>Firewalls and WAFs function as intended</li><li>The cloud provider and PaaS environment are secure and highly available</li><li>Payment providers are partially trusted and subject to regular audits</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DgYM8fMGlArpicRDHkcoog.jpeg" /><figcaption>Our assumptions</figcaption></figure><p>Writing down these assumptions ensured that identified threats were grounded in real architectural dependencies. Threat modeling focuses on making<strong> underlying beliefs visible, examinable, and revisable</strong>, which helps teams understand how design decisions and trust boundaries could be abused under realistic attacker behavior.</p><p><em>Terms:</em></p><p><strong>Stride</strong> is a mnemonic that represents six high-level classes of threats:</p><ul><li><strong>Spoofing</strong> — impersonating users or system components</li><li><strong>Tampering</strong> — modifying data or system state without authorization</li><li><strong>Repudiation</strong> — denying actions due to insufficient logging or evidence</li><li><strong>Information Disclosure</strong> — exposing sensitive data to unauthorized parties</li><li><strong>Denial of Service</strong> — degrading or preventing system availability</li><li><strong>Elevation of Privilege</strong> — gaining higher permissions than intended</li></ul><p><strong><em>Why it matters:</em></strong></p><p>Using STRIDE turns assumptions into explicit points of risk.</p><p>An <strong>assumption </strong>is a statement the system relies on being <strong>true </strong>for security to hold. If an assumption is <strong>false</strong>, the security controls built on top of it may fail silently. Rather than treating assumptions as background context, the workshop required them to be <strong>documented, examined, and challenged.</strong></p><p><strong><em>Why assumptions matte</em></strong>r:</p><p><strong>Each assumption represents a potential failure point</strong>.</p><p>Assumptions: clarify where trust is placed and why, constrain the scope of analysis, expose hidden dependencies and control reliance, and allow underlying beliefs to be <strong>challenged</strong> and <strong>revisited </strong>as systems evolve.</p><p><em>Example</em>: Hidden belief: <em>“The bike can only unlock when the cloud says so.” </em>If that belief is wrong or later changes, security breaks <strong>silently</strong>. Writing it down lets the team say:<em>“Is this still true? What if it isn’t?”</em></p><p>Adam Shostack emphasizes that assumptions are <strong>unavoidable in any system</strong>;<strong> every design depends on them</strong>. The risk comes from leaving assumptions implicit and unexamined. When they are undocumented, security decisions may rest on fragile or incorrect foundations without the team realizing it. In Shostack’s methodology, assumptions are treated as <strong>first-class inputs </strong>to the threat model. They actively shape which threats are <strong>considered plausible</strong>, which attack paths are worth analyzing, and where defensive effort is justified.</p><h3>Kill Chain Analysis</h3><h3>How could an attacker realistically succeed?</h3><p>In Adam Shostack’s threat modeling framework, kill chain analysis helps answer a refinement of the second core question: <em>What can go wrong?</em></p><p>The kill chain asks<strong> how an attacker could progress through the system over time</strong>, combining multiple weaknesses into a coherent attack.</p><h4><strong><em>Why this question matters</em></strong></h4><p>Individual threats rarely cause real-world failures on their own. Most successful attacks unfold as <strong>sequences</strong>, where one capability enables the next. Without understanding these sequences, teams risk <strong>over-investing</strong> in controls that do not meaningfully slow or<strong> stop attacker progress</strong>, while under-protecting paths that lead directly to high-impact outcomes.</p><h3>What my team and I did:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YiaGl22MvEpctF4VXiHkrQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fQ2nLg5oIGy60eOYMKBMuw.jpeg" /></figure><p><strong>Reconnaissance (implicit)</strong><br>The chain began with a <strong>valid user account</strong>, intentionally reflecting an earlier assumption: <strong>authentication does not imply benign intent</strong>. The attacker does not need to bypass login controls; instead, they operate within <strong>expected system functionality</strong>.</p><p><strong>Weaponization and Delivery</strong><br>Using this access, the attacker submits a <strong>malicious SVG file</strong> through the <strong>issue reporting feature</strong>. This step represents delivery and weaponization combined, exploiting the assumption that uploaded images are well formed and safe. The SVG format was chosen deliberately because it can contain executable behavior while appearing innocuous (for example, <strong>embedded JavaScript or malicious event handlers</strong>).</p><p><strong>Exploitation</strong><br>Next, the attacker exploits an <strong>SVG parsing vulnerability</strong> that allows <strong>embedded JavaScript execution</strong>. This moves the chain into the exploitation phase, where untrusted input is interpreted in a privileged context. From there, the attacker leverages script or command execution paths to call functions<strong> not intended to be exposed</strong>.</p><p><strong>Installation and Command and Control</strong><br>With execution established, the attacker proceeds to <strong>copy images and sensitive data</strong> to an <strong>attacker-controlled server</strong>. This stage represents both installation and command and control, as the attacker gains repeated ability to extract data rather than performing a single action.</p><p><strong>Actions on Objectives</strong><br>The final stage, actions on objectives, is reached when the attacker successfully accesses stored images and credit card–related data. This ties directly back to earlier <strong>information disclosure threats</strong> and demonstrates how relatively <strong>small design decisions</strong> can combine into a <strong>high-impact outcome</strong>.</p><h3>How assumptions shaped the kill chain</h3><p>The chain depends on the assumption that authenticated users are trusted enough to upload files, that SVG images are accepted without strict sanitization, and that backend components will process uploaded content as expected. It also relies on the assumption that internal services trust data once it passes initial validation.</p><p>Kill chain analysis transformed a list of STRIDE <strong>threats into a story</strong> about attacker behavior grounded in the actual architecture and assumptions of the system. The analysis showed where attacker progress could be interrupted most effectively. This made subsequent mitigation discussions more focused, more defensible, and more aligned with real risk. In this way, kill chains served as the bridge between threat identification and informed security decision-making.</p><p><em>Example:</em></p><ul><li>If the attack depends on <strong>file upload</strong>, then strict file validation could stop it or if it depends on <strong>script execution</strong>, then sandboxing or input sanitization could stop it.</li></ul><p><em>Term:</em></p><p><strong>Kill chain-</strong> is a structured way of describing attacker behavior across stages, typically including<strong> reconnaissance, weaponization, delivery, exploitation, installation, command and control, and actions on objectives. </strong>The purpose is to understand <strong>how threats interact over time</strong> and where defenses can interrupt attacker momentum.</p><p><em>Why it matters:</em><br>Attackers do not act in isolated steps. A kill chain shows how individual threats depend on earlier successes and unfold in a specific order over time.</p><p><em>As a reminder:</em></p><h4>STRIDE identifies <em>what kinds</em> of threats exist;</h4><h4>Kill chains explain <em>how those threats are combined</em> to achieve attacker goals.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aKmYcC5_oa1Ieis3PrLZ4w.jpeg" /><figcaption>Tour at MxD</figcaption></figure><p>The third question in Adam Shostack’s<strong> threat modeling framework</strong> is:</p><h3><strong>3. What are we going to do about it?</strong></h3><p><strong>Why this question matter?</strong></p><p>Threat modeling only becomes actionable when teams decide <strong>how to respond</strong>, <strong>what to prioritize</strong>, and <strong>what risks will be consciously accepted</strong>. This question forces the transition from analysis to action by requiring teams to decide which risks to <strong>mitigate</strong>, which to <strong>reduce</strong>, which to <strong>transfer</strong>, and which to explicitly <strong>accept</strong>. Without this step, threat modeling remains descriptive rather than operational, producing insight without <strong>impact</strong>.</p><p>In Adam Shostack’s approach, this question is about making <strong>deliberate, documented trade-offs</strong> based on system assumptions, attacker behavior, and practical constraints. The goal is <strong>informed decision-making that aligns defenses with real risk.</strong></p><h3>What my team and I did:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nesfD1u6elmcdV6CGUpang.jpeg" /><figcaption>Our mitigations and risk treatment decisions</figcaption></figure><p>For spoofing threats, particularly phishing and impersonation risks, we applied <strong>risk reduction</strong> by identifying user training as a necessary mitigation, acknowledging that technical controls alone cannot fully prevent social engineering. This was complemented by <strong>risk reduction</strong> through email authentication mechanisms such as DMARC and DKIM to reduce the effectiveness of spoofed messages.</p><p>In response to credit card–related risks, we chose <strong>risk transfer</strong> by explicitly assigning responsibility for securing credit card data to the payment processor. This decision reduced the system’s exposure to sensitive financial information and established a clear trust boundary by <strong>delegating</strong> financial data protection to a specialized third party.</p><p>For physical tampering and hardware risks, including lock manipulation and GPS interference, we accepted that full prevention was not feasible. The chosen approach combined <strong>risk acceptance</strong> with <strong>risk transfer</strong>, using insurance coverage for lock and GPS failures and further analysis of physical hardware access.</p><p>To address repudiation threats such as false issue reporting or denial of actions, my team applied <strong>risk reduction</strong> through stronger identity validation and behavioral controls, including flagging users who repeatedly submit false reports to improve accountability.</p><p>For information disclosure risks, we applied <strong>risk reduction</strong> by identifying encryption as a baseline mitigation to protect sensitive data both in transit and at rest. While encryption does not eliminate disclosure risk entirely, it significantly reduces impact and aligns with the assumption that perimeter defenses may eventually be bypassed.</p><p>Denial of service threats were addressed through <strong>risk reduction</strong> by implementing infrastructure-level mitigations such as firewall strengthening, cellular service redundancy, and load balancing.</p><p>Finally, elevation of privilege risks, particularly those involving administrative access, were addressed through <strong>risk reduction</strong> by enforcing multi-factor authentication requirements, password rotation policies, and encryption. These mitigations reflect the understanding that high-privilege paths represent concentrated risk and therefore require stronger, layered protections.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RHIZAHGUXvGzgO1Kvq_PJw.jpeg" /><figcaption>Our class</figcaption></figure><h3>Why mitigations matter in threat modeling</h3><p>This phase demonstrated that threat modeling is about <strong>owning </strong>the risk. By tying mitigations directly to documented threats and assumptions, my team transformed abstract analysis into actionable decisions.<strong> Some risks were reduced, some were transferred, some were mitigated </strong>through non-technical means, and some were <strong>consciously accepted</strong>. Threat modeling provided the structure to make those choices<strong> explicit, reviewable, and defensible.</strong></p><p>In this way, the question <em>“What are we going to do about it?”</em> serves as the bridge between understanding threats and building resilient systems.</p><p>It ensures that security effort is focused where it produces the greatest reduction in meaningful risk, rather than where it is easiest to add controls.</p><p><em>Terms:</em></p><p><strong>DKIM</strong> proves that an email really came from the domain it claims to come from and that it was not changed in transit.</p><p><strong>DMARC</strong> tells email providers what to do if an email fails DKIM or SPF checks, such as reject it or mark it as spam.</p><p><strong><em>Why they matter:</em></strong><br> They <strong>reduce </strong>email spoofing and phishing by making it harder for attackers to send fake emails that appear to come from a trusted organization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/376/1*jdf3pXI4i9-EhpBV8UgEGw.png" /><figcaption>The team</figcaption></figure><h4>FLYWHEEL CONCEPT (discussed by our teammate David)</h4><p>Each step of threat modeling<strong> builds on the previous one</strong>. Starting with the Data Flow Diagram, then identifying assumptions, applying STRIDE, analyzing kill chains, and finally deciding on mitigations creates a continuous loop of understanding and improvement.<strong> Each phase strengthens the next,</strong> ensuring that threat modeling remains structured, cumulative, and grounded in real system behavior rather than isolated analysis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/633/1*hp7eds03yVR8Gwb1Fy3i0Q.png" /><figcaption>Threat Modeling Flywheel (Team-Derived Concept)</figcaption></figure><p>Lastly, according to Adam Shostack’s<strong> threat modeling framework</strong>, the fourth and final question to answer is:</p><h3>4. Did we do a good job?</h3><p><strong>Why this question matters in threat modeling</strong></p><p>Threat modeling is an<strong> iterative process</strong>. A retrospective helps teams <strong>evaluate </strong>not only technical outcomes, but also collaboration, clarity, and effectiveness. Without reflection, teams risk repeating the same gaps and missing opportunities to<strong> improve future threat modeling efforts.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/618/1*U_wTCvzSL1RSTC5vgy_d4Q.png" /><figcaption>Manufacturing × Digital</figcaption></figure><h3>What we liked about our teamwork</h3><p>The team worked in a highly collaborative and active manner. Everyone participated and contributed ideas, which helped surface different perspectives early. When we were unclear about what to write or how to proceed, we asked for help instead of making assumptions in isolation. This openness improved the quality of the discussion.</p><p>Our group was well structured in how ideas were gathered and organized. We listed components clearly, mapped services in detail, and ensured that ownership of different system areas was understood. This made the threat modeling process more efficient and reduced confusion.</p><h3>What we would change</h3><p>We would improve how data flows and operations were documented by adding a sequence to each process. This would make the Data Flow Diagram easier to follow and would better support later stages such as kill chain analysis.</p><p>We would also consider using a different approach to organizing diagrams and drawings to improve clarity, especially when transitioning from brainstorming to formal threat analysis.</p><h3>What we are going to do about it</h3><p>In future threat modeling sessions, we would more deliberately structure data flows from the beginning and explicitly tie them to attacker steps in the kill chain.</p><p>Finally, we would ensure that records of decisions, risks, and assumptions are preserved for future teammates. This allows future teams to understand not only <strong>what decisions were made</strong>, but <strong>why they were made</strong>, strengthening continuity and long-term security reasoning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t9i3qYig6ZSiqNr8ik5J6w.jpeg" /><figcaption>Another team’s presentation.</figcaption></figure><h3>Personal Opinion about 2-day workshop</h3><p>Beyond the technical depth of the material, the way the workshop was delivered had a significant impact on its effectiveness. <strong>Adam Shostack </strong>is not only a highly experienced security practitioner and instructor, but also an exceptionally clear communicator. As a non-native English speaker, I found his diction, pacing, and sentence structure particularly effective. His explanations were deliberate and well-organized, making complex concepts easy to follow without oversimplification.</p><p>Adam consistently chose precise language, paused at the right moments, and framed ideas in a way that supported understanding rather than intimidation. His tone was steady and engaging, never rushed or monotonous, which made it easy to remain focused throughout long, technically dense sessions. This clarity of expression materially improved my ability to absorb and apply the material in real time.</p><p>The combination of technical rigor and careful communication is rare in security education. Adam’s ability to teach how to think while expressing those ideas with clarity and discipline significantly strengthened the learning experience. His command of language is strong enough that, independent of his security expertise, he would make <strong>an excellent English instructor as well.</strong></p><p><strong>I enjoyed the whole two-day workshop and working within my team.</strong></p><p>Dear Reader, thank you for your time !</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/466/1*_HW6OKDRZVtTBNNrWE0BJg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=200eec3b2693" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Big Dogs — read if you’re one of them.]]></title>
            <link>https://medium.com/@angelska327/big-dogs-read-if-youre-one-of-them-a7c48528e018?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/a7c48528e018</guid>
            <category><![CDATA[reading]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[top-10]]></category>
            <category><![CDATA[social-impact]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Sat, 22 Nov 2025 22:19:37 GMT</pubDate>
            <atom:updated>2026-08-30T05:45:35.180Z</atom:updated>
            <content:encoded><![CDATA[<h3>Big Dogs -read if you’re one of them.</h3><p>True resilience is not bending yourself to be seen -it’s creating something so solid that visibility <strong>finds you</strong> on its own.</p><p>There’s a difference between resilience and performing for approval. Real growth isn’t about begging big names to see you -it’s about building something big enough that they can’t ignore you.</p><p>Big Dogs protect their energy from:</p><ul><li>low-value connections</li><li>draining dynamics</li><li>noise</li><li>chaotic people</li><li>emotional spirals</li><li>unnecessary obligations</li></ul><p>Instead they are:</p><p>• direct<br> • intelligent<br> • ambitious<br> • <strong>intense</strong><br> • a builder<br> • a leader<br> • a creator<br> • a force<br> • a visionary</p><p>A Big Dog formula:</p><p>Mastery + Building + Consistency + Confidence + Vision = Big Dog</p><p>When you try to <em>sell yourself</em> to famous, influential people, you end up in a position of <strong>seeking approval</strong>.<br> Your value becomes dependent on someone else saying “yes.”<br> That creates pressure, insecurity, and a feeling of being <em>smaller</em> than them.</p><p>But when you build something <strong>real</strong>, <strong>valuable</strong>, <strong>original</strong>, and <strong>strong</strong>, two things happen:</p><ol><li>Your work becomes th<strong>e magnet.</strong><br> You don’t have to pull people in-they are pulled in naturally.</li><li><strong>The power dynamic flips.</strong><br> You’re no longer asking for attention;<br> <strong>you’re earning it without effort</strong>, because what you built is meaningful.</li></ol><p>This is why the biggest founders, artists, engineers, and creators don’t beg to be noticed they create something that forces the world to pay attention.</p><p>If this what you are reading now resonates with you internally it means</p><h4>you recognized a truth that your body already understands:</h4><ol><li>You’ve spent years proving yourself.</li><li>You’ve worked hard to impress, to be seen, to be validated.</li><li>You’ve chased attention from leaders, mentors, companies, “big dogs.”</li><li>And you’ve outgrown that phase. (Bravo baby! You made it!)</li></ol><p>Your internal reaction came from realizing:</p><h4>You don’t need to be chosen anymore. You’re becoming the chooser. You are the big dog now!</h4><p>Your system recognized the shift. It’s the feeling of stepping into power, not performance.</p><p>Visibility finds you — means that <strong>your work becomes the magnet</strong> instead of <em>you</em> chasing recognition.</p><h3>1. Build something REAL, not performative</h3><p>This could be:</p><ul><li>your skills</li><li>your portfolio</li><li>an original project</li><li>your voice</li><li>your reputation</li><li>your consistency</li></ul><p>People don’t chase noise.<br>People chase substance what makes someone magnetic instead of just loud?</p><h3>2. Master one lane</h3><p>Not 10 things lightly- <strong>One thing deeply.</strong></p><p>When you master a lane, the world starts seeing you as:</p><ul><li>“the person who’s killing it in …”</li><li>“the founder building a new …”</li><li>“the person who leads …. with heart and intelligence”</li></ul><p>Clarity = visibility.</p><h3>3. Show your work</h3><p>Visibility comes from:</p><ul><li>writing</li><li>documenting</li><li>sharing progress</li><li>speaking</li><li>posting</li><li>teaching others</li><li>publishing your code</li><li><strong>shipping projects</strong>: launching a feature, deploying an app, posting a project, releasing a tool, submitting a product, making something <em>real and visible. </em>It’s the opposite of:</li></ul><ol><li>starting and abandoning</li><li>planning forever</li><li>keeping ideas in your notebook</li><li>waiting for perfection</li></ol><p>Shipping = finishing + delivering + making public.</p><p>People trust what they see repeatedly. This is not “performing.” This is<strong> producing evidence</strong>.</p><h3>4. Build reputation through consistency</h3><p>Not by: feigning, flattering, begging, asking for exposure,<strong> forcing connections,</strong></p><p>But by:</p><ul><li>showing up</li><li>delivering</li><li>improving</li><li>becoming dependable</li><li>staying in your lane long enough for momentum to build</li></ul><p>Reputation is earned in the dark before it’s seen in the light.</p><h3>5. Create value</h3><p>Ask yourself:<br> <strong>“What can I build that actually solves a problem?”</strong></p><p>It is a <strong>solution</strong>. People chase value.</p><h3>6. Grow your network authentically</h3><p>By:</p><ul><li>meeting</li><li>learning</li><li>exchanging ideas</li><li>building mutual respect</li></ul><p>High-powered people can smell “pick me” energy.</p><p>But they <strong>respect builders</strong>.</p><h3>So do you “need to create create create”?</h3><p><strong>No.</strong></p><p><strong>You need to build what matters, consistently, in your lane, with clarity, until the gravitational pull starts working in your favor.</strong></p><p>When you do that <strong>visibility becomes a natural</strong> side-effect, not the goal.</p><h3>Examples of what this looks like in real life:</h3><p>1. Recruiters find your projects and reach out before you apply. You didn’t chase them.<br> Your work pulled them in.</p><p>2. Senior people hear about you through someone else. Your name travels without you.</p><p>3. People read your posts, ideas, scripts, or projects — quietly. They already know you before you ever introduce yourself.</p><p>4. Someone says, “I saw your work on ___.” Not because you gave it to them because they found it.</p><p>5. You walk into a room and people recognize your contributions. Not your voice, pitch <strong>or impact.</strong></p><h3>Signs the gravitational pull has already started:</h3><ul><li>People invite you to speak.</li><li>You receive without begging.</li><li>Leaders respond to you seriously.</li><li>You keep getting noticed.</li><li>You’re building a real thing.</li><li>Your writing is getting sharper and deeper.</li><li>People remember your name.</li><li>You’re in rooms you never thought you’d enter.</li><li>High-level people talk to you <em>as an equal</em>.</li><li>People ask for your help.</li><li>Companies reach out.</li><li>Strong people follow your work.</li><li>Opportunities appear. Invitations come.</li><li>People talk about you highly in rooms you’re not in.</li></ul><h3><strong>SUBSTANCE</strong></h3><p><strong>What makes someone magnetic instead of just loud?</strong></p><h4>Substance = What You Build, What You Know, and Who You Are When No One Is Looking.</h4><h3>1. Substance = Skill = Respect.</h3><p>People respect: mastery, competence, expertise and depth.</p><h3>2. Substance = Results</h3><p>People don’t chase talkers. They chase <strong>doers</strong>. Results create gravity.</p><p>Results look like:</p><ul><li>shipped projects</li><li>real tools</li><li>certifications</li><li>solved incidents</li><li>measurable improvements</li><li>contributions to teams</li><li>things you can point at and say:<br> <strong>“I built this.”</strong></li></ul><h3>3. Substance = Character</h3><p>Character is rare. Rarity is valuable.</p><p>People follow people who are:</p><ul><li>consistent</li><li>honest</li><li>reliable</li><li>disciplined</li><li>grounded</li><li>self-aware</li><li>emotionally intelligent</li></ul><h3>4. Substance = Vision= Clarity.</h3><p>Vision is what makes people believe in you.</p><p>Not vague dreams.</p><ul><li>Where you’re going</li><li>What you’re building</li><li>Why you’re doing it</li><li>How it will change something</li></ul><h3>5. Substance = Presence</h3><p>A calm, stable energy. Not anxious, needy or performative.</p><p>Presence says: <strong>“I’m here because I belong, not because I need to be chosen.” </strong>People feel that energy instantly.</p><h3>6. Substance = Value Creation= Influence.</h3><p>If you want to become someone people chase: <strong>solve real problems.</strong></p><p>Value could be:</p><ul><li>a tool, framework, method, script, product, perspective, emotional support, leadership or clarity.</li></ul><h3>7. Substance = Identity</h3><p>Identity is:</p><ul><li>experience, story, struggle, resilience and/or transformation.</li></ul><p>A story must have<strong> gravity</strong>.</p><p>People chase stories that move them.</p><h4><strong>Substance means depth.</strong> Substance is unforgettable.</h4><h3>Why this may hit so deeply for you</h3><p>Because you’re done with: seeking approval, overperforming, chasing and proving.</p><p>You are entering the phase of: building, mastering, becoming and owning your lane. This is maturity, power and identity.</p><p><em>If this was valuable for you, and you learned something new about yourself — share this post with others.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/664/1*E5rC5b3-UvDL78h7TR3ukw.png" /><figcaption>Thanks for your time!</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a7c48528e018" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[One Million Degrees Panel]]></title>
            <link>https://medium.com/@angelska327/one-million-degrees-panel-547bc2f917b5?source=rss-3a644af46fc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/547bc2f917b5</guid>
            <category><![CDATA[internships]]></category>
            <category><![CDATA[students]]></category>
            <category><![CDATA[schools]]></category>
            <category><![CDATA[jim-rohn]]></category>
            <category><![CDATA[user-experience]]></category>
            <dc:creator><![CDATA[Angelska]]></dc:creator>
            <pubDate>Mon, 17 Nov 2025 06:47:27 GMT</pubDate>
            <atom:updated>2025-11-22T23:01:38.698Z</atom:updated>
            <content:encoded><![CDATA[<p>On Saturday November 15 I was invited to the One Million Degrees (OMD) panel at Wilbur Wright College in Chicago. Our<strong> topic </strong>was how community college students can successfully transfer to a four year university and position themselves for academic, social, and career success.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N3ywE5TWI73Lj85k1in3lg.png" /><figcaption>The One Million Degrees panelists.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j_8IzUDpXCHj1xBPNtw5Mg.jpeg" /><figcaption>Five panelists after the OMD panel.</figcaption></figure><p>Along with four other panelists, I was asked the following questions:</p><h3><strong>1. What factors matter most when choosing a four-year school to transfer to? What should students focus on first?</strong></h3><p>When I was choosing where to transfer, I focused on schools that were truly transfer-friendly and accepted the most credits from City Colleges of Chicago. I didn’t want to waste time or money retaking courses I had already completed.</p><p>Southern Illinois University immediately stood out — not just for its well-built cybersecurity program, but for its online flexibility and its articulation agreement with Wilbur Wright College. That partnership made my entire transfer path smooth and fast.</p><p>Transferring isn’t just about location or picking the school with the loudest reputation —<strong> it’s about alignment with your goals.</strong></p><p><strong><em>Choose a school that fits your life. </em></strong>If you need online options, make sure they’re actually available. If you’re working, look for programs designed with working students in mind. Always ask yourself:<br> <strong>“Will this program support my growth, my schedule, and my financial reality?”</strong></p><p>Some schools look prestigious or offer large scholarships, but still end up costing twice as much once you calculate the full picture.</p><p>Do the groundwork before you transfer: research, compare, ask questions, read agreements, talk to advisors, and visit campuses.<strong> Every step brings you closer</strong> to the school that fits you best.</p><p>You’re not just transferring , you’re expanding and<strong> stepping into your next, higher chapter.</strong></p><h3><strong>2. How can community college experiences — academic, leadership, or service — make a transfer application stand out? Any strategies that work especially well?</strong></h3><p>I didn’t start in cybersecurity. I began in ESL, then moved into Business Management, and eventually found my way into Cybersecurity.</p><p>I openly share my journey from being a nail technician to becoming a cybersecurity professional, and<strong> that honesty always resonates</strong>. Employers and universities notice authentic growth, not polished perfection.</p><p>Every class teaches you something valuable. Claim it. Even jobs that seem unrelated build transferable skills: communication, discipline, problem solving, attention to detail. These matter far more than impressive titles.</p><p>The moment you take ownership of your story and <strong>reframe it as skill building, </strong>you set yourself apart, and<strong> that is exactly what you want.</strong></p><p><strong>Good storytelling turns your journey into an advantage.</strong> Curiosity opens doors faster than perfection ever will.</p><p><strong>Take leadership opportunities wherever they appear</strong>. Join programs like One Million Degrees, apply for internships, volunteer, raise your hand in class. It all counts.</p><p>My advice: stay curious, keep asking questions, and always connect your past experiences to the person you are becoming now.</p><h3><strong>3. What challenges do transfer students usually face at four-year universities, and what strategies help them adjust successfully?</strong></h3><p>My biggest challenge was learning how to navigate everything on my own. At first I struggled to find classes, understand degree audits, or even locate my transcript. Community college felt easier because there was more hands on support. But when I transferred fully online, I had to teach myself how to manage new platforms and a new schedule.</p><p>Now it is much easier, but it took patience.</p><p>What helped me most was <strong>community</strong>. I stayed connected to friends from Wilbur Wright College, colleagues from my first IT job, and people I met at cybersecurity conferences like CyberCon, ChiBrrCon, CypherCon, SecureWorld, THOTCON, ISSA Chicago Chapter, and Black Hat. Those spaces helped me find mentors, real cyber security professionals who guide me whenever I need it. Through them, I have been invited to<strong> private industry events,</strong> which has<strong> opened even more doors.</strong></p><p>If you ever feel lost or unsure, remember you are not alone. Everyone has to figure things out.<strong> Confusion is normal and be patient with yourself. </strong>It is part of transferring and part of growth.</p><p>Build friendships early, join clubs,<strong> find mentors, and attend events that align with your interests. You want to be in those communities because being part of them will do good for you.</strong> Ask for help before you feel overwhelmed.</p><p>Do not wait for confidence to appear. Act first, and confidence will rise to meet you.</p><h3><strong>4. What common resume mistakes do you see students make, and what advice would you give to make a resume stronger and aligned with industry expectations?</strong></h3><p>The biggest mistake I see on resumes is listing tasks instead of impact. For example, instead of saying “Worked on lab projects,” say, “Created, implemented, and secured a lab network.”</p><p>Start every bullet with a<strong> strong verb like designed</strong>, <strong>deployed</strong>, <strong>configured</strong>, or <strong>implemented</strong>. Show results or skills gained. Keep your resume clear, concise, action driven, and no more than one or two pages.</p><p>Another mistake is using the same resume for every job. Customize it to match the language of the job description. <strong>Highlight the certifications, tools, and skills the employer is asking for. </strong>If a posting mentions attention to detail, problem solving, or communication, make sure those qualities are reflected on your resume.</p><p>If you really want to land a job and stand out,<strong> write a cover letter.</strong> Use it to explain <strong>why </strong>you want that specific role, how your experience aligns with the position, and which certification you are currently working on that supports the job you are applying for<strong>. A cover letter adds a personal touch and shows real intention and motivation.</strong></p><p>Do the work before you have the job.<strong> Build projects, join clubs, and share your progress online</strong>. Instead of sending hundreds of applications, try <strong>connecting directly with hiring managers</strong> or recruiters on LinkedIn. A short, thoughtful message often opens more doors than dozens of applications because it shows initiative and communication skills.</p><p>Employers also check LinkedIn to see your achievements and consistency. Add an objective at the top of your resume<strong> so recruiters immediately understand who you are, what you bring, and what you want.</strong></p><h3><strong>5. What are the best ways for students to showcase transferable skills from community college so they stand out on resumes, applications, and interviews?</strong></h3><p>Start by identifying the real value behind your experiences.<br> For example, my time as a nail technician taught me precision, communication, and client management. These are the same skills I now use in my current role for teamwork engagement.</p><p><strong>During interviews, connect your classroom projects to real world outcomes.</strong><br> Instead of saying “Completed a database assignment,” say:<br> “Developed and optimized SQL queries for member data, simulating real business reporting.” This shows employers you understand practical impact, not just theory.</p><p><strong>Use stories </strong>to explain how community college taught you adaptability, time management, and resilience. Employers value grit and curiosity, and community college builds both.</p><p>Mention how you balanced work, school, and life while achieving strong results. That story translates into <strong>professional reliability.</strong></p><p>Take small actions, build something, and share your progress publicly. <strong>Visibility matters. You want to be seen- being seen in this case will do good for you! Stand out.</strong></p><p><strong>Storytelling is your personal brand,</strong> and it makes your skills unforgettable.</p><p>Translate your past into power. Let your journey become<strong> your competitive edge.</strong></p><p><strong>And no matter where you go, act like you’ve been there before.</strong></p><p>We talked with students a lot about how to choose the right school and where to go next, but the truth is, none of it matters unless you understand the <strong>reason </strong>behind your actions.</p><h3>Reflect on your WHY behind everything you do.</h3><p>Make sure you always have a <strong>game plan to match your wishes,</strong> otherwise they will always remain wishes.</p><p>Why study that much? Why learn all that?<br> Why get so many certificates?<br> Why take so many courses? Why study for your bachelor?<br> Why try all that ? Why suffer?</p><p>Why read that many books?<br> Why be ambitious?<br> Why go that far?</p><p>Why earn that much? Why work for yourself? Why wake up at 4 or 5 in the morning? Why go to sleep early?</p><p>Why attend so many conferences?<br>Why network with many people?<br>Why put yourself through so much?<br>Why leave your comfort zone and struggle?</p><p><strong>The answer is simple: Why not?</strong></p><p><strong>What else are you going to do with your life if not checking how good you can become.</strong></p><p>Why not be curious and see how much you can earn, how much you can learn, and how strong and resilient you can become? Why not?</p><p>The alternative is staying in your comfort zone and living the same life until you decide to change it. <strong>Stay until you decide to go.</strong></p><p>You get to see the world.<br>You get to learn new cultures.<br>You get to expand.<br>You get to try new things and learn things you never expected.<br>You get to read the books, meet big people, and experience more.<br>Because why not?</p><p>This all takes excellent discipline. Discipline equals freedom. Do not postpone the better version of yourself or your better future any longer.</p><p><strong>Without a why, you are busy — not growing.</strong><br> Busy doing what others expect.<br> Busy chasing random goals.<br> Busy reacting instead of creating.<br> Busy moving, but not moving forward.</p><p>When you do not know your why, your days fill up with tasks that do not match your direction. You work hard but feel empty. You push yourself but see little progress. You check boxes instead of building a life.</p><p><strong>A busy life is not the same as a meaningful life.</strong><br>Busy is noise. Your why is clarity.</p><p><strong>With a why, you are intentional.</strong><br>1. Your time has purpose.<br>2. Your choices have direction.<br>3. Your discipline has meaning.</p><p>Your why turns motion into momentum. Your why turns effort into elevation.</p><p>Stay curious. Attend conferences. Set new goals. Read new books. Find new mentors. Journal your experiences. Have conversations. Make contact and stay in touch. Write a book. Build something and publish it. Read more. Go on walks. Become a creator and a mentor. Follow your passion. Inspire youth. Open yourself to this world and do it <strong>now</strong>. Let your new self emerge and stay.</p><p>This will take you into a new journey.<strong> Each day ask God for guidance and let Him do His part</strong> while you work on yours. That is when real magic happens. That is what happened in my case, and I speak from experience.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sBHwY9V0Ti2rtPQUFT3d8A.jpeg" /><figcaption>The OMD Panel at Wilbur Wright College.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=547bc2f917b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>