<?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 Rei on Medium]]></title>
        <description><![CDATA[Stories by Rei on Medium]]></description>
        <link>https://medium.com/@rei21220?source=rss-e3cc961721f9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*bWFoAlxu5em8ZtN5</url>
            <title>Stories by Rei on Medium</title>
            <link>https://medium.com/@rei21220?source=rss-e3cc961721f9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 00:17:35 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rei21220/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[Is a Secure Online Exam Really Secure?]]></title>
            <link>https://medium.com/@rei21220/is-a-secure-online-exam-really-secure-fca24db06dd9?source=rss-e3cc961721f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/fca24db06dd9</guid>
            <dc:creator><![CDATA[Rei]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 09:56:37 GMT</pubDate>
            <atom:updated>2026-11-03T09:56:37.543Z</atom:updated>
            <content:encoded><![CDATA[<p>The university believes SEB, camera proctoring and university logins make its exams secure. The four incidents suggest otherwise: these tools protect one moment, the student’s screen during the exam, while the incidents happened elsewhere.</p><p>Question 1 – Security beyond technology</p><p>A system is not secure because it has security tools. It is secure when it meets defined objectives against realistic threats, at a risk level the organization accepts. The question is not “what do we have?” but “what must not go wrong, and does it?”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1017/1*g5baAvcTI1MU_lE1lcQDdA@2x.jpeg" /></figure><p>None of the existing tools addresses these. SEB and proctoring watch the student’s device during the exam, but A happened before it, B in the staff dashboard, and C in the infrastructure. Login, the one identity control, fails in D because passwords can be shared.</p><p>Question 2 – Thinking like an analyst</p><p>A vulnerability is a weakness; a threat is a potential source of harm; an attack is an actual attempt to exploit the weakness; risk combines likelihood and impact.</p><p>Incident A – Question leakage</p><p>Asset: the question bank before the exam.</p><ul><li>Vulnerabilities: many staff can view questions early; the dashboard likely uses passwords alone; views and exports are probably not logged.</li><li>Threats and attack: an insider, or a student who phishes a lecturer with a fake “LMS login” email, opens the dashboard and screenshots the questions.</li><li>Attack surfaces: human (phishing) and software (dashboard).</li><li>Consequences: a cancelled exam, disadvantaged honest students, and no way to find out how it happened.</li></ul><p>Incident B – Unauthorized grade change</p><ul><li>Assets: grade records and the logs that should explain them.</li><li>Vulnerabilities: no per-user audit trail, possibly shared admin accounts, possibly weak access control on the grade function.</li><li>Threats and attack: a student uses a stolen staff session, or edits a request to point at their own record, and changes the grade.</li><li>Attack surfaces: software (dashboard and APIs) and human (credential theft).</li><li>Consequences: once one change is unexplained, every grade becomes questionable, which matters for appeals and accreditation.</li></ul><p>In my judgment, B carries the higher risk: its impact spreads beyond one exam, and nothing currently detects it.</p><p>Question 3 – Three priorities</p><p>1. Zero Trust access for staff (preventive). Every request to questions or grades is checked against identity and role, not just a valid session [2]. Staff use phishing-resistant MFA. Applying least privilege, only the exam owner sees questions, shortly before release; applying separation of privilege, a published grade needs a second approver to change [3]. Reduces: A and B. Trade-off: friction for lecturers and MFA support effort.</p><p>2. Tamper-evident audit logs with alerts (detective, response). Individual accounts only; every view, export and grade change goes to append-only logs, with alerts on bulk exports or late grade edits. Reduces: restores accountability for B and makes A investigable. Trade-off: storage cost, staff-privacy concerns, and someone must watch the alerts.</p><p>3. Exam-day resilience (preventive, recovery). Load-test before major exams, add redundant capacity, and let SEB save encrypted, time-stamped answers locally until the connection returns, backed by a written extension and re-sit plan. Reduces: C, turning an outage into a delay rather than lost work. Trade-off: cost, and local copies must be protected.</p><p>I did not prioritize heavier proctoring for D. More surveillance costs privacy and risks false accusations, which conflicts with Saltzer and Schroeder’s principle of psychological acceptability [3]. A proportionate alternative is an ID check at the start, randomized question pools, and short oral follow-ups for a random sample.</p><p><strong>Conclusion</strong></p><p>Together, these measures form defense in depth across the human, software and infrastructure layers, covering the Protect, Detect, Respond and Recover functions of the NIST Cybersecurity Framework [1]. The university’s mistake was not choosing bad tools, but treating their presence as proof of security. An online exam is secure only when every objective holds, before, during and after the exam.</p><p><strong>References</strong></p><p>1. NIST. (2024). The NIST Cybersecurity Framework (CSF) 2.0 (NIST CSWP 29). nist.gov</p><p>2. Rose, S., Borchert, O., Mitchell, S., &amp; Connelly, S. (2020). NIST SP 800–207: Zero Trust Architecture. NIST. csrc.nist.gov</p><p>3. Saltzer, J. H., &amp; Schroeder, M. D. (1975). The protection of information in computer systems. Proceedings of the IEEE, 63(9), 1278 – 1308. doi.org/10.1109/PROC.1975.9939</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fca24db06dd9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sending Large Files Safely over an Untrusted Network]]></title>
            <link>https://medium.com/@rei21220/sending-large-files-safely-over-an-untrusted-network-a54161e699be?source=rss-e3cc961721f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/a54161e699be</guid>
            <dc:creator><![CDATA[Rei]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 07:51:26 GMT</pubDate>
            <atom:updated>2026-11-02T07:51:26.505Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>1. Introduction</strong></p><p>Alice must send large files to Bob over a network an attacker controls: messages may be read, modified, replaced, dropped, reordered or replayed. The design must provide confidentiality (only Alice and Bob read the data), integrity (Bob detects changes), authentication (Bob knows the data came from Alice) and non-repudiation (Alice cannot credibly deny sending it, while her signing key stays under her control). No single primitive provides all four.</p><p><strong>2. Security Analysis</strong></p><p>Encryption alone is not enough. It hides content, but modes such as CTR are malleable: a flipped ciphertext bit flips the same plaintext bit, undetected. It also says nothing about the sender.</p><p>Three different guarantees. Integrity concerns the data itself, authentication its origin, and non-repudiation whether a third party can be convinced of that origin.</p><p>Symmetries encryption for bulk data. AES is fast and hardware-accelerated; public-key operations are far slower, so they protect only small items (one key exchange, one signature).</p><p>HMAC cannot give non-repudiation. Anyone holding the shared key, including Bob, can produce a valid HMAC tag, so a judge cannot tell who made it. Only a signature made with a key that Alice alone holds can [2].</p><p>Key assumptions. Alice’s private key is secret and only hers. Bob has her authentic public key via a university certificate authority (CA). Session keys are random, single-use and destroyed afterwards. Nonces never repeat [1][3].</p><p><strong>3. Proposed Architecture</strong></p><ol><li>Alice holds a signing key pair (SK_A, PK_A); Bob holds a key-agreement pair (SK_B, PK_B), both certified by the university CA.</li><li>1. Key establishment. Alice generates an ephemeral key pair, runs ECDH with PK_B, and uses HKDF to derive a one-time AES-256 session key K [4][6]. She signs her ephemeral public key with SK_A. Without it, a man-in-the-middle could agree a key with each side.</li><li>2. Chunking. Files are split into numbered, fixed-size chunks.</li><li>3. Signed manifest. Alice lists the transfer ID, sender, recipient, timestamp, chunk count and each chunk’s SHA-256 hash, signs this with SK_A, and encrypts it under K. Without the signature, there is no non-repudiation.</li><li>4. Authenticated encryption. Each chunk is encrypted with AES-256-GCM under K, using the chunk counter as nonce. The associated data (authenticated, not encrypted) binds the transfer ID, index, chunk count and a final flag [1]. Without the tag, ciphertext could be altered undetected.</li><li>5. Bob’s verification. Bob checks the key-exchange signature, derives K and opens the manifest: Alice’s signature must verify, he must be the recipient, and the transfer ID must be new and the timestamp fresh. Every chunk must pass its tag check, in order, with none missing, and every hash must match. Any failure means reject; otherwise he accepts and keeps the signed manifest as evidence.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/761/1*fSXtYxeIFpiztMiYZz5xLA@2x.jpeg" /></figure><p>Alice’s four steps (left) produce three messages across the untrusted network. Bob (right) runs four checks in order; any failure rejects the whole transfer.</p><p><strong>4. Critical Evaluation</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/731/1*GLihCBsWdptAERxDJTIbhQ@2x.jpeg" /></figure><p>Best option. Only C meets all four. B, with the MAC applied after encryption [7], proves only that someone holding the key sent the data.</p><p>Trade-offs. C needs a certificate infrastructure and more code, but performs close to B because bulk data still uses AES.</p><p>If K leaks. The attacker can read that transfer but cannot forge Alice’s signature, so altered chunks fail the hash check; other transfers use different keys. If Bob’s long-term SK_B leaks, recorded transfers become readable; ephemeral keys on both sides would prevent this.</p><p>If SK_A leaks. The attacker can sign as Alice, so non-repudiation fails until the certificate is revoked; trusted timestamps keep earlier signatures valid [3].</p><p>Replay, loss and ordering. Handled by the fresh transfer ID and timestamp, and by the index, count and final flag bound into each chunk.</p><p><strong>5. Final Engineering Judgment</strong></p><p>For confidential university records, I would approve Option C, built from standard parts: TLS 1.3 with mutual certificate authentication for the channel, plus a signed manifest per file [5][2].</p><ul><li>Security effectiveness. TLS 1.3 secures the connection, with forward secrecy, but cannot give non-repudiation, since both ends share its keys. The signed manifest fills that gap and stays with the stored file.</li><li>Performance. AES-GCM is hardware-accelerated and signing is one operation per file, so large transfers stay fast.</li><li>Key management and trust. Signing keys live in a hardware security module; certificates from the university CA are rotated and revocable under a written policy [3].</li><li>Implementation complexity. In my view, vetted libraries beat custom code, where bugs like nonce reuse usually appear.</li></ul><p>What remains. The design assumes uncompromised endpoints: malware on Alice’s machine could sign anything. File sizes and timing stay visible, and everything depends on a working CA and revocation. These need operational controls, not more cryptography.</p><p><strong>6. References</strong></p><p>1. Dworkin, M. (2007). NIST SP 800–38D: Recommendation for Block Cipher Modes of Operation: Galois/Counter Mode (GCM) and GMAC. NIST. csrc.nist.gov</p><p>2. NIST. (2023). FIPS 186–5: Digital Signature Standard (DSS). csrc.nist.gov</p><p>3. Barker, E. (2020). NIST SP 800–57 Part 1 Rev. 5: Recommendation for Key Management. NIST. csrc.nist.gov</p><p>4. Barker, E., et al. (2018). NIST SP 800–56A Rev. 3: Recommendation for Pair-Wise Key-Establishment Schemes Using Discrete Logarithm Cryptography. NIST. csrc.nist.gov</p><p>5. Rescorla, E. (2018). RFC 8446: The Transport Layer Security (TLS) Protocol Version 1.3. IETF. rfc-editor.org</p><p>6. Krawczyk, H., &amp; Eronen, P. (2010). RFC 5869: HMAC-based Extract-and-Expand Key Derivation Function (HKDF). IETF. rfc-editor.org</p><p>7. Bellare, M., &amp; Namprempre, C. (2000). Authenticated encryption: Relations among notions and analysis of the generic composition paradigm. ASIACRYPT 2000, LNCS 1976, 531 – 545. doi.org/10.1007/3-540-44448-3_41</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a54161e699be" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Logged In Is Not the Same as Allowed In]]></title>
            <link>https://medium.com/@rei21220/logged-in-is-not-the-same-as-allowed-in-dcd67eee56c9?source=rss-e3cc961721f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/dcd67eee56c9</guid>
            <category><![CDATA[application-security]]></category>
            <category><![CDATA[owasp]]></category>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[web-development]]></category>
            <dc:creator><![CDATA[Rei]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 07:38:09 GMT</pubDate>
            <atom:updated>2026-11-02T07:38:09.862Z</atom:updated>
            <content:encoded><![CDATA[<p>You log in to your bank’s website and open a statement. The address bar ends in …/statements/1042. Out of curiosity, you change it to 1043. Should you see a stranger’s bank statement?</p><p>Obviously not. Yet this is one of the most common security failures on the web.</p><p><strong>The issue: being logged in is not the same as being allowed</strong></p><p>A web application has to answer two different questions. Authentication asks who are you? That is the login screen. Authorization, or access control, asks are you allowed to see or do this?</p><p>Broken access control happens when an application answers the first question but forgets the second. It checks that you are logged in, but not whether record 1043 belongs to you. It mainly affects web applications. OWASP, a non-profit security community, ranks it the number one web application risk in its 2025 Top 10. In OWASP’s data, every application tested had some form of it [1].</p><p><strong>Why it matters: 885 million documents, one digit apart</strong></p><p>A documented incident. In May 2019, security journalist Brian Krebs reported that First American Financial, a major U.S. title-insurance company, had left about 885 million mortgage-related documents open on its website. Anyone with one valid document link could change a single digit in the address and open someone else’s file. The files included Social Security numbers, bank statements and driver’s license images [3]. In 2021, the U.S. Securities and Exchange Commission found that the company’s security staff had spotted the problem months earlier but never fixed or escalated it. The company paid a penalty of $487,616 [4].</p><p>A hypothetical example. Imagine a university portal where you download your transcript from …/transcript?id=12345. If the server only checks that you are a logged-in student, a classmate could read everyone’s grades by counting upward.</p><p>For users, this means identity theft and fraud; for organizations, fines and lost trust, all from one missing check.</p><p><strong>Which controls reduce the risk?</strong></p><p>The OWASP Top 10 tells us what tends to go wrong. OWASP’s Application Security Verification Standard (ASVS) turns that into specific requirements a team can test [1][2]. Three controls matter most here:</p><ul><li>Deny by default. Nothing is accessible unless a rule explicitly allows it [1].</li><li>Check ownership on the server, for every request. ASVS 5.0 requires that each data item is available only to users with explicit permission (requirement 8.2.2), enforced on the server rather than in code a user can change (8.3.1) [2]. Think of a hotel key card: getting into the building is not enough. It should open only your room.</li><li>Write the check once and reuse it. OWASP advises building access control in one place and using it everywhere [1]. Scattered copies are where one gets forgotten.</li></ul><p>In my analysis, these fit the risk because the root cause is a missing question: is this yours? Making record numbers random and hard to guess helps a little, but a leaked link still works. Only the ownership check actually closes the hole.</p><p><strong>How do we know the controls work?</strong></p><p>A rule on paper proves nothing. The simplest method is a two-account test: create two users, Alice and Bob, and try to make Alice reach Bob’s data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iFeICwfEst0SSNvhmzjFEw@2x.jpeg" /></figure><p>The automated row matters most. The First American flaw was found and then left in place [4]; a failing test makes a missing check impossible to ignore. In my view, automated scanners are weak here because they cannot know who owns which record.</p><p><strong>My judgment</strong></p><p>Access control should be treated as a feature with its own acceptance tests, not as something we hope is there. My recommendation is simple: no page or API that returns personal data should be released without a passing two-account test. Developers should write these tests, and security or QA teams should check that every sensitive endpoint has one.</p><p>This has trade-offs. Writing tests for every endpoint takes time, and apps with many roles (student, lecturer, admin) multiply the combinations. Start with the most sensitive endpoints. Tests only prove what they cover, so occasional manual testing still matters. And as First American shows, finding a flaw is not enough: someone must own fixing it.</p><p>A login screen proves who you are. Only access control decides what you get to see.</p><p><strong>References</strong></p><ol><li>OWASP. (2025). OWASP Top 10:2025, A01 Broken Access Control. top10.owasp.org</li><li>2. OWASP. (2025). Application Security Verification Standard 5.0.0, V8 Authorization. github.com/OWASP/ASVS</li><li>3. Krebs, B. (2019, May 24). First American Financial Corp. Leaked Hundreds of Millions of Title Insurance Records (republished by Security Boulevard). securityboulevard.com</li><li>4. FOX 29. (2021, June 15). SEC charges mortgage title issuer First American with cybersecurity vulnerability violation. fox29.com</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dcd67eee56c9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Code Nobody on Your Team Wrote]]></title>
            <link>https://medium.com/@rei21220/the-code-nobody-on-your-team-wrote-74dd215484a7?source=rss-e3cc961721f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/74dd215484a7</guid>
            <category><![CDATA[supply-chain-security]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[log4j]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Rei]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 07:24:28 GMT</pubDate>
            <atom:updated>2026-11-02T07:24:28.158Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine building a house where most of the bricks, pipes and wiring were made by strangers you have never met. You checked that the lights turn on. You never checked who made the wires.</p><p>That is how most modern software is built.</p><p>The issue: you ship code you never wrote</p><p>Today’s apps are mostly assembled, not written from scratch. Developers combine ready-made pieces of code called third-party libraries, often free open-source projects maintained by volunteers. A library might handle dates, images, passwords or logging, so the team can focus on its own features.</p><p>This is not laziness. NIST’s Secure Software Development Framework (SSDF) actually recommends reusing well-secured existing software instead of reinventing it [3]. The catch is that every library brings its own bugs into your product, plus the bugs of the libraries it depends on. You inherit risks you did not write, and often do not know you have.</p><p>Why it matters: two wake-up calls</p><p>Log4Shell (2021). In December 2021, a flaw was found in Log4j, a free Java library that records what a program is doing, like a diary. If an attacker could get a server to write down a specially crafted piece of text, such as a username or chat message, the server would fetch and run the attacker’s code [4]. It received the maximum severity score. The U.S. Cyber Safety Review Board later found that Log4j sits inside thousands of other software packages, and that many companies could not quickly tell where they were even using it. The Board called it an “endemic” vulnerability that could linger for a decade or more [1].</p><p>The xz Utils backdoor (2024). On March 29, 2024, Microsoft engineer Andres Freund noticed that remote logins on a Linux test machine were using unusual CPU and running slowly [5]. He traced it to hidden malicious code in xz Utils, a small compression library used across Linux. The person behind it, using the name “Jia Tan,” had spent about two years earning the trust of the project’s overworked main maintainer before slipping the backdoor into versions 5.6.0 and 5.6.1 [2][6]. It had already reached test versions of major Linux distributions. One curious engineer stopped it, not any formal process.</p><p>The deeper problem: three hidden assumptions</p><p>The incidents above are facts. What follows is my analysis of why they keep happening.</p><p>“We know what’s in our software.” Usually we don’t. A developer picks a handful of libraries, but each one pulls in others. It is like inviting one friend to a party and having them bring five more you never met. The Log4j response showed this clearly: teams could not patch what they could not find [1].</p><p>“Popular means well looked after.” Many critical libraries rest on one or two unpaid volunteers. The review board pointed to this thinly resourced open-source community as a core risk [1]. The xz attacker did not break any code at first. He exploited a tired person who needed help.</p><p>“If it passes the tests, it’s fine.” Log4j’s dangerous behavior was a feature working as designed. Functional tests check that software does what it should. They rarely check what else it can be tricked into doing. And code reviews usually cover the team’s own code, not the thousands of lines it imports.</p><p>My judgment: treat dependencies like a food supply chain</p><p>A good restaurant knows every ingredient it serves and who supplied it. In my view, software teams should hold themselves to the same standard. Three practical changes:</p><ol><li>Keep an automatic ingredient list. Every build should produce a Software Bill of Materials (SBOM), a list of every library inside the product. The review board recommended better SBOM tooling and accurate inventories, and the SSDF lists SBOMs as a way to track components [1][3]. Trade-off: a list nobody reads is useless, so it must be connected to automatic alerts when a listed library gets a known vulnerability.</li><li>2. Check a library’s health before adding it. Is it still maintained? By how many people? When was the last release? The SSDF asks teams to confirm components are still maintained and to plan for ones that are not [3]. Trade-off: this slows development, and sometimes there is no healthier option. Writing your own replacement is often riskier, especially for things like encryption.</li><li>3. Invest in what you depend on. Companies that profit from open source should fund maintainers or give staff time to help. The review board called for more resources for open-source security [1]. Limitation: money alone will not stop a patient attacker like “Jia Tan,” so teams must also practice updating quickly when the next alarm comes.</li><li>Secure software is not only about the code you write. It starts with knowing whose code you are running.</li><li>References</li><li>1. Cyber Safety Review Board. (2022, July). Review of the December 2021 Log4j Event: Key Findings and Recommendations. CISA. cisa.gov</li><li>2. Cato Networks. (2024, April 1). XZ Backdoor RCE (CVE-2024–3094) Is the Biggest Supply Chain Attack Since Log4j. catonetworks.com</li><li>3. NIST. (2022). SP 800–218: Secure Software Development Framework (SSDF) Version 1.1, practice PW.4. csrc.nist.gov</li><li>4. NIST National Vulnerability Database. CVE-2021–44228 (Log4Shell). nvd.nist.gov</li><li>5. Freund, A. (2024, March 29). backdoor in upstream xz/liblzma leading to ssh server compromise. oss-security mailing list. openwall.com</li><li>6. NIST National Vulnerability Database. CVE-2024–3094 (xz Utils). nvd.nist.gov</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74dd215484a7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Calling Users the Weakest Link]]></title>
            <link>https://medium.com/@rei21220/stop-calling-users-the-weakest-link-564d49f65505?source=rss-e3cc961721f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/564d49f65505</guid>
            <category><![CDATA[information-security]]></category>
            <category><![CDATA[multifactorauthentication]]></category>
            <category><![CDATA[social-engineering]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[usability]]></category>
            <dc:creator><![CDATA[Rei]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 07:17:51 GMT</pubDate>
            <atom:updated>2026-11-02T07:17:51.911Z</atom:updated>
            <content:encoded><![CDATA[<p>It is late at night and your phone buzzes: “Are you trying to sign in? Approve / Deny.” You tap Deny. It buzzes again. And again, for an hour. Then a message arrives from someone who says they are from your company’s IT team: approve one, and the alerts will stop.</p><p>Would you tap Approve? A lot of tired people would. And when they do, the headline usually calls it “human error.”</p><p>The issue: blaming the person who clicked</p><p>In cybersecurity, people are often called “the weakest link.” When an attack succeeds because someone clicked a link, shared a code or approved a login, the investigation tends to stop at that person. The fix that follows is usually more rules and more training for users.</p><p>Is that blame fair, and does it make us safer? I argue it often does neither, because many “user mistakes” are predictable results of system design.</p><p>Why it matters: the Uber breach of 2022</p><p>The opening scene is not imaginary. In September 2022, Uber disclosed that an attacker had likely bought a contractor’s work password on the dark web, after malware infected the contractor’s personal device [3]. The account was also protected by multi-factor authentication (MFA): each login sent an “approve” notification to the contractor’s phone.</p><p>The attacker kept trying to log in, flooding the phone with notifications. Then, posing as Uber IT on WhatsApp, he told the contractor that accepting one request was the only way to stop them [3][4]. The contractor accepted, and the attacker reached Uber’s internal systems. Security teams now call this technique MFA fatigue, or push bombing.</p><p>It is easy to say the contractor should have known better. But this pattern is old. In a 1999 study, researchers Anne Adams and Angela Sasse found that people who broke password rules were often reacting to how the security mechanisms were built, for example creating linked passwords like tom1, tom2, tom3 to cope with too many rules [1]. Their conclusion, more than 25 years ago, was in the title of their paper: Users Are Not the Enemy.</p><p>Anyone who has been asked to “just forward the OTP code” has seen the same trick aimed at ordinary people.</p><p>The deeper problem: a design choice, not just a bad click</p><p>Look at what the system asked of the contractor. A single “Approve” button, with no context, repeated without limit, at whatever hour the attacker chose. That design turns security into a test of willpower that the attacker can retake as many times as he likes.</p><p>Better designs already existed. In October 2022, the U.S. Cybersecurity and Infrastructure Security Agency (CISA) urged organizations to adopt phishing-resistant MFA, such as security keys. Where that is not yet possible, CISA recommended number matching: the user must type a number shown on the login screen into the app [2]. A tired person in bed cannot type a number they never saw.</p><p>So is this a technology problem or a human problem? In my view, it is a human problem created by a technology choice. People will always get tired; design decides whether that becomes a breach.</p><p>The law and professional ethics point the same way. Indonesia’s Personal Data Protection Law (UU No. 27 Tahun 2022, Article 35) places the duty to protect and ensure the security of personal data on the data controller, meaning the organization, not the individual employee [5]. The ACM Code of Ethics asks computing professionals to design systems that are “robustly and usably secure” (principle 2.9) [6]. Security that only works for perfectly alert users is not usably secure.</p><p>My judgment: change the question from “who clicked?” to “why could one click do this?”</p><p>Users are not blameless. But blame should follow power, and the people who choose and configure systems have far more power than the person holding the phone. Three changes follow:</p><ol><li>Design out single-tap failures (responsibility: security and IT teams). No important account should fall to one careless moment. That means phishing-resistant MFA where possible, number matching where not, and limits on how many prompts can be sent [2].</li><li>2. Treat repeated mistakes as bug reports (responsibility: organizations). If many employees fall for the same trick, the system is producing that error. Reporting should also be blame-free, so strange prompts get reported, not hidden.</li><li>3. Give users a few clear duties, not dozens of vague ones (responsibility: users). “Never approve a login you did not start, and report it” is a rule people can actually follow. Adams and Sasse showed that rules which clash with real work get bypassed [1].</li><li>This is reasonable because it gives each fix to the party best able to make it, which is where Indonesian law already places the duty [5]. Training should support good design, not replace it.</li><li>A question for you</li><li>If a security system only works when a tired person makes a perfect decision at midnight, is the person failing the system, or is the system failing the person?</li><li>References</li><li>1. Adams, A., &amp; Sasse, M. A. (1999). Users are not the enemy. Communications of the ACM, 42(12), 40 – 46. doi.org/10.1145/322796.322806 (open copy: UCL Discovery)</li><li>2. CISA. (2022, October). Implementing Number Matching in MFA Applications (fact sheet). cisa.gov</li><li>3. InfoQ. (2022, September). Multi-Factor Authentication Fatigue Key Factor in Uber Breach. infoq.com</li><li>4. KPMG India. (2022, September 27). Cyber Threat Notification: Lessons to learn from the Uber security breach. kpmg.com</li><li>5. Republic of Indonesia. (2022). Undang-Undang Nomor 27 Tahun 2022 tentang Pelindungan Data Pribadi, Article 35. jdih.kemenkeu.go.id</li><li>6. ACM. (2018). ACM Code of Ethics and Professional Conduct, principle 2.9. acm.org</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=564d49f65505" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When the Help Desk Becomes Part of the Security Perimeter]]></title>
            <link>https://medium.com/@rei21220/when-the-help-desk-becomes-part-of-the-security-perimeter-4519c50d09e0?source=rss-e3cc961721f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/4519c50d09e0</guid>
            <dc:creator><![CDATA[Rei]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 07:12:33 GMT</pubDate>
            <atom:updated>2026-11-02T08:10:34.648Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine fitting your front door with three locks, then telling the building manager to hand a spare key to anyone who calls and says they live there. That is roughly how many organizations treat their IT help desk.</p><p><strong>The problem: a side door around strong login</strong></p><p>Companies have spent years strengthening authentication, the step where a system checks that you are who you claim to be. Long passwords and multi-factor authentication (MFA), such as a code on your phone, make it hard for a stranger to log in as you.</p><p>But every login system also needs a way back in for people who forget passwords, lose phones or get locked out. Very often, that way back is a person: the help desk. An agent who resets your password or registers a new phone for your MFA is also performing authentication, just by phone call instead of software.</p><p>If that agent can be convinced by a confident voice and a few personal details, the attacker never needs to break MFA. They simply ask for it to be moved to their own phone. The core problem is that a system is only as strong as its weakest path to access, and the recovery path is often much weaker than the login it replaces.</p><p><strong>Why it matters: two very expensive phone calls</strong></p><p>In September 2023, MGM Resorts, one of the biggest casino and hotel operators in Las Vegas, was hit by a cyberattack. Slot machines, digital room keys and booking systems failed for days, and MGM told U.S. regulators the incident would cost about $100 million in that quarter [3]. The criminal group Scattered Spider publicly claimed it got in by finding an MGM employee on LinkedIn and calling the help desk while pretending to be that person [4]. MGM never confirmed how the attackers entered, so this claim deserves caution. But the method itself is well documented.</p><p>The U.S. Cybersecurity and Infrastructure Security Agency (CISA) and the FBI describe exactly this pattern in a joint advisory on Scattered Spider [1]. The group collects usernames, personal data and phone numbers, then uses social engineering to convince help desk staff to reset passwords and MFA. The 2025 update warns that members pose as employees to get the help desk to transfer MFA to a device the attacker controls.</p><p>The same pattern appeared in April 2025 at UK retailer Marks &amp; Spencer. Its chief executive said attackers entered through a third-party contractor using social engineering, which was reported as impersonating staff to fool a help desk [5]. The company expected about £300 million in lost operating profit, and online orders were disrupted for weeks [6].</p><p>Neither case needed a clever software exploit. In both, the costliest moment was a person doing their job well: helping someone who sounded like a stressed, locked-out colleague.</p><p><strong>What we should learn: treat recovery as authentication</strong></p><p>The usual reaction to incidents like these is “more security awareness training.” Training helps, but I think it puts the burden in the wrong place. Help desk agents are rewarded for solving problems quickly, and attackers deliberately sound urgent, senior and frustrated. Asking a busy agent to detect a lie over the phone is a weak control. The better fix is to design the process so that a convincing voice is simply not enough.</p><p><strong>The evidence points to three changes:</strong></p><ol><li>Make recovery at least as strong as login. If an account normally needs a password plus a phone, a reset should not succeed with a name and date of birth that anyone can find online. NIST’s updated Digital Identity Guidelines (SP 800–63B-4, 2025) treat account recovery as a formal process with defined methods, such as saved recovery codes or repeated identity proofing, not as an informal favor [2].</li><li>Verify through a channel the caller does not control. Among CISA’s mitigations is contacting the employee’s manager when someone asks to reset both their password and MFA at once [1]. In the same spirit, agents should call back on the number already on file, never one the caller provides.</li><li>Notify and slow down risky changes. NIST requires that the account holder be notified whenever recovery takes place [2]. A message like “Your MFA device was changed. Not you?” lets the real employee raise the alarm. For administrator accounts, a second approver or a short waiting period costs minutes but could save millions.</li></ol><p>This lesson applies at home too. Your email, bank and phone accounts all have “forgot password” links and support lines. Saving backup codes, protecting your recovery email and asking your mobile carrier for a SIM-change PIN close the same side door.</p><p>Strong security is not one strong lock. It is every path that ends in “access granted,” including the friendly voice on the support line.</p><p><strong>Final reflection</strong></p><p>What assumption are we making about security, and what could happen if that assumption is wrong?</p><p>We assume that the person asking for help is the person they claim to be, because they sound right and know the right details. If that assumption is wrong, we do not just fail to stop an attacker: we hand them a legitimate, fully working account, and every security tool afterward sees a normal employee logging in.</p><p><strong>References</strong></p><ol><li>CISA &amp; FBI. Scattered Spider (Joint Cybersecurity Advisory AA23–320A), first published Nov 16, 2023, updated Jul 29, 2025. cisa.gov</li><li>NIST. SP 800–63B-4: Digital Identity Guidelines, Authentication and Authenticator Management, Jul 2025, Section 4.2 (Account Recovery). pages.nist.gov</li><li>CDC Gaming Reports. SEC filing by MGM: Cyberattack will cost $100 million in adjusted earnings, Oct 5, 2023. cdcgaming.com</li><li>KTNV Channel 13 Las Vegas. MGM says hotel reservations back online as company deals with cyber attack aftermath, Sep 2023. ktnv.com</li><li>The Guardian. M&amp;S expects cyber-attack to last into July and cost £300m in lost profits, May 21, 2025. theguardian.com</li><li>Reuters (via The Star). M&amp;S says cyberattack to cost $400 million, May 21, 2025. thestar.com.my</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4519c50d09e0" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>