<?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 Jorge Castaneda, D.Eng (c), MLS on Medium]]></title>
        <description><![CDATA[Stories by Jorge Castaneda, D.Eng (c), MLS on Medium]]></description>
        <link>https://medium.com/@jorge.castaneda13?source=rss-fbd41c6aa29f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Zk8lRhW6pIo-vHLPEA7Zfw.png</url>
            <title>Stories by Jorge Castaneda, D.Eng (c), MLS on Medium</title>
            <link>https://medium.com/@jorge.castaneda13?source=rss-fbd41c6aa29f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 10:48:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jorge.castaneda13/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[HHS Rewrites the HIPAA Privacy Rule]]></title>
            <link>https://medium.com/@jorge.castaneda13/hhs-rewrites-the-hipaa-privacy-rule-de9b05811d41?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/de9b05811d41</guid>
            <category><![CDATA[hipaa-compliance]]></category>
            <category><![CDATA[hipaa]]></category>
            <category><![CDATA[hipaa-privacy-rule]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 13:05:04 GMT</pubDate>
            <atom:updated>2026-11-17T13:05:04.692Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*95oDcKITCKktp-3XxOu8tw.jpeg" /></figure><p>Yesterday, November 16, 2026, HHS published a final rule that rewrites large sections of the HIPAA Privacy Rule for the first time in over a decade. The rule, titled <em>HIPAA Privacy Rule: Changes to Support Coordinated Care and Individual Engagement and Reduce Regulatory Burdens</em> (<a href="https://www.federalregister.gov/public-inspection/2026-23429/hipaa-privacy-rule-changes-to-support-coordinated-care-and-individual-engagement-and-reduce">Federal Register document 2026–23429</a>), takes effect 60 days after Federal Register publication. Regulated entities then have 180 days from the effective date to comply.</p><p>This isn’t the reproductive health privacy rule from 2024 that a Texas court vacated in June 2025. This is the modernization rule that’s been working through the rulemaking process since a 2021 Notice of Proposed Rulemaking. It touches patient access timelines, fee transparency, care coordination permissions, and several administrative requirements that haven’t changed since the original Privacy Rule took effect.</p><p>If you’re an IT executive, compliance officer, or part of a legal team, the operational impact is quite significant. The access request response window just got cut in half. Fee schedules now have to be posted publicly. And the care coordination provisions create new data-sharing pathways that security programs need to account for.</p><p><strong>The Access Request Overhaul</strong></p><p>The change with the sharpest teeth is to the individual right of access under <a href="https://www.govinfo.gov/content/pkg/CFR-2011-title45-vol1/pdf/CFR-2011-title45-vol1-sec164-524.pdf">45 CFR Section 164.524</a>. The response window drops from 30 calendar days to 15 calendar days. Covered entities must respond “as soon as practicable, but no more than 15 calendar days after receipt of the request.” The rule retains one extension of up to 30 calendar days with written notice, but the baseline just got cut in half.</p><p>That alone will force workflow changes at most covered entities. Organizations that route access requests through legal review, or that rely on manual retrieval from multiple systems, don’t have the luxury of a 30-day buffer anymore. If your average turnaround was 25 days, you’re now out of compliance by default.</p><p>The rule also establishes that individuals have the right to obtain their electronic PHI through a personal health application, where the covered entity can produce it in that form and format. This isn’t entirely new, but the final rule makes the obligation explicit and closes the gap that many organizations used to avoid API-based data sharing.</p><p>On fees, covered entities must now post estimated fee schedules on their websites for copies of PHI provided under the right of access and under authorization. The rule specifies when electronic PHI must be provided at no charge and requires advance notice of fees before the entity begins processing a request. The days of using opaque or inflated fee structures to discourage access requests are over, at least on paper.</p><p>The rule also reduces identity verification burdens on individuals exercising access rights. This is a practical change. The verification requirements under the existing rule were routinely used as a friction layer, sometimes unintentionally, sometimes not, to slow or deny legitimate access requests.</p><p>Finally, individuals can now direct a covered entity to send an electronic copy of PHI from an electronic health record to a designated third party. This codifies a right that existed in guidance but hadn’t been spelled out in the regulatory text with this level of specificity.</p><p><strong>Care Coordination Gets Explicit Permission</strong></p><p>The Privacy Rule has always permitted uses and disclosures for treatment, payment, and health care operations (TPO) under <a href="https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/disclosures-treatment-payment-health-care-operations/index.html">45 CFR Section 164.506</a>. What it didn’t do clearly enough was address the growing universe of care coordination that happens outside the walls of a traditional covered entity.</p><p>The final rule fixes that in two ways. First, it clarifies that health plans may use and disclose PHI for both population-based and individual-focused care coordination and case management as part of health care operations. This was arguably permitted before, but the lack of explicit regulatory language made compliance teams nervous, and nervous compliance teams tend to say no.</p><p>Second, and more consequentially, the rule expressly permits covered entities to disclose PHI for individual-level care coordination and case management to social services agencies, community-based organizations, home and community-based services (HCBS) providers, and similar third parties offering health-related services. The HCBS definition follows the Medicaid framework at <a href="https://www.ecfr.gov/current/title-42/chapter-IV/subchapter-C/part-441">42 CFR Part 441</a>.</p><p>This is important for anyone working in integrated care delivery, value-based care arrangements, or social determinants of health programs. Under the old rule, sharing PHI with a food bank to coordinate nutrition support for a patient with diabetes, or with a housing authority to stabilize a patient’s living situation, occupied a gray area. Organizations either did it and hoped they were covered under TPO, or they didn’t do it, and the patient’s care suffered. The new rule settles the question.</p><p>From a security perspective, these new data-sharing pathways mean new data flows. Community-based organizations receiving PHI need to be evaluated for how they handle that data, even if they aren’t business associates in the traditional sense. IT Executives should expect compliance teams to come asking about risk assessments for these relationships.</p><p><strong>What Goes Away and What Gets Added</strong></p><p>The rule eliminates the requirement for direct-treatment providers to make a good faith effort to obtain written acknowledgment that an individual received the Notice of Privacy Practices. Every compliance professional knows this was a formality. Patients sign or don’t sign, the form goes in a file, and nobody reads it. Removing the requirement doesn’t change how NPPs work in practice. It just stops generating paper that served no protective purpose.</p><p>The NPP itself, however, gets substantive updates. Content requirements are revised to better explain individuals’ rights and how to exercise them. The rule adds a right to discuss the NPP with a designated contact person or office. It removes the exception that previously excused correctional institutions from providing inmates with an NPP. That last change reflects a broader trend in health care privacy law toward closing gaps in coverage that disproportionately affect incarcerated populations.</p><p>The rule also creates a carve-out for Telecommunications Relay Services. TRS communications assistants who facilitate calls for individuals who are deaf, hard of hearing, deaf-blind, or have a speech disability can now receive PHI disclosures without triggering business associate requirements. TRS providers are expressly excluded from the business associate definition, which means no BAA is required. This is a practical fix for a problem that shouldn’t have existed in the first place, where accessibility infrastructure was being treated as a data-sharing relationship.</p><p>Finally, the rule expands the existing permission to disclose PHI to military command authorities for mission readiness. The previous rule covered only the Armed Forces. The final rule extends the same permission to all Uniformed Services, adding the U.S. Public Health Service Commissioned Corps and the NOAA Commissioned Corps.</p><p><strong>What should IT Executives and Cybersecurity Teams Do?</strong></p><p>The 15-day access request window changes the technical requirements for how PHI retrieval systems work. If your organization relies on batch processing, manual record pulls from multiple systems, or queued workflows that assume a 30-day window, those need to be re-engineered. Automated retrieval from EHR systems, with direct patient portal delivery, is the most reliable way to meet the compressed timeline at scale.</p><p>The personal health app provision means your systems need to support authenticated API-based data export to third-party applications. If you’re running on certified EHR technology, the FHIR-based Patient Access API should already be in place. If you’re not, or if the API is technically available but operationally neglected, this rule makes it a compliance requirement with teeth. Audit your API endpoints, test them against real-world request patterns, and make sure your logging captures who requested what and where it went.</p><p>The care coordination provisions create new outbound data flows to organizations that may have limited security maturity. Social services agencies and community-based organizations receiving PHI need to be assessed. Even where no BAA is required, the covered entity remains responsible for applying the minimum necessary standard under <a href="https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.514">45 CFR Section 164.514(d)</a>. That means your disclosure policies and technical controls need to account for these new recipients. Work with compliance to define what PHI elements are appropriate for each type of care coordination disclosure and enforce those boundaries in your systems.</p><p>Review your BAA inventory. The rule provides some entities with additional time to update business associate agreements, but the specifics depend on the circumstances. Get ahead of this by identifying which BAAs reference Privacy Rule provisions that this final rule changes, particularly around access timelines and fee structures.</p><p>Finally, update your audit logging. The access request changes mean more requests are processed faster, potentially through new channels like personal health apps. Your audit controls under <a href="https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C/section-164.312">45 CFR Section 164.312(b)</a> need to capture the full lifecycle of these requests: receipt, processing, delivery, and any denials or extensions.</p><p><strong>What about Legal Compliance Teams?</strong></p><p>The Notice of Privacy Practices needs a rewrite. The content requirements have changed, the rights explanation needs to be updated to reflect the new access timelines and personal health app provisions, and the new right to discuss the NPP with a designated contact must be included. If your organization is also a covered entity subject to the <a href="https://www.ecfr.gov/current/title-42/chapter-I/subchapter-A/part-2">42 CFR Part 2</a> alignment rule from February 2024, the NPP updates from both rules should be coordinated into a single revision cycle.</p><p>Fee schedules for PHI access requests must be posted on the covered entity’s website. This means actually calculating what your fees are, documenting the methodology, and publishing them in a place patients can find. If your current fee structure doesn’t comply with the cost-based limitations (labor, supplies, postage only, no retrieval fees), fix that before you post anything.</p><p>Care coordination disclosure policies need to be written or revised. The rule expressly permits disclosures to social services agencies, community-based organizations, and HCBS providers for individual-level care coordination. Your policies should define which workforce members can make these disclosures, what PHI elements are appropriate under the minimum necessary standard, and how these disclosures are documented. The fact that a disclosure is now expressly permitted doesn’t mean it’s unrestricted. Minimum necessary still applies.</p><p>Identity verification procedures for access requests should be reviewed. The rule reduces the burden on individuals, which means your current verification requirements may need to be loosened. If your process requires notarized identification or in-person verification for routine access requests, it’s probably more restrictive than the rule intends.</p><p>Do the compliance timeline math. The rule takes effect 60 days after Federal Register publication, with a 180-day compliance window after that. Watch for the publication date, then count forward. That gives you roughly eight months from publication to full compliance. Some BAA updates may have additional time, but the core provisions, particularly the access request timeline change, don’t get extra runway.</p><p>The full rule text is available as <a href="https://www.federalregister.gov/public-inspection/2026-23429/hipaa-privacy-rule-changes-to-support-coordinated-care-and-individual-engagement-and-reduce">Federal Register document 2026–23429</a>.</p><p><em>This post is for informational purposes only and does not constitute legal advice. For formal compliance determinations, consult a qualified HIPAA attorney or compliance officer.</em></p><p>I hope you find this article helpful and informative. Thanks for stopping by!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de9b05811d41" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Quantifiable Risk and Defensible Governance in OT Security]]></title>
            <link>https://medium.com/@jorge.castaneda13/quantifiable-risk-and-defensible-governance-in-ot-security-8b2c6a822b5e?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/8b2c6a822b5e</guid>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[quantifiable-risk]]></category>
            <category><![CDATA[operational-technology]]></category>
            <category><![CDATA[sto]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 23:06:32 GMT</pubDate>
            <atom:updated>2026-11-14T23:06:32.461Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d5BZTRnLifc1uZ-oCa5jGQ.jpeg" /></figure><p>OT cybersecurity has operated on an unspoken truce: corporate networks handle the data, operational technology keeps the physical world moving, and an air gap keeps the two from ever having to negotiate. That truce doesn’t really apply anymore. Every automated turbine, substation, and assembly line now sits on an interconnected balance sheet. Yet most organizations still govern their operational environments using qualitative color charts and compliance checklists that fall apart the moment an auditor, an insurer, or a board member asks a direct question: what is our actual exposure in dollars, and can we prove we took reasonable care?</p><p>Moving from reactive compliance to resilient operations requires replacing vague technical posturing with three pillars: quantifiable risk, regulatory defensibility, and unified IT/OT governance.</p><p><strong>Replacing Heat Maps with Financial Probability</strong></p><p>Heat maps make for comforting slide decks, but they’re terrible decision-making tools. Calling a vulnerability “critical” tells a Chief Financial Officer nothing about capital allocation. When plant uptime costs thousands of dollars per minute, security leadership has to speak the language of financial probability, not technical severity.</p><p>Quantifiable risk shifts the conversation from what could break to what a disruption costs and how likely it is to occur. Instead of cataloging thousands of theoretical exposures, mature teams model concrete operational scenarios. That means tying specific asset failures directly to lost production hours, spoiled batch materials, and contractual SLA penalties. It means using historical incident frequency and control maturity to estimate annualized loss expectancy, giving executives a clear return-on-investment metric for every dollar spent on mitigation. And it means focusing engineering hours on the specific pathways that lead to safety incidents or prolonged plant shutdowns, rather than chasing every routine patch cycle.</p><p>When risk is expressed in financial terms, security stops being a cost center asking for blank checks and becomes a measurable lever for protecting operating margins.</p><p><strong>Regulatory Defensibility Over Point-in-Time Compliance</strong></p><p>Passing an annual audit is no longer a shield against liability. Across energy, manufacturing, water, and transportation, regulatory frameworks are shifting from prescriptive checklists to outcome-based accountability. Regulators, cyber insurers, and plaintiffs’ counsel aren’t asking whether you had a policy on paper. They’re asking whether your leadership exercised continuous, demonstrable due care.</p><p>Regulatory defensibility means building an evidentiary trail that holds up under scrutiny after an adverse event occurs. It rests on three habits.</p><p>The first is documented risk acceptance. When an aging programmable logic controller can’t be patched without halting production for six months, defensibility isn’t pretending the issue doesn’t exist. It’s formally documenting the operational constraint, implementing compensating controls, and having an accountable executive sign off on the residual risk.</p><p>The second is continuous telemetry over annual snapshots. Point-in-time assessments decay the day they’re printed. Defensible programs maintain verifiable records of asset inventories, configuration baselines, and access reviews as living operational artifacts, not binder tabs that collect dust between audit cycles.</p><p>The third is traceable decision-making. Every architectural exception, vendor connection, and incident response drill should leave a clear record of why a choice was made and who authorized it.</p><p>In a post-incident investigation, organizations rarely fail because they lacked the newest tool. They fail because they can’t prove they governed the tools they already had.</p><p><strong>Unified IT/OT Governance Without Operational Friction</strong></p><p>The fastest way to break an operational environment is to force corporate IT policies onto the plant floor without translation. Enterprise networks prioritize confidentiality and rapid change. Operational environments prioritize physical safety, deterministic timing, and continuous availability. Treating them as identical creates friction, and friction breeds shadow workarounds.</p><p>Effective IT/OT governance doesn’t mean merging two distinct engineering cultures into one. It means establishing a shared accountability model with clear boundaries.</p><p>The primary mandates differ. Corporate IT focuses on data confidentiality and integrity. OT focuses on physical safety and continuous uptime. The bridge between them is shared business-impact tiering that forces both sides to rank what matters in the same language.</p><p>Change management looks different on each side. IT can automate frequent patching. OT operates within scheduled maintenance windows. The unified approach relies on compensating controls between outages to maintain security posture without disrupting production.</p><p>Access and identity follow a similar split. Enterprise directories handle corporate IT. Local operator roles and vendor sessions handle OT. Brokered remote access with audit trails connects the two without compromising either.</p><p>Incident response is where the gap matters most. IT defaults to containment and system isolation. OT defaults to safe shutdown and manual failover. Joint playbooks with clear plant authority ensure that when something goes wrong, nobody’s guessing who makes the call.</p><p>When IT leaders respect the physics and safety constraints of the plant floor, and OT leaders embrace structured visibility and access accountability, governance starts functioning as operational reliability.</p><p>Resilience isn’t something you buy at the end of a procurement cycle. It’s the natural byproduct of knowing what your operations are worth, governing the boundary between enterprise data and physical machinery, and maintaining clear proof of your decisions along the way. Organizations that master quantifiable risk and defensible governance won’t just satisfy regulators. They’ll build operations capable of handling disruption without losing their footing.</p><p>I hope you find this post helpful and informative. Thanks for stopping by!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8b2c6a822b5e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Legal Architecture of Cybersecurity]]></title>
            <link>https://medium.com/@jorge.castaneda13/the-legal-architecture-of-cybersecurity-52fa9809b554?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/52fa9809b554</guid>
            <category><![CDATA[statute]]></category>
            <category><![CDATA[cybersecurity-law]]></category>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[legal-framework]]></category>
            <category><![CDATA[legal]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 15:51:27 GMT</pubDate>
            <atom:updated>2026-11-07T15:51:27.928Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LqgXGuwrevjgddd26e5Vfg.jpeg" /></figure><p>There is no unified cybersecurity code in the United States. I’ve mentioned this in a <a href="https://cybercasta.com/2026/06/20/congress-uses-the-word-cybersecurity-hundreds-of-times-it-never-actually-defines-it/">previous post</a>, the word “cybersecurity” shows up more than 200 times across federal legislation, but Congress never defined it in the U.S. Code. The closest thing to a statutory anchor is the <a href="https://www.cisa.gov/topics/cyber-threats-and-advisories/federal-information-security-modernization-act">Federal Information Security Modernization Act of 2014</a>, codified at <a href="https://www.congress.gov/bill/113th-congress/senate-bill/2521/text">44 U.S.C. 3552(b)(3)</a>, which defines “information security” around the familiar CIA triad: confidentiality, integrity, and availability.</p><p>What we actually have is a patchwork. Common law contract and negligence doctrines. <a href="https://www.federalreserve.gov/boarddocs/supmanual/cch/200806/ftca.pdf">Section 5 of the Federal Trade Commission Act</a> and its prohibition on unfair trade practices, dating back over a century. Sector-specific privacy regulations. State statutory mandates that multiply every legislative session. Criminal statutes that weren’t written with modern technology in mind but get applied to it anyway.</p><p>For IT leaders and CISOs, this fragmented legal architecture creates a problem that no single compliance framework solves. Your technical controls, your incident response plans, your vendor contracts, and your public disclosures all sit under different legal regimes with different enforcement mechanisms. Understanding how those regimes interact is not really optional. It could actually be the difference between a defensible security program and one that looks good until a regulator or a plaintiff’s attorney starts asking questions.</p><p><strong>Criminal Law and Corporate Security</strong></p><p>There is <a href="https://www.justice.gov/jm/jm-9-48000-computer-fraud">The Computer Fraud and Abuse Act (CFAA), 18 U.S.C. 1030</a>, which is the primary federal cybercrime statute. Enacted in 1984, it penalizes unauthorized access to computers and conduct that exceeds authorized access. Its <a href="https://www.congress.gov/crs-product/R46536">seven core prohibitions</a> span everything from obtaining national security information and government computer trespass to password trafficking and computer-based extortion.</p><p>The statute’s reach is broad. Under <a href="https://uscode.house.gov/us/usc/t18/s1030/">18 U.S.C. 1030(e)(2)</a>, a “protected computer” includes any computer used in or affecting interstate or foreign commerce or communication. The Seventh Circuit confirmed in <a href="https://btlj.org/data/articles2015/vol21/21_1_AR/21-berkeley-tech-l-j-0562-0562.pdf">United States v. Mitra (2005)</a> that this definition expands with technology. In that case, an emergency services radio system with a microchip qualified. Any device with built-in processing intelligence connected to interstate regulatory domains falls within the statute’s protective scope.</p><p>Penalties scale sharply. Intentionally damaging a protected computer through a knowing transmission can carry up to 10 years for a first offense and 20 years for a subsequent conviction under <a href="https://www.thefederalcriminalattorneys.com/federal-computer-hacking">Section 1030(a)(5)(A)</a>. Even negligently causing damage by intentional access carries a one-year maximum. Extortion involving computers tops out at five years, with 10 years on a second conviction.</p><p>For years, the hardest question under the CFAA wasn’t about outsiders breaking in. It was about insiders walking out.</p><p><strong>The Authorization Problem</strong></p><p>Federal circuits split sharply on whether an employee who has legitimate credentials violates the CFAA by accessing data for unauthorized purposes. The narrow view, held by the Second, Fourth, Sixth, and Ninth Circuits, said the statute only covers accessing information or system areas that are off-limits entirely. The Ninth Circuit’s decision in <a href="https://law.justia.com/cases/federal/appellate-courts/ca9/10-10038/10-10038-2012-04-10.html">United States v. Nosal (2012)</a> held that employees who downloaded customer lists they were authorized to access, then transferred them to a competitor in violation of corporate policy, didn’t exceed authorized access under the CFAA. The statute targets access restrictions, not use restrictions. The Second Circuit reached a similar result in <a href="https://law.justia.com/cases/federal/appellate-courts/ca2/14-2710/14-2710-2015-12-03.html">United States v. Valle (2015)</a>, ruling that a police officer who searched law enforcement databases for personal reasons did not violate the Act because he had permission to use the system.</p><p>The broad view, advanced by the Fifth, Seventh, and Eleventh Circuits, treated purpose-based violations of corporate use policies as federal computer crimes. In the <a href="https://www.casemine.com/judgement/us/591464f0add7b0493428265d">United States v. John (2010)</a>, the Fifth Circuit held that an employee who accessed customer data to commit fraud violated the CFAA because the employer’s authorization only covered legitimate business purposes. The Seventh Circuit went further in <a href="https://law.justia.com/cases/federal/appellate-courts/F3/455/749/569600/">International Airport Centers v. Citrin (2006)</a>, holding that a breach of fiduciary duty terminated an employee’s agency authorization, making subsequent data deletion an unauthorized act.</p><p>The Supreme Court resolved this split in <a href="https://www.supremecourt.gov/opinions/20pdf/19-783_k53l.pdf">Van Buren v. United States, 141 S. Ct. 1648 (2021)</a>. The Court adopted the narrow, access-based interpretation. An individual “exceeds authorized access” only when accessing information or system areas that are entirely off-limits to them, not when misusing authorized access for an improper purpose. The Court framed it as a “gates up” versus “gates down” distinction. If the gate to the data is up for a given user, using that data for a bad purpose is not a CFAA violation.</p><p>This is a significant operational consequence for IT leaders. After Van Buren, corporate acceptable use policy violations and terms of service breaches can’t form the basis of federal criminal CFAA prosecutions. Insider threat enforcement shifted from a criminal law backstop to a matter of contractual, civil, and employment remedies. If your insider threat program relied on the CFAA as its legal hammer, the hammer is now smaller.</p><p><strong>Web Scraping and Public Data</strong></p><p>The CFAA application to automated web scraping raised a parallel question. Can a website operator criminalize access to publicly available data by sending a cease-and-desist letter?</p><p>The Ninth Circuit drew the line in <a href="https://www.jlcw.org/case-law/hiq-labs-v-linkedin-9th-cir-2022">hiQ Laboratories, Inc. v. LinkedIn Corp. (2019)</a>. For private information behind authentication barriers, like username and password gates, bypassing those controls constitutes unauthorized access. That’s analogous to breaking and entering. But for information on open websites accessible to the public without authentication, the concept of access “without authorization” doesn’t apply. A cease-and-desist letter can’t create CFAA liability for scraping data that anyone can see.</p><p>If you’re protecting proprietary data, the CFAA backstop requires actual technical gatekeeping, not just a terms-of-service page. And if your business model depends on aggregating publicly available data, the hiQ decision provides important protection against CFAA overreach.</p><p><strong>Trade Secrets, Wire Fraud, and How Prosecutors Stack the Deck</strong></p><p>Digital trade secret theft falls under the <a href="https://www.congress.gov/104/plaws/publ294/PLAW-104publ294.pdf">Economic Espionage Act of 1996, 18 U.S.C. 1831–1839</a>, expanded by the <a href="https://www.congress.gov/bill/114th-congress/senate-bill/1890">Defend Trade Secrets Act of 2016</a>, which added a federal private civil right of action under 18 U.S.C. 1836(b). Section 1831 covers theft intended to benefit a foreign government. Section 1832 covers commercial misappropriation for economic advantage.</p><p>The critical threshold for IT Leaders such as CISOs sits in <a href="https://www.law.cornell.edu/uscode/text/18/1839">18 U.S.C. 1839(3)(A)</a>. Proprietary digital information qualifies as a trade secret only if the owner has taken “reasonable measures” to keep it secret. Courts evaluating this standard look at specific technical and operational controls: multi-factor authentication and strong password policies, enterprise firewalls and intrusion prevention systems, role-based logical access controls tied to job authorization, and physical and technical access controls on digital storage media. In <a href="https://www.justice.gov/archive/usao/iln/chicago/2008/pr0402_01a.pdf">United States v. Hanjuan Jin (N.D. Ill. 2012)</a>, the court examined precisely these kinds of controls when evaluating whether an enterprise met the reasonable measures bar.</p><p>This means your security controls do double duty. They’re not just protecting data from unauthorized access. They’re establishing the legal predicate for trade secret protection. If your access controls are weak, your encryption is inconsistent, or your data classification program is nonexistent, you may not be able to prove the information qualified as a trade secret at all.</p><p>Federal prosecutors don’t limit themselves to one statute when charging digital intrusions. <a href="https://www.govinfo.gov/app/details/USCODE-2024-title18/USCODE-2024-title18-partI-chap63-sec1343">The Wire Fraud Act, 18 U.S.C. 1343</a>, applies to any fraudulent scheme executed via interstate wire communications, and courts have held that accessing web pages or sending internal network emails qualifies. In <a href="https://oge.gov/Web/OGE.nsf/Resources/U.S.+v.+Selby,+557+F.3d+968+(9th+Cir.+2009)%C2%A0">United States v. Selby (9th Cir. 2009)</a>, the court confirmed that standard network activity falls within wire fraud’s jurisdictional reach.</p><p>Prosecutors pair wire fraud counts with CFAA charges for practical reasons. Wire fraud carries a maximum sentence of 20 years, compared to 5 or 10 years for most first-time CFAA offenses. Wire fraud also serves as a predicate offense under the <a href="https://www.justice.gov/archives/jm/criminal-resource-manual-109-rico-charges">Racketeer Influenced and Corrupt Organizations Act,</a> opening the door to enhanced penalties and criminal asset forfeiture. When a cyber intrusion results in federal prosecution, the CFAA charge is often just the entry point. Wire fraud is where the sentencing exposure lives.</p><p><strong>Breach Notification: 50 States, No Single Answer</strong></p><p>The United States has no national data breach notification standard covering all data types. Federal requirements are sector-specific. Financial institutions report under the <a href="https://www.ftc.gov/business-guidance/privacy-security/gramm-leach-bliley-act">Gramm-Leach-Bliley Act’s</a> interagency guidance when unauthorized access compromises sensitive customer information like SSNs, driver’s licenses, or account numbers combined with names. Healthcare organizations follow <a href="https://www.hhs.gov/hipaa/index.html">HIPAA</a> and the <a href="https://www.hhs.gov/hipaa/for-professionals/special-topics/hitech-act-enforcement-interim-final-rule/index.html">HITECH Act,</a> reporting breaches of unsecured Protected Health Information under <a href="https://www.hhs.gov/hipaa/for-professionals/breach-notification/index.html">45 C.F.R. 164.400–414</a>. Telecom providers notify under FCC rules when Customer Proprietary Network Information is compromised. Public companies face SEC disclosure requirements for material cybersecurity incidents. Bulk electric power operators report under <a href="https://www.nerc.com/standards/reliability-standards/cip/cip-008-6">FERC Reliability Standard CIP-008–6</a>.</p><p>The real compliance burden, though, sits at the state level. Every state has a breach notification law, starting with <a href="https://law.justia.com/codes/california/code-civ/division-3/part-4/title-1-81/section-1798-82/">California Civil Code 1798.82</a>. These laws apply based on the state of residence of the affected individual, not where the organization is headquartered. A company with customers across 30 states faces 30 sets of rules.</p><p>The variations are substantive, not cosmetic. States define personal information differently. The baseline standard covers a name combined with an unencrypted SSN, driver’s license number, or financial account number with access codes. But California, Arkansas, New York, Vermont, and Washington include biometric data. California, Missouri, Vermont, and Washington include health and medical data. North Dakota includes mother’s maiden name when paired with birthdate. Vermont and California treat login credentials as standalone triggers.</p><p>Harm thresholds diverge too. Most states let organizations skip notification if a formal investigation determines there’s no reasonable likelihood of harm to consumers. California doesn’t offer that safe harbor. If unencrypted personal information was, or is reasonably believed to have been, acquired by an unauthorized person, notification is mandatory regardless of assessed harm.</p><p>The encryption safe harbor is nearly universal but not uniform. Most states exempt organizations from notice if compromised data was encrypted and the encryption key wasn’t also compromised. Wyoming uses redaction standards instead. Tennessee requires compliance with <a href="https://csrc.nist.gov/pubs/fips/140-2/upd2/final">FIPS 140–2</a>.</p><p>Third-party vendor duties add another layer. Vendors holding data for a primary data owner must report breaches to the owner immediately under California law, or within 10 days under Oregon law. Oregon also requires vendors to notify the state Attorney General directly when a breach involves more than 250 consumers, unless the primary data owner handles it first.</p><p>State AG notification triggers vary. California and Illinois require AG notice when breaches exceed 500 residents. Illinois authorizes the AG to publicly publish breach details, including entity names and compromised data categories.</p><p><strong>What the Notice Letter Must Contain</strong></p><p>California’s statute provides a useful template for what state-compliant breach notices require. The letter must explain what happened in plain language, including the date of the breach, the date of discovery, and the basic circumstances. It must specify the categories of personal information compromised. It must describe what the organization is doing: containment measures, forensic investigations, and security enhancements. It must tell affected individuals what they can do to protect themselves, including fraud alerts, credit freezes, and free credit reports. And it must provide contact information for the organization and relevant regulatory resources.</p><p><strong>Rolling Notices and Enforcement Risk</strong></p><p>State statutes impose strict notification windows. Some are subjective (“most expedient time possible and without unreasonable delay”), but several states set hard deadlines: 30 days in Maine, Colorado, and Washington; 45 days in Arizona, Oregon, and Alabama; 60 days in Texas, Louisiana, and South Dakota. Delayed notices are permissible only when law enforcement certifies that notification would impede a criminal investigation.</p><p>The temptation to delay notification while finishing forensic work can be a trap. In a regulatory enforcement action against <a href="https://oag.ca.gov/news/press-releases/attorney-general-bonta-announces-49-million-settlement-kaiser-illegal-disposal">Kaiser, the California Attorney General secured a $150,000</a> settlement because the organization held consumer notices for three months while conducting scope investigation after initially identifying 30,000 compromised records.</p><p>The operational answer is rolling notifications. Instead of waiting until forensic work identifies every affected individual, organizations dispatch formal statutory notices incrementally as batches of verified victims are confirmed. CISOs and IT Leaders, along with legal counsel, coordinate this by establishing a 72-hour batch processing window for forensic victim logs. Legal compliance teams reconcile those batch logs against state AG notification thresholds, like Oregon’s 250-resident trigger and Illinois’ 500-resident reporting requirement, while maintaining strict adherence to 30-to-45-day state statutory windows.</p><p>Rolling notices require discipline. They also require that your forensic team and your legal team have a shared workflow before an incident happens, not one they’re building during it.</p><p><strong>Litigation: Standing, Circuit Splits, and the Important Data</strong></p><p>Federal data breach class actions require plaintiffs to establish constitutional standing under <a href="https://www.hinshawlaw.com/en/insights/blogs/consumer-crossroads-where-financial-services-and-litigation-intersect/navigating-standing-considerations-data-breach-class-actions">Article III</a>. <a href="https://www.reminger.com/report-555">Under Spokeo, Inc. v. Robins (2016)</a>, alleging a statutory violation isn’t enough. Plaintiffs must demonstrate an injury-in-fact that’s concrete, particularized, and actual or imminent, not conjectural. The injury must be traceable to the defendant’s conduct and redressable by a favorable judicial decision.</p><p><a href="https://www.supremecourt.gov/opinions/20pdf/20-297_4g25.pdf">TransUnion LLC v. Ramirez, 141 S. Ct. 2190 (2021)</a>, tightened the standard further. The Court held that a risk of future harm alone doesn’t confer Article III standing for monetary damages claims unless that risk materializes into concrete emotional or financial harm.</p><p>Defendants in breach litigation deploy a four-stage procedural defense. First, they move to compel individual arbitration based on mandatory terms of service. Second, they file Rule 12(b)(1) motions challenging subject matter jurisdiction for lack of standing. Third, they bring Rule 12(b)(6) motions arguing the complaint fails to state a recognized claim. Fourth, they oppose class certification under Rule 23 based on individual variations in damages.</p><p><strong>The Circuit Split on Future Harm</strong></p><p>Federal circuits remain divided on whether an increased risk of future identity theft constitutes a concrete injury when stolen data hasn’t yet been misused.</p><p>Several circuits recognize standing based on risk and mitigation costs. The Sixth Circuit held in <a href="https://law.justia.com/cases/federal/appellate-courts/ca6/15-3386/15-3386-2016-09-12.html">Galaria v. Nationwide Mutual Insurance Co. (2016)</a> that a substantial risk of future harm plus reasonably incurred mitigation costs satisfies Article III. The Seventh Circuit, in decisions including <a href="https://law.justia.com/cases/federal/appellate-courts/ca7/14-3122/14-3122-2015-07-20.html">Remijas v. Neiman Marcus Group (2015)</a> and <a href="https://law.justia.com/cases/federal/appellate-courts/ca7/14-3700/14-3700-2016-04-14.html">Lewert v. P.F. Chang’s (2016)</a>, ruled that a hacker’s intent to steal private data creates a reasonable inference of imminent fraud risk. The Ninth Circuit found in <a href="https://law.justia.com/cases/federal/appellate-courts/ca9/09-35823/09-35823-2011-04-18.html">Krottner v. Starbucks Corp. (2010)</a> and In <a href="https://cdn.ca9.uscourts.gov/datastore/opinions/2018/04/20/16-16860.pdf">re Zappos.com, Inc. (2018)</a> that theft of sensitive unencrypted data creates a credible threat of real and immediate harm. The D.C. Circuit agreed in <a href="https://law.justia.com/cases/federal/appellate-courts/cadc/16-7108/16-7108-2017-08-01.html">Attias v. CareFirst, Inc. (2017)</a> and In <a href="https://law.justia.com/cases/federal/appellate-courts/cadc/17-5117/17-5117-2019-06-21.html">re U.S. Office of Personnel Management (2019)</a>, holding that exfiltration of sensitive data like SSNs is an imminent concrete injury.</p><p>Other circuits take the opposite view. The First Circuit in <a href="https://law.justia.com/cases/federal/appellate-courts/ca1/11-1983/11-1983-2012-02-28.html">Katz v. Pershing, LLC (2012)</a> held that speculative fear of potential future harm fails to establish standing. The Third Circuit in <a href="https://law.justia.com/cases/federal/appellate-courts/ca3/11-1738/111738p-2011-12-12.html">Reilly v. Ceridian Corp. (2011)</a> found hypothetical future misuse relies on a chain of speculation. The Fourth Circuit in <a href="https://law.justia.com/cases/federal/appellate-courts/ca4/15-1395/15-1395-2017-02-06.html">Beck v. McDonald (2017)</a> ruled that stolen hardware without proven intentional data targeting doesn’t meet the imminence threshold. The Eighth Circuit in In <a href="https://caselaw.findlaw.com/court/us-8th-circuit/1872597.html">re SuperValu, Inc. (2017)</a> held that payment card breaches without SSN loss lack identity theft risk. The Eleventh Circuit in <a href="https://law.justia.com/cases/federal/appellate-courts/ca11/18-14959/18-14959-2021-02-04.html">Tsao v. Captiva MVP Restaurant Partners, LLC (2021)</a> found that payment card data loss without immutable identifiers fails standing requirements.</p><p><strong>Which Data Gets You Standing</strong></p><p>The type of data compromised drives standing outcomes more than the fact of a breach. Social Security numbers and birthdates are immutable identifiers. Their compromise creates a substantial risk of identity theft and new-account fraud, satisfying standing requirements across the pro-plaintiff circuits. Payment card numbers alone generally fail to establish standing in pro-defendant circuits because issuing banks cancel compromised numbers and reverse unauthorized charges, eliminating direct out-of-pocket loss.</p><p>Plaintiffs who can’t show actual identity theft have developed two alternative theories. The first is loss of property value in personal information, alleging that PII carries objective commercial value on dark web markets and that the breach diminished that value. Courts in In <a href="https://www.govinfo.gov/app/details/USCOURTS-cand-5_16-md-02752/context">re Yahoo! Customer Data Security Breach Litigation (N.D. Cal. 2017)</a> and In re <a href="https://www.uschamber.com/assets/documents/Opinion-In-re-Marriott-International-Customer-Data-Security-Breach-Litigation-Fourth-Circuit.pdf">Marriott International, Inc. (D. Md. 2020)</a> accepted this theory. The second is loss of the benefit-of-the-bargain, alleging that consumers overpaid because a portion of their payments was intended to cover data security measures the defendant failed to deliver. The Seventh Circuit recognized this theory in <a href="https://law.justia.com/cases/federal/appellate-courts/ca7/17-2408/17-2408-2018-04-11.html">Dieffenbach v. Barnes &amp; Noble, Inc. (2018)</a>.</p><p>For IT Leaders and CISOs, the litigation landscape reinforces a basic truth: the type of data you lose determines how much legal exposure you face. A payment card breach and an SSN breach are different incidents with different litigation profiles. Your data classification and protection priorities should reflect that distinction.</p><p><strong>What This Means of Your Security Program</strong></p><p>The legal architecture of cybersecurity isn’t abstract. It shapes how you build controls, how you respond to incidents, and how well your organization holds up when regulators or plaintiff’s counsel come knocking.</p><p>Public statements about your security posture need to be precise and verifiable. SEC reporting guidance requires transparent disclosure of material cybersecurity risks and operational disruptions. Boilerplate assurances and marketing language about security create real exposure under federal securities fraud statutes when a breach reveals that the reality didn’t match the claims.</p><p>Documentation of security controls isn’t just operational hygiene. Under <a href="https://www.law.cornell.edu/uscode/text/18/1839">18 U.S.C. 1839(3)(A)</a>, your records of access controls, firewall architecture, and logical access limits are the evidence you need to prove trade secret protection. Contemporaneous records of baseline controls are required to demonstrate that the organization took reasonable measures to maintain confidentiality. If those records don’t exist or are incomplete, you may lose the ability to pursue trade secret claims regardless of whether the theft actually occurred.</p><p>Incident response requires a dual-track structure from the start. External digital forensic response teams should be retained through outside legal counsel. IT remediation logs and technical system status reports must be maintained separately from legal risk assessments. This separation preserves attorney-client and work-product privileges under established discovery precedents. Mixing operational documentation with legal analysis in the same communications risks waiving those protections.</p><p>Encryption isn’t optional. Standard-compliant encryption for sensitive data at rest and in transit secures safe harbors under state breach notification statutes. Tennessee’s requirement for <a href="https://csrc.nist.gov/pubs/fips/140-2/upd2/final">FIPS 140–2</a> compliance represents the most demanding standard, and meeting it effectively covers the encryption requirements across all states. The safe harbor holds only if the decryption keys remain uncompromised.</p><p>Rolling notification procedures need to be operational before the incident happens. The pipeline from forensic handlers to legal compliance teams, the 72-hour batch processing windows, the reconciliation against state AG thresholds, and the adherence to fixed statutory windows all require pre-established workflows. Building this infrastructure during an active incident is how organizations end up in enforcement actions like Kaiser’s $150,000 settlement.</p><p>None of this is new information. But the gap between knowing these obligations exist and having operational processes that actually meet them is where legal exposure lives. The organizations that face the worst outcomes aren’t usually the ones with the weakest technical controls. They’re the ones that didn’t connect their security operations to their legal obligations until it was too late.</p><p>I hope you find this post helpful and informative. Thanks for stopping by!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52fa9809b554" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Data Breach Notification Laws Require From You]]></title>
            <link>https://medium.com/@jorge.castaneda13/what-data-breach-notification-laws-require-from-you-4f9754969aa6?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f9754969aa6</guid>
            <category><![CDATA[privacy]]></category>
            <category><![CDATA[privacy-law]]></category>
            <category><![CDATA[data-breach]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 15:06:08 GMT</pubDate>
            <atom:updated>2026-11-02T15:06:08.578Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dXiQCrTO55aK3bwHzbBJzg.png" /></figure><p><strong>The Core Obligation</strong></p><p>Data breach notification laws require organizations that experience an unauthorized acquisition of, or access to, personal information to notify affected individuals and, in most cases, state regulators within a specified timeframe. Every U.S. state, the District of Columbia, and most U.S. territories have enacted their own version of this requirement. Federal sector-specific laws (HIPAA, GLBA, FERPA) layer additional obligations on top of state law for healthcare, financial services, and education, respectively.</p><p>If your organization holds personal information about individuals and that information is compromised, you have a legal obligation to tell people what happened and what they should do about it. Failing to do so exposes the organization to regulatory enforcement, civil litigation, and reputational damage.</p><p><strong>What Triggers the Obligation</strong></p><p>Not every security incident is a notifiable breach. The trigger depends on the jurisdiction, but the general formula is: unauthorized acquisition of, or access to, unencrypted personal information that creates a reasonable risk of harm to the affected individual.</p><p><strong>Elements Control Whether Notification Is Required</strong></p><p><strong>Personal information.</strong> State statutes define this term, and the definitions vary. At a minimum, nearly all states cover a person’s name combined with a Social Security number, driver’s license number, or financial account number with access credentials. Many states have expanded coverage to include medical information, biometric data, email credentials, taxpayer identification numbers, and insurance policy numbers. Some states, such as California, also cover geolocation data and genetic information.</p><p><strong>Unauthorized acquisition or access.</strong> The compromise must involve an unauthorized party actually acquiring or accessing the data. Some states distinguish between acquisition (taking possession) and access (viewing), and a few require only that data was “reasonably believed” to have been accessed. An encrypted laptop lost at an airport may not trigger notification if the encryption meets recognized standards. The same laptop unencrypted almost certainly does.</p><p><strong>Risk of harm.</strong> A growing number of states apply a risk-of-harm analysis. If the organization can demonstrate through a documented investigation that the breach is unlikely to result in harm to individuals, notification may not be required. Other states have no such safe harbor and require notification regardless of assessed risk. Know which model your governing statutes follow.</p><p><strong>Who You Must Notify and When</strong></p><p><strong>Affected individuals.</strong> Every state requires direct notification to the people whose personal information was compromised. The method is typically first-class mail or email, though substitute notice (website posting and media notification) is permitted when the cost of direct notice would exceed a statutory threshold or when the organization lacks sufficient contact information.</p><p><strong>State regulators.</strong> Most states require notification to the state attorney general or another designated agency, either before or concurrent with individual notification. Thresholds vary. Some states require regulatory notice only when the breach affects more than a specified number of residents (500 in many states, 1,000 in others). A handful require regulatory notification for every breach regardless of size.</p><p><strong>Consumer reporting agencies. </strong>When a breach affects a large number of residents in a single state (typically 500 or 1,000+), several states require the organization to notify the major consumer reporting agencies.</p><p><strong>Timing. </strong>Deadlines range from “as expeditiously as possible” with no fixed number of days (the federal HIPAA standard is 60 calendar days) to specific statutory windows. Several states have moved to hard deadlines: 30 days in some jurisdictions, 45 or 60 in others. The clock typically starts when the organization “discovers” or “becomes aware” of the breach, not when the incident itself occurred. Law enforcement may request a delay if notification would impede an active investigation.</p><p>The practical constraint is that an organization operating in multiple states must identify the shortest applicable deadline and treat it as the working timeline. If you have employees, customers, or clients in states with a 30-day window, that is effectively your deadline for everyone, because the operational cost of staggered notifications rarely makes sense.</p><p><strong>What the Notification Must Contain</strong></p><p>Statutory requirements differ, but a compliant notification letter generally includes the following.</p><p>A description of the incident in general terms, including the date or estimated date range. What categories of personal information were involved (Social Security numbers, financial account numbers, medical records, etc.). What the organization has done in response, including steps taken to contain the breach and prevent recurrence. What the affected individual should do to protect themselves, including contact information for the three major credit bureaus and instructions for placing a fraud alert or credit freeze. Contact information for the organization, typically a toll-free number or dedicated email address staffed to handle inquiries. A reference to the individual’s right to file a complaint with the state attorney general or the Federal Trade Commission, where required by statute.</p><p>Some states mandate that the notification include an offer of free credit monitoring or identity theft protection services for a specified period (12 to 24 months is common). HIPAA-covered entities must include specific language about individuals’ rights to file complaints with the U.S. Department of Health and Human Services.</p><p>The notification must be written in plain language. Several states explicitly prohibit including marketing materials or promotional content in breach notifications.</p><p><strong>Variations Across Jurisdictions</strong></p><p>There is no single federal data breach notification law that covers all sectors. The result is a patchwork of 50+ state and territorial statutes, each with its own definitions, timelines, and requirements.</p><p><strong>Key differences to track.</strong></p><p><strong>Definition of personal information.</strong> Some states cover only the traditional triad (SSN, driver’s license, financial account). Others have added biometrics, health data, email credentials, and taxpayer identification numbers. California and Illinois are among the broadest.</p><p><strong>Encryption safe harbor.</strong> Most states exempt encrypted data from notification requirements, provided the encryption key was not also compromised. The standard of encryption that qualifies varies. Some statutes reference “generally accepted” standards without specifying them.</p><p><strong>Notification timeline.</strong> Ranges from 30 days (e.g., Colorado, Florida) to 60 days (most HIPAA-covered entities) to “without unreasonable delay” with no fixed number (the original California model that many states adopted).</p><p><strong>Risk-of-harm threshold.</strong> Some states require notification only if the breach poses a reasonable likelihood of harm. Others require it whenever unauthorized access to personal information occurs, regardless of assessed risk.</p><p><strong>Private right of action. </strong>Several states, including California under the CCPA/CPRA, allow individuals to bring private lawsuits for statutory damages following a breach involving certain categories of personal information. This changes the liability calculus significantly.</p><p><strong>Federal overlay.</strong> HIPAA (healthcare), GLBA (financial services), FERPA (education), and SEC rules (public companies) impose their own requirements that run parallel to state law. An organization subject to both must satisfy all applicable obligations.</p><p>If your organization operates across state lines or handles data from residents of multiple states, you must map which statutes apply to each population of affected individuals and comply with each one. There is no shortcut. A single breach can require compliance with a dozen different statutes simultaneously.</p><p><strong>Consequences of Noncompliance</strong></p><p>Failing to comply with breach notification requirements creates exposure on multiple fronts.</p><p><strong>Regulatory enforcement.</strong> State attorneys general can bring enforcement actions for violations. Penalties vary by state but can reach thousands of dollars per affected individual per day of delay. The FTC has pursued enforcement against organizations that failed to notify consumers or misrepresented their security practices. HIPAA penalties for covered entities can reach $2.067 million per violation category per year (the exact figure is adjusted periodically for inflation).</p><p><strong>Civil litigation.</strong> Affected individuals may bring class action lawsuits alleging negligence, breach of contract, or violations of state consumer protection statutes. In states with a private right of action for data breaches, statutory damages can apply even without proof of actual harm. The cost of defending and settling these actions frequently exceeds the cost of compliant notification by orders of magnitude.</p><p><strong>Regulatory scrutiny. </strong>A botched notification process draws additional regulatory attention to the organization’s overall security posture. Regulators who discover that an organization delayed or avoided notification are more likely to investigate the underlying security program, which can lead to consent decrees, ongoing monitoring requirements, and mandated security improvements.</p><p><strong>Reputational cost.</strong> Late or evasive notifications erode trust with customers, clients, employees, and business partners. The reputational impact of being perceived as having concealed a breach is consistently worse than the impact of the breach itself.</p><p><strong>The calculation is straightforward.</strong> The cost of timely, accurate, legally compliant notification is a known, manageable expense. The cost of noncompliance is open-ended and compounding. Every hour spent debating whether to notify is an hour closer to a deadline you may have already missed.</p><p>I hope you find this post helpful and informative. Thanks for stopping by!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f9754969aa6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You Signed the NDA. Then You Uploaded the Report to AI.]]></title>
            <link>https://medium.com/@jorge.castaneda13/you-signed-the-nda-then-you-uploaded-the-report-to-ai-4d1fc0906d19?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/4d1fc0906d19</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[nda]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 18:02:57 GMT</pubDate>
            <atom:updated>2026-10-24T18:02:57.441Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2lBj7ZP3cjfJ6Z65FfgC3Q.jpeg" /></figure><p>Your company told everyone to use AI. Maybe it came from the CEO during a town hall meeting, maybe it was part of the digital transformation strategy, or maybe IT just rolled out licenses and sent an email to the entire org informing them that employees can now use AI. Either way, the message was: adopt AI, move faster, do more with less. What nobody mentioned was what not to put into it.</p><p>So now you have your risk analyst dropping a vendor’s full SOC 2 Type 2 report into Claude to summarize the exceptions. Your security manager is uploading a supplier’s penetration test results to find the critical findings. Your procurement lead is pasting a vendor’s incident response plan to compare it against a questionnaire. Every one of them is doing exactly what the company asked them to do, and every one of them may have just violated the non-disclosure agreement that gave them access to those documents.</p><p>The employee is not trying to do anything malicious, the employee simply doesn’t understand governance because the company doesn’t have any AI governance in place. The org pushed AI adoption without building the guardrails first, and now people who are trying to do their job well are creating contractual exposure they don’t know exists. They aren’t being careless. They’ve never been told that a vendor’s SOC 2 report isn’t theirs to upload wherever they want, or that the click-through terms on a trust portal create binding obligations, or that their personal Claude Pro subscription handles data in ways that are fundamentally incompatible with the NDA sitting in the contract file.</p><p>The AI tools themselves are genuinely good at this work. They can sift through a 90-page SOC 2 report in minutes, surface the control exceptions and carve-outs that matter, and spot inconsistencies between a vendor’s questionnaire answers and its audited controls that a tired reviewer will miss on a Friday afternoon. The problem isn’t the technology. The problem is that nobody told the people using it where the lines are. The 2026 Verizon Data Breach Investigations Report put shadow AI squarely in scope for SOC 2 audits, and auditors are now asking how vendor-confidential materials are being processed. If your organization doesn’t have an answer, your employees are building one for you, one upload at a time.</p><p>AI tools like Claude are genuinely good at this work. They can sift through a 90-page SOC 2 report in minutes, surface the control exceptions and carve-outs that matter, and spot inconsistencies between a vendor’s questionnaire answers and its audited controls that a tired reviewer will miss on a Friday afternoon. The problem isn’t the technology. The problem is that users are treating confidential vendor materials as if they were their own documents, when in fact they hold them under contractual obligations that were rarely written with AI in mind.</p><p>Those obligations matter. An NDA breach can expose your organization to legal liability, damage a vendor relationship that took years to build, and undermine the trust that makes vendors willing to share sensitive security information at all. If vendors come to believe their reports will end up in someone’s personal AI account, they’ll share less, and your third-party risk program will be weaker for it. The 2026 Verizon Data Breach Investigations Report put shadow AI squarely in scope for SOC 2 audits, and auditors are paying attention. The answer isn’t to ban AI from third-party risk management. It’s to use it deliberately.</p><p><strong>The Contract Exposure Nobody Told Them About</strong></p><p>The contractual question has to be answered before the technical one, because no amount of configuration can cure a disclosure the NDA doesn’t permit. But in most organizations, nobody is asking the question at all. Employees are uploading vendor documents the same day they receive them, and the NDA review that should happen first is happening never.</p><p>Every NDA, confidentiality clause, and access agreement should be reviewed for three provisions before any vendor document touches an AI tool.</p><p>The first is who may receive the information. Many NDAs limit disclosure to “Representatives,” a term usually defined as employees and advisors who have a need to know and are bound by confidentiality obligations. Whether an AI service provider acting as a data processor falls within that definition is often unclear, and many agreements don’t address it at all. Your employees aren’t parsing that distinction. They’re uploading the report.</p><p>The second is how the information may be used. Clauses that limit use to evaluating the business relationship are generally compatible with a risk assessment, but clauses that prohibit processing by third-party tools, automated systems, or artificial intelligence are not. These clauses are showing up more frequently, and your team probably hasn’t noticed.</p><p>The third is what happens when the relationship or assessment ends. Return-or-destroy clauses determine how long vendor content may remain in any system, including an AI platform, and they should drive your retention settings. When nobody has configured those settings, the default retention may violate the clause on its own.</p><p>The first is who may receive the information. Many NDAs limit disclosure to “Representatives,” a term usually defined as employees and advisors who have a need to know and are bound by confidentiality obligations. Whether an AI service provider acting as a data processor falls within that definition is often unclear, and many agreements don’t address it at all.</p><p>Two categories of material deserve particular scrutiny, and they’re exactly the ones employees upload most often. SOC 2 reports carry AICPA restricted-use language that limits their distribution to specified parties. The AICPA reinforced this point in September 2026 when it released Technical Questions and Answers Section 9561, addressing AI’s effect on SOC 1 and SOC 2 examinations and reaffirming the auditor’s obligation to control report distribution. Your analyst uploading that report to a consumer AI account isn’t just violating the NDA. They’re running afoul of the professional standards that govern the report itself.</p><p>Vendor trust portals such as SafeBase, Vanta Trust Center, and Whistic frequently include click-through terms, and those terms increasingly prohibit uploading the content to AI tools. Clicking “accept” creates a binding obligation even when no separate NDA was ever signed. Employees click through these agreements reflexively, download the documents, and upload them to AI without reading a word of the terms they just accepted.</p><p>Where an agreement is silent or ambiguous, the most defensible approach is to request the vendor’s written consent. Going forward, organizations should add language to their NDA templates, contract templates, and questionnaire cover letters that permits processing by approved AI providers operating under commercial terms that prohibit training on customer data. This establishes consent at the outset rather than case by case. Final interpretation of any agreement belongs with your legal counsel.</p><p>The third is what happens when the relationship or assessment ends. Return-or-destroy clauses determine how long vendor content may remain in any system, including an AI platform, and they should drive your retention settings.</p><p>Two categories of material deserve particular scrutiny. SOC 2 reports carry AICPA restricted-use language that limits their distribution to specified parties. The AICPA reinforced this point in September 2026 when it released Technical Questions and Answers Section 9561, addressing AI’s effect on SOC 1 and SOC 2 examinations and reaffirming the auditor’s obligation to control report distribution. If your organization is feeding SOC 2 reports into AI tools without considering the restricted-use designation, you’re on the wrong side of that guidance.</p><p>Vendor trust portals such as SafeBase, Vanta Trust Center, and Whistic frequently include click-through terms, and those terms increasingly prohibit uploading the content to AI tools. Clicking “accept” creates a binding obligation even when no separate NDA was ever signed. Your analysts need to know that.</p><p>Where an agreement is silent or ambiguous, the most defensible approach is to request the vendor’s written consent. Going forward, organizations should add language to their NDA templates, contract templates, and questionnaire cover letters that permits processing by approved AI providers operating under commercial terms that prohibit training on customer data. This establishes consent at the outset rather than case by case. Final interpretation of any agreement belongs with your legal counsel.</p><p><strong>The Subscription Your Employees Are Actually Using</strong></p><p>Here’s what happens in practice. The company says “use AI.” Some employees already have a personal Claude Pro or Max subscription. Others sign up for one with their personal email. They start uploading vendor documents that afternoon. Nobody told them it matters which plan they’re on, so nobody checked.</p><p>It matters. Claude Free, Pro, and Max are consumer plans governed by Anthropic’s Consumer Terms of Service. On those plans, each user controls whether conversations may be used for model training, and data may be retained for up to five years when that setting is enabled. Consumer plans come with no data processing agreement and no organizational controls. Upgrading from Pro to Max increases usage limits, but it doesn’t change how your data is handled. Your employee’s personal subscription is not your organization’s AI tool, and the vendor’s NDA didn’t contemplate its SOC 2 report living in someone’s personal account with five-year retention and potential training use.</p><p>For NDA-protected vendor material, organizations need a commercial offering governed by Anthropic’s Commercial Terms of Service, which don’t permit training on customer content. Claude Team Standard at $25 per user per month provides an acceptable baseline supported by a data processing agreement. Claude Team Premium at $150 per user per month adds enhanced features. Claude Enterprise adds single sign-on, administrative oversight, and custom retention periods. For the most sensitive materials, the Claude API with an approved zero-data-retention arrangement offers the strongest posture, although it requires approval and technical integration.</p><p>But most organizations pushing AI adoption haven’t made this distinction for their employees. They haven’t provisioned commercial accounts. They haven’t told anyone that using a personal subscription for vendor-confidential work is a problem. The gap between “use AI” and “use AI on this plan, configured this way” is where the contractual exposure lives.</p><p>For NDA-protected vendor material, organizations should use a commercial offering governed by Anthropic’s Commercial Terms of Service, which don’t permit training on customer content. Claude Team Standard at $25 per user per month provides an acceptable baseline supported by a data processing agreement. Claude Team Premium at $150 per user per month adds enhanced features. Claude Enterprise adds single sign-on, administrative oversight, and the ability to set a custom retention period, which is valuable when vendor agreements contain destruction obligations. For the most sensitive materials, the Claude API with an approved zero-data-retention arrangement offers the strongest posture, although it requires approval and technical integration.</p><p>Even on commercial plans, the fine print has shifted in 2026 in ways that matter for vendor-confidential work. When Anthropic launched its Mythos-class models in June 2026 under the Covered Models policy, it required 30-day prompt and output retention for safety evaluation purposes, even under zero-data-retention arrangements. Anthropic stated that this data wasn’t used for training and that human access was logged, but the mandatory retention created a contractual wrinkle for organizations whose vendor agreements required immediate destruction.</p><p>Anthropic addressed this in September 2026 with the release of Fable 5.1 and Mythos 5.1 under its Enterprise Frontier Safeguards program. EFS restored zero-data-retention for qualifying enterprise customers, removing the mandatory 30-day hold. If your organization negotiated ZDR before June 2026 and hasn’t revisited the arrangement since, now is the time to confirm whether EFS applies to your account.</p><p>Content flagged by automated systems for usage-policy violations may still be retained for up to two years regardless of your retention settings. And if your staff use desktop agent tools such as Claude Code under an organizational account, session transcripts may be retained under a separate, longer retention schedule unless your organization has negotiated a shorter custom period. None of these exceptions is likely to derail legitimate risk work, but all of them belong in a complete risk assessment. The question is whether anyone in your organization has done that assessment, or whether employees are operating in a policy vacuum.</p><p>Anthropic addressed this in September 2026 with the release of Fable 5.1 and Mythos 5.1 under its Enterprise Frontier Safeguards program. EFS restored zero-data-retention for qualifying enterprise customers, removing the mandatory 30-day hold for organizations that meet Anthropic’s enterprise safety requirements. If your organization negotiated ZDR before June 2026 and hasn’t revisited the arrangement since, now is the time to confirm whether EFS applies to your account.</p><p>Content flagged by automated systems for usage-policy violations may still be retained for up to two years regardless of your retention settings. And if your staff use desktop agent tools such as Claude Code under an organizational account, session transcripts may be retained under a separate, longer retention schedule unless your organization has negotiated a shorter custom period. None of these exceptions is likely to derail legitimate risk work, but all of them belong in a complete risk assessment and should be documented in the data processing agreement review.</p><p><strong>The Vendor You Forgot to Assess and the Settings Nobody Configured</strong></p><p>There’s an irony that most organizations haven’t caught yet: the tool they’re using to assess third-party risk is itself a third party they haven’t assessed. Your AI provider should go through the same TPRM process as any vendor handling confidential information. That means obtaining its SOC 2 Type 2 report and ISO certifications, reviewing its data processing addendum and subprocessor list, and documenting how vendor materials flow through the system. This step demonstrates that your processor meets the standard of care most NDAs require, and it gives you a documented answer when a vendor asks how its SOC 2 report was handled. But when the company’s AI push didn’t include provisioning a commercial plan, it almost certainly didn’t include a vendor assessment of the AI provider either.</p><p>Even when an organization does get the commercial plan right, default settings are rarely the right settings for vendor-confidential work, and nobody is changing them. Access should be controlled through single sign-on and limited to personnel who actually perform third-party risk assessments. Each vendor’s materials should live in a separate project with restricted sharing, so one vendor’s information never informs or appears in another vendor’s assessment. Memory and past-chat search features should be turned off for these accounts, so vendor details don’t carry into unrelated conversations. Connectors, integrations, and artifact publishing should be disabled unless there’s a specific, approved need, because connectors can move content into other systems and published artifacts generate shareable links.</p><p>On Enterprise, the retention period should align with the shortest return-or-destroy obligation in your vendor agreements. When an assessment closes, the vendor’s project should be deleted and the deletion date recorded in the assessment file as evidence of compliance. In practice, none of this is happening unless someone has built the governance structure that makes it happen.</p><p>Commercial plans provide the right foundation, but default settings are rarely the right settings for this use case. Access should be controlled through single sign-on and limited to personnel who actually perform third-party risk assessments. Each vendor’s materials should live in a separate project with restricted sharing, so one vendor’s information never informs or appears in another vendor’s assessment. Memory and past-chat search features should be turned off for these accounts, so vendor details don’t carry into unrelated conversations. Connectors, integrations, and artifact publishing should be disabled unless there’s a specific, approved need, because connectors can move content into other systems and published artifacts generate shareable links.</p><p>On Enterprise, the retention period should align with the shortest return-or-destroy obligation in your vendor agreements. When an assessment closes, the vendor’s project should be deleted and the deletion date recorded in the assessment file as evidence of compliance.</p><p><strong>What Should Have Gone In and What Shouldn’t Have</strong></p><p>Data minimization remains the most effective control, and it’s the one that requires the least organizational overhead to implement. Even when governance is still catching up, individual analysts can reduce exposure immediately. Before uploading materials, they should redact or pseudonymize the vendor’s name and remove IP addresses, hostnames, credentials, network diagrams, and screenshots from penetration test reports.</p><p>Wherever possible, analysts should provide relevant excerpts rather than entire documents. For a SOC 2 report, the auditor’s opinion, control exceptions, complementary user entity controls, subservice organization carve-outs, and management’s responses usually contain the information that drives a risk decision. Raw penetration test findings with exploitable technical detail deserve the most conservative treatment, and a sanitized summary of findings, severity ratings, and remediation status is almost always sufficient to support a risk rating. But without guidance, employees are uploading the whole document every time, because nobody told them not to.</p><p>Wherever possible, analysts should provide relevant excerpts rather than entire documents. For a SOC 2 report, the auditor’s opinion, control exceptions, complementary user entity controls, subservice organization carve-outs, and management’s responses usually contain the information that drives a risk decision. Raw penetration test findings with exploitable technical detail deserve the most conservative treatment, and a sanitized summary of findings, severity ratings, and remediation status is almost always sufficient to support a risk rating.</p><p>There’s an emerging development worth watching that could change this dynamic. In August 2026, SafeBase and Drata both launched MCP-based Trust Center integrations that allow AI tools to connect directly to a vendor’s Trust Center through Anthropic’s Model Context Protocol. These integrations provide scoped, read-only access to vendor security documentation through an authenticated channel that the vendor controls. Instead of downloading a PDF and uploading it to an AI tool, the AI connects to the Trust Center directly, queries specific sections, and works within the access boundaries the vendor has set. The vendor retains control over what’s shared and can revoke access at any time.</p><p>This approach doesn’t eliminate every NDA concern. The AI still processes the vendor’s confidential information, and the underlying commercial-plan and retention considerations still apply. But it removes the manual download-and-upload step that creates the most contractual risk, keeps the vendor in the loop on how its materials are accessed, and provides an audit trail that neither party had before. If your vendors use SafeBase or Drata, ask whether they’ve enabled MCP access. This is the direction the industry is moving, and it’s a path that bypasses the employee-behavior problem entirely by building the access controls into the workflow.</p><p>This approach doesn’t eliminate every NDA concern. The AI still processes the vendor’s confidential information, and the underlying commercial-plan and retention considerations still apply. But it removes the manual download-and-upload step that creates the most contractual risk, keeps the vendor in the loop on how its materials are accessed, and provides an audit trail that neither party had before. If your vendors use SafeBase or Drata, ask whether they’ve enabled MCP access and whether your AI provider supports the integration. This is the direction the industry is moving.</p><p><strong>The Governance That Should Have Come First</strong></p><p>Your AI Acceptable Use Policy should specify which vendor data classifications may be processed, on which platform, and under what conditions. It should also require that a qualified assessor review and validate every AI-generated output before it enters the risk register or reaches stakeholders. If your organization doesn’t have an AI AUP yet, that’s the root cause of everything described in this post.</p><p>AI is excellent at identifying qualified opinions, gaps in bridge-letter coverage, questionnaire answers that contradict audited controls, and incident response plans that fall short of regulatory requirements. It can also misread technical context or miss nuance, so human review is an essential control rather than a formality. Keeping a record of which materials were processed, when, and by whom will further strengthen your position during vendor inquiries and audits.</p><p>AI is excellent at identifying qualified opinions, gaps in bridge-letter coverage, questionnaire answers that contradict audited controls, and incident response plans that fall short of regulatory requirements. It can also misread technical context or miss nuance, so human review is an essential control rather than a formality. Keeping a record of which materials were processed, when, and by whom will further strengthen your position during vendor inquiries and audits.</p><p>The organizations that told everyone to use AI weren’t wrong. They were incomplete. The speed and analytical power of AI-assisted third-party risk management is real, and organizations that don’t adopt it will fall behind. But adoption without governance is just exposure with extra steps. Review the NDA before anything gets uploaded. Provision commercial accounts and stop pretending personal subscriptions are enterprise tools. Assess your AI provider the way you’d assess any vendor handling confidential data. Configure the environment to enforce the boundaries your agreements require. Tell your analysts what to redact and what to excerpt. Put a qualified human between the AI’s output and your risk register. And write the AI Acceptable Use Policy that should have existed before anyone uploaded the first document.</p><p>Your employees aren’t the problem. They’re doing what you asked them to do. The problem is that you asked them to do it without telling them where the lines are.</p><p><em>This post reflects my professional analysis and does not constitute legal advice. Vendor agreements, AI provider terms, and regulatory requirements vary. Consult qualified legal counsel before making compliance decisions based on this content. Information about Anthropic’s products, pricing, and policies is current as of October 2026 and subject to change.</em></p><p>I hope you find this post helpful and informative. Thanks for stopping by!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4d1fc0906d19" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Attackers Exploit Your Inbox]]></title>
            <link>https://medium.com/@jorge.castaneda13/how-attackers-exploit-your-inbox-531e565d044e?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/531e565d044e</guid>
            <category><![CDATA[cybersecurity-awareness]]></category>
            <category><![CDATA[email-security]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 13:58:02 GMT</pubDate>
            <atom:updated>2026-10-18T15:53:45.167Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mmqghL3fZoPFme8vFzoDYQ.jpeg" /></figure><p>Since October is Cybersecurity Awareness Month, I wanted to write about a common threat vector that organizations always warn their employees about, not just during Cybersecurity Awareness Month, but throughout the entire year.</p><p>Our inbox is one of the most legally and operationally significant surfaces in our organization. It’s where contracts get negotiated, instructions get issued, and communication moves. It’s also where most attacks start. Threat actors know this, and they’ve spent years refining their techniques to match the way professional communication actually looks.</p><p>The advice I always give: “don’t click suspicious links and NEVER trust and ALWAYS verify”, yet this advice can fail because modern phishing doesn’t look suspicious anymore. A well-crafted attack arrives as a plausible invoice, a routine HR notice, or a message that looks exactly like your CEO’s writing style. Telling people to stay alert isn’t a defense strategy; it’s more of a wish now.</p><p>What works is understanding how attackers actually get in. There are several primary vectors: malicious attachments, weaponized links, and pure social engineering through text alone. Each one exploits a different gap in your defenses, and each one demands a different countermeasure.</p><p><strong>Attachments: The Delivery Mechanism</strong></p><p>A file attached to an email looks like a file. It could be a Word document, a PDF, a zip archive: the kind of thing you’d receive from a client or a colleague without thinking twice. What the recipient doesn’t see is that the file is functioning as a delivery container for something else entirely.</p><p>The most persistent technique here is the macro exploit. You open a Word or Excel file, and it shows a blurred preview with a banner asking you to click “Enable Content.” The moment you do, a hidden script fires, connects out to an attacker-controlled server, pulls down a payload, and establishes a presence on your machine. The document is just the wrapper.</p><p>Attackers also hide payloads inside compressed archives: .ZIP or .ISO files, because the outer container isn’t the thing being scanned. They use shortcut files (.LNK) that carry executable code but display a document icon. Some PDFs embed scripts that silently redirect to exploit sites. One of the more elegant evasion techniques is the password-protected archive. The attacker locks the file, puts the decryption key in the email body, and watches automated scanners fail to look inside. Security tools can’t read what’s encrypted; people can. The human becomes the unintentional decryption layer.</p><p>A subset of attachment attacks carries no malware at all. Opening the file launches a convincing login page (a replica of your Microsoft 365 portal, your firm’s VPN, your cloud storage provider) and waits for you to type in your credentials. No exploit needed. You hand them over voluntarily.</p><p><strong>Technical Countermeasure</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/851/1*wiCr94bO-WbBtOpENpED8Q.png" /></figure><p>One practical note: CDR and aggressive file-type blocking have real operational consequences. Finance teams run complex macros in Excel. Law firms use PDFs with embedded form fields. Security settings calibrated in isolation will create friction for the people doing legitimate work. Effective attachment security requires ongoing collaboration with the business units affected, not a one-time configuration decision.</p><p><strong>Links: The Misdirection Layer</strong></p><p>If attachments are about hiding what a file does, links are about hiding where they go. Plenty of phishing campaigns skip file delivery entirely and send a URL that takes you straight to a convincing replica of a cloud portal, a VPN login, or a file-sharing service. By the time you realize the page isn’t real, you’ve already typed in your password.</p><p>The manipulation techniques here are varied and creative. Attackers register domains that look nearly identical to real brands, sometimes using characters from other alphabets that appear visually identical to Latin letters (a technique called a homoglyph attack). They embed malicious destinations inside long, cluttered URLs that are hard to parse at a glance, or use URL shorteners to hide the final destination entirely.</p><p>Some campaigns exploit the trust reputations of legitimate platforms. They build fake login forms hosted on real cloud storage services, survey tools, or compromised content management systems. Because the hosting domain is legitimate, basic reputation filters let it through. Others use multi-stage redirects: the link in the email points to a clean page, which then dynamically forwards you to a malicious destination based on your device, location, or browser fingerprint. The initial URL scans as safe. The final destination isn’t.</p><p>QR codes, what the industry calls quishing, add another layer. An image embedded in the email body carries a QR code. When you scan it with your phone, the request bypasses every desktop email filter entirely. Your phone isn’t on the corporate network. It isn’t running enterprise security tools. The link lands in a browser with no protection between you and the attacker’s page.</p><p><strong>Click-Time Protection</strong></p><p>The fundamental challenge with links is that their destination can change after the email is delivered. An attacker can send a link that scans as clean, wait for it to reach inboxes, and then flip the destination to a malicious page. Scanning at delivery time misses this entirely.</p><p>Click-time protection addresses this by rewriting every URL in the email. When you click, the request routes through a security service that inspects the destination in real time: it expands shortened URLs, follows redirect chains to their final endpoint, and checks the actual landing page for signs of credential harvesting or brand impersonation. The URL you clicked and the URL you land on are treated as two different questions, because they often are.</p><p><strong>Text Alone: Business Email Compromise</strong></p><p>Business Email Compromise (BEC) is in a different category. There’s no attachment. There’s no link. It’s a text message, written in plain language, that impersonates someone with authority and asks you to do something: transfer funds, share sensitive documents, update payment routing information. The request appears to come from your CEO, your CFO, your managing partner, or an established vendor.</p><p>This attack vector is particularly expensive. The FBI’s Internet Crime Complaint Center has reported multi-billion-dollar annual losses attributed to BEC globally, and those numbers likely undercount the actual impact because many incidents go unreported due to reputational concerns. Law firms and professional services organizations are high-value targets precisely because they handle large financial transactions and have access to confidential client information.</p><p>Technical controls struggle against BEC because the message itself is legitimate text. There’s no malware signature to match, no suspicious domain to flag. A filter doesn’t know that your CFO is traveling and wouldn’t send a wire transfer request at 11 PM. It doesn’t know your firm’s standard operating procedure for vendor payment changes. Context is the defense, and context lives with people.</p><p><strong>Defense Requires Both Technical and Human Layers</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/846/1*MdL3Av3r9k6e4aSwCbvOcQ.png" /></figure><p><strong>What Happens When Something Goes Wrong</strong></p><p>No filter catches everything. The question isn’t whether a phishing email will eventually reach someone. It’s how fast you detect and contain it when it does. That requires a reporting process your people will actually use.</p><p>Incident Triage Flow:</p><ul><li>User reports via one-click buttonFull message headers attach automatically. No technical steps required from the reporter.</li><li>Analyst triages: benign, malicious, or uncertainUncertain defaults to malicious until proven otherwise.</li><li>Mass remediation across all mailboxesConfirmed threats get purged organization-wide and the sending domain gets blocked at the gateway.</li><li>Close the loop with the reporter”That was real, and we removed it.” Every confirmed report that gets acknowledged is an incentive for the next one.</li></ul><p>The reporting mechanism matters. If your people have to forward a suspicious email to a generic IT inbox, remember a specific email address, or manually copy header information, most of them won’t. A one-click “Report Phishing” button integrated directly into the email client removes all friction. Clicking it should automatically package the full message headers without requiring the reporter to do anything technical.</p><p>Closing the feedback loop is equally important and often skipped. When someone reports a legitimate threat, they should hear back that it was real and that the security team acted on it. That confirmation turns a passive user into an active participant in your detection process. It also builds the kind of trust that makes people report the next one instead of quietly second-guessing themselves.</p><p>Training supports all of this, but “training” shouldn’t mean an annual hour-long video. Short, frequent sessions — fifteen minutes, four times a year — with realistic examples drawn from actual campaign types work better. Fake invoices, package tracking updates, executive requests, HR notices. When you run phishing simulations, use the data to identify which departments and which message types are most effective against your organization, and tailor your follow-up to those gaps. Shame is not a security control.</p><p><strong>Measuring Whether It’s Working</strong></p><p>You don’t need elaborate dashboards to track email security effectiveness. A few consistent metrics tell you most of what you need to know. Gateway metrics (volume of malicious attachments and links blocked per month) show you campaign activity and give you a baseline for spotting unusual spikes. Endpoint and identity alerts (malware detections on devices, unusual authentication events) highlight the cases where email controls failed, and something got through. Human metrics (report volume, triage resolution time, simulation results by department) tell you whether your users are functioning as part of the detection layer or not.</p><p>The value of integrating these signals is significant. If a user lands on a credential harvesting page and enters their password, that event is visible to your identity systems. A connected environment can trigger an immediate password reset or force multi-factor re-authentication before the attacker has time to use what they collected. The email filter didn’t stop the click, but the downstream systems contained the damage.</p><p><strong>Building Out The Stack Over Time</strong></p><p>Mature email security doesn’t require everything at once. It builds in layers, and the early layers create the foundation the later ones depend on.</p><p><strong>Baseline controls.</strong> Standard spam filtering and domain authentication: SPF, DKIM, and DMARC properly configured. This stops direct spoofing and filters the high-volume commodity attacks.</p><p><strong>User reporting.</strong> A single, integrated one-click reporting mechanism that automatically captures message headers. This turns your user base into a detection asset.</p><p><strong>Triage playbook.</strong> A documented process for security teams to categorize reported emails and execute mass remediation quickly when a threat is confirmed.</p><p><strong>Targeted training.</strong> Short, frequent sessions with realistic examples. Simulation results used to improve training, not to rank or shame employees.</p><p><strong>Advanced technical layers.</strong> Attachment sandboxing, CDR, and click-time link protection. These tools address the sophisticated evasion techniques that baseline filters miss.</p><p><strong>Cross-functional review.</strong> Regular sessions with Security, IT, Finance, and HR to refine workflows. The BEC threat in particular requires operational processes that cut across departments.</p><p><strong>The Mental Model That Holds It Together</strong></p><p>We must treat email security as a set of overlapping controls: technical, procedural, and cultural, that each address a different part of how attacks actually work.</p><p>Attachments are delivery containers. Block high-risk formats at the perimeter, use sandboxing to observe behavior in isolation, and strip active code through CDR before it reaches the user’s device. Links are misdirection. Expand shortened URLs, follow redirect chains to their final endpoint, and apply click-time inspection so that a destination that changes after delivery still gets checked. Message bodies are pressure. Authenticate the sending domain with SPF, DKIM, and DMARC, verify payment changes through a separate channel, and build an environment where slowing down to ask a question is treated as good judgment rather than inefficiency.</p><p>None of these layers is sufficient on its own. All of them together, supported by metrics that track whether they’re working, give you something close to a real defense.</p><p>I hope you find this post helpful and informative. Thanks for stopping by!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=531e565d044e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Shodan And How Does It Work?]]></title>
            <link>https://medium.com/@jorge.castaneda13/what-is-shodan-and-how-does-it-work-f06dc2e01190?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/f06dc2e01190</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[sto]]></category>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[shodan-search-engine]]></category>
            <category><![CDATA[shodan]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Wed, 14 Oct 2026 00:41:31 GMT</pubDate>
            <atom:updated>2026-10-14T00:58:29.560Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hrqc7HbEeHr2ZAnw" /></figure><p><strong>What is Shodan?</strong></p><p>For anyone starting out in cybersecurity, learning how assets are exposed online is one of the most fundamental skills you can develop. When most people think of a search engine, they picture platforms like Google or Bing, which are indexers built to catalog web page text, HTML documents, and media file contents. Shodan, however, operates on an entirely different plane.</p><p>Shodan is a specialized search engine designed to discover and index internet-connected devices rather than web pages. Developed in 2009 by computer programmer John Matherly out of San Diego, Shodan was created to continuously sweep global network infrastructure. While traditional search engines send out web crawlers to read text content by following hyperlinks, Shodan sends out port requests to find servers, webcams, routers, industrial hardware, and smart infrastructure.</p><p>If Google helps you locate information published on the web, Shodan helps you locate the underlying physical and virtual hardware that forms the internet itself. Demystifying this tool is an essential milestone for novice security practitioners. It encourages a mental shift from seeing the internet as a simple collection of websites to viewing it as a vast, interconnected ecosystem of live hardware devices.</p><p><strong>How Does It Work?</strong></p><p>To understand how Shodan operates, you must examine its underlying scanning mechanics. The global IPv4 address space contains roughly 4.3 billion individual addresses. Shodan constantly probes this expanse by sending initial connection requests across dozens of standard TCP and UDP ports. These range from standard web channels like Port 80 (HTTP) to remote management endpoints like Port 3389 (RDP) and specialized operational hardware ports like Port 502 (Modbus).</p><p>When Shodan contacts an open port, the connected host returns a metadata packet known as a “banner.” A banner is a service response containing technical details about the system. Depending on how the host is configured, this banner can reveal operating system details, exact software build numbers, hardware model lines, system unit IDs, serial numbers, and internal IP address schemes (such as 192.168.x.x). Shodan collects these returned banners, organizes them, and indexes them into a searchable database accessible at shodan.io. Users must always ensure they navigate to shodan.io rather than the .com domain, as the .com domain is unassociated with the project and is frequently flagged or blocked by web browsers as a security risk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/737/1*-ZpAAb0MefMwwqZWiGVbVQ.png" /></figure><p>Because scanning billions of addresses across thousands of service ports requires significant time and computing power, Shodan data is typically a few weeks old. Search results reflect what was online when the crawler last visited that specific IP. As a senior practitioner, you must treat Shodan as an intelligence repository rather than a real-time monitor. Active verification, such as conducting an authorized, targeted Nmap scan or direct port check, is always required to confirm if a service remains accessible in real time.</p><p>Modern security workflows often automate this discovery process. Beyond the web interface, security teams and researchers interact with Shodan programmatically using command-line interface (CLI) tools or custom Python API scripts. Executing commands like shodan init with an API key, shodan search for automated queries, or shodan stats --facets allows defenders to programmatically sweep subnets and integrate exposure data directly into their internal vulnerability management platforms.</p><p><strong>How Cybersecurity Professionals Use Shodan for Defense</strong></p><p>Penetration testers, security analysts, and defenders rely on Shodan for passive reconnaissance. Active reconnaissance, such as running an Nmap sweep directly from your testing machine against a target company, sends active network packets across the wire. This noisy traffic alerts firewalls, triggers Intrusion Detection Systems (IDS), and generates log entries on the target host. Passive reconnaissance via Shodan allows you to inspect an organization’s public attack surface safely and silently, querying historical data without generating a single packet of traffic to the target machine.</p><p>Defenders structure their passive reconnaissance workflows through specific analytical phases:</p><h4>Subnet and Perimeter Auditing</h4><p>Defensive teams begin by identifying their organization’s public IP footprint. By using the net: search operator with Classless Inter-Domain Routing (CIDR) notation (e.g., net:153.156.222.0/24), analysts can pull every indexed IP address, open port, and service banner associated with their network blocks. Similarly, the hostname: filter allows defenders to discover exposed IP addresses tied directly to corporate domain names.</p><h4>Identifying Misconfigurations and End-of-Life Systems</h4><p>Security analysts routinely search for legacy operating systems and unpatched software that remain exposed to the public internet. Queries targeting unsupported operating systems, such as os:&quot;Windows XP&quot; (which reveals over 5,700 connected instances worldwide) or os:&quot;Windows 7&quot; (which uncovers over 42,000 public instances), highlight assets that no longer receive manufacturer security patches. Analysts also query for outdated database deployments, such as unpatched MySQL installations running on Port 3306 (including legacy builds like version 5.6.5 or 5.7.42) or SQL Server running on Port 1433, ensuring legacy systems are removed from public exposure.</p><h4>Uncovering Insecure Access Points</h4><p>Directly exposed management protocols create immediate vectors for unauthorized entry. Defenders search Shodan for exposed Remote Desktop Protocol (RDP) screens on Port 3389, VNC instances operating without authentication (identified by returned status codes like rfb003008), open Server Message Block (SMB) file shares on Port 445 indicating authentication disabled, or anonymous FTP servers operating on Port 21 (which return status code 230 upon successful anonymous login).</p><h4>Deploying Honeypots for High-Fidelity Defense</h4><p>Security teams also use Shodan concepts to design threat intelligence sensors known as honeypots. Frameworks like Conpot or Cowrie simulate vulnerable services, industrial controllers, or operating systems to attract scanning traffic. Publicly exposed honeypots allow defenders to observe attacker methodologies, exploit scripts, and scanning frequencies. When deployed on internal networks, honeypots serve as high-fidelity intrusion detection sensors. Because legitimate users have no operational reason to contact a dummy host, any connection attempt to an internal honeypot generates an immediate, high-confidence alert indicating active internal lateral movement. However, defenders must configure honeypots carefully. If a decoy host enables every available industrial protocol simultaneously, Shodan and experienced threat actors will easily fingerprint the anomalous signature, tag it as a decoy (tag:honeypot), and bypass it.</p><p><strong>How Hackers Exploit Shodan for Target Discovery</strong></p><p>From an offensive perspective, threat actors use Shodan to gather intelligence on potential targets without alerting defensive monitoring tools. By leveraging passive queries, attackers identify vulnerable entry points, missing patches, and unauthenticated management screens across the globe.</p><h4>Protocol Targeted Scanning and Facility Infrastructure</h4><p>Malicious actors routinely target standard operational ports associated with Industrial Control Systems (ICS), Operational Technology (OT), and critical facility infrastructure. These targeted ports include Modbus (Port 502), BACnet for building automation (Port 47808), Ethernet/IP (Port 44818), Siemens S7 (Ports 102 and 161), DNP3 for electrical utilities, and Niagara Fox or Tritium protocols. Beyond factory PLCs, attackers search for critical data center and facility support hardware. Equipment such as uninterruptible power supplies (UPS systems manufactured by Eaton, APC, or Tripp Lite), Power Distribution Units (PDUs), Automatic Transfer Switches (ATS), and Computer Room Air Conditioning (CRAC) units are frequently connected to networks via web management cards. If an attacker disables a facility’s CRAC units or trips an ATS remotely, they can shut down an entire server facility without writing a single line of malware.</p><h4>Hardware Vendor and Model Filtering</h4><p>To locate systems vulnerable to specific exploit payloads, attackers filter queries by vendor names and product lines. Searching for vendors such as Schneider Electric (including Modicon PLCs), Siemens (including SIMATIC controllers and SCALANCE S615 routers), Rockwell Automation / Allen-Bradley, or Honeywell yields granular device lists. Banners returned from these queries leak exact firmware versions, equipment serial numbers, and internal network IP schemes (such as 192.168.100.x). This granular metadata provides threat actors with the exact specifications required to select a matching exploit module.</p><h4>Visual Interface Reconnaissance and Default Credentials</h4><p>Visual queries offer another shortcut for target discovery. Using images.shodan.io or queries like screenshot.label:ics or screenshot.label:webcam, attackers browse indexed visual captures of exposed Human-Machine Interfaces (HMIs), industrial control dashboards, and facility webcams. Exposed webcams watching solar farms, water treatment tanks, or transportation hubs provide visual confirmation of physical target layouts. Once exposed management portals are located, attackers attempt access using factory default administrative credentials (such as admin/admin or vendor defaults like apc/admin on UPS management cards).</p><h4>Navigating Search Noise: False Positives and Facet Analysis</h4><p>Navigating Shodan effectively requires understanding search noise and data distribution. A common trap for beginners is searching raw string keywords like ICS. Querying ICS directly does not return a clean list of industrial control systems. Instead, it frequently returns software libraries like Delphi’s <em>Overbyte Internet Component Suite</em> (Overbyte ICS TCP Server), creating false positives. Skilled analysts bypass this noise using Facet Analysis. By using the view report feature in the web interface or running command-line queries like shodan stats --facets, analysts break down large datasets across specific property categories such as os, port, org, or country. Facet breakdowns allow practitioners to summarize global dataset distributions instantly without manually filtering through thousands of irrelevant search results.</p><p><strong>Industrial Control Systems and IoT Risks</strong></p><p>Directly connecting smart devices, facility infrastructure, and industrial control hardware to the open internet creates severe cyber-physical risks. Many legacy industrial protocols were designed decades ago for air-gapped, isolated internal networks. Consequently, standard protocols like Modbus or BACnet lack native encryption, password authentication, or packet verification mechanisms. When these systems are attached to internet-facing IP addresses, any remote host that can reach the port can read operational registers or transmit control commands directly to the hardware.</p><p>The primary operational danger stems from lateral movement. A peripheral, seemingly minor asset, such as an internet-exposed IoT security camera, an environmental controller, or a building HVAC sensor at a water treatment plant, solar farm, or refinery, can serve as an initial landing pad. If an attacker breaches the peripheral camera and discovers it sits on a dual-homed network shared with operational technology, they can pivot through that peripheral device to reach critical Programmable Logic Controllers (PLCs) or safety instrumented systems.</p><p>The real-world impact of these compromises extends far beyond digital data theft. Manipulating physical industrial hardware can result in immediate physical service outages, catastrophic equipment failure, environmental chemical contamination, or severe safety hazards for plant operators and surrounding communities.</p><p>These critical risks expose several critical operational technology (OT) vulnerabilities, starting with inherent protocol security deficits in legacy OT systems that lack native authentication and encryption, which enables unauthorized remote users to execute control commands if systems are exposed. Also, peripheral entry points, such as unsecured IoT devices, webcams, and facility management cards, offer easy footholds for attackers to pivot laterally into internal IT and OT networks. This type of unrestricted control system exposure carries severe cyber-physical consequences, potentially resulting in physical equipment destruction, prolonged utility blackouts, environmental harm, and significant real-world safety hazards.</p><p><strong>Ways You Can Secure Your Perimeter</strong></p><p>Securing an organization against passive discovery requires continuous hygiene, clear banner management, and strict network isolation. Implementing the following defensive steps will significantly reduce your visible attack surface:</p><p><strong>Conduct Routine Self-Audits on Shodan:</strong> Regularly query Shodan using your organization’s public IP subnets (net: operator with CIDR notation) and domain names (hostname: operator). Identifying your exposed assets first allows you to close accidental exposures before external actors exploit them.</p><p><strong>Isolate Internal OT, ICS, and Facility Infrastructure:</strong> Operational technology, PLCs, HMIs, and data center support hardware (UPS, PDU, ATS, CRAC units) must never be exposed directly to the public internet. Isolate all operational hardware behind strict firewall boundaries, isolated VLANs, and complete physical air gaps where necessary.</p><p><strong>Require Encrypted VPNs with Multi-Factor Authentication:</strong> Eliminate direct public access for remote administrative interfaces, including RDP, VNC, HTTP management portals, and Telnet. Force all remote management traffic through an encrypted Virtual Private Network (VPN) secured by Multi-Factor Authentication (MFA).</p><p><strong>Harden Banners and Change Default Credentials:</strong> Immediately change all factory default administrative usernames and passwords upon deployment. Disable unauthenticated guest access on file shares (SMB, FTP), and sanitize service banners across all public servers to prevent leaking exact operating system, software, and firmware build numbers.</p><p><strong>Implement Continuous Monitoring and Internal Sensors:</strong> Establish automated scripts or API queries to alert your security team whenever new IP addresses or open ports within your subnets appear on Shodan. Deploy single-purpose honeypots internally as early warning sensors to detect unauthorized scanning or lateral movement within your internal network boundaries.</p><p>Shodan is fundamentally a neutral mirror reflecting the raw reality of what is connected to the global internet. It does not introduce software vulnerabilities, nor does it alter the operational state of exposed hardware. It simply collects, organizes, and indexes the service banners that internet-connected devices continuously broadcast to the world.</p><p>For anyone pursuing a career in cybersecurity, learning how to navigate and interpret Shodan is an essential early step. By understanding and practicing passive reconnaissance, you can locate misconfigurations, identify unsupported end-of-life systems, and remediate exposed access points long before malicious actors can exploit them. Adopting this defensive, practitioner-focused mindset is key to building a secure digital infrastructure.</p><p>I hope you find this post helpful and informative. Thanks for stopping by!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f06dc2e01190" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Vendor Risk Changes Faster Than The Approval Process]]></title>
            <link>https://medium.com/@jorge.castaneda13/ai-vendor-risk-changes-faster-than-the-approval-process-dab6362a14a6?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/dab6362a14a6</guid>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-privacy]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 21:45:58 GMT</pubDate>
            <atom:updated>2026-09-13T21:45:58.736Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/487/1*wdDE3Ako2bxIkLjie4Q8qg.png" /></figure><p>It’s easy to look at the recent coverage of Anthropic’s legal plugin and assume it is mainly a concern for legal teams. But that misses the bigger issue. Around the same time, Anthropic introduced similar capabilities for sales, marketing, and data analysis. The legal plugin received more attention, but the underlying risk applies to almost any organization that has already approved an AI vendor for a specific use case.</p><p>The issue is scope.</p><p>A vendor may have been reviewed and approved for one purpose, but over time it adds new features, integrations, plugins, or connectors that significantly expand what the platform can access. Those new capabilities often fall under the original vendor approval instead of triggering another security or privacy review.</p><p>That creates a problem. The original data flow may no longer be accurate. New integrations could provide access to sensitive or regulated information that was never part of the initial assessment. The vendor may not have done anything wrong. The organization’s vendor risk and change management processes simply may not have been designed to catch this type of change.</p><p>Consider healthcare. An AI platform originally approved for internal documentation could later introduce an integration with a system containing protected health information. Before enabling that integration, the organization needs to determine whether the appropriate safeguards are in place and whether its Business Associate Agreement covers the new data flow. From a HIPAA perspective, what matters is how PHI is being handled, not whether the new functionality was considered a minor feature update.</p><p>Financial services organizations face a similar issue. A new integration could expose customer financial information covered by the Gramm-Leach-Bliley Act Safeguards Rule or records subject to SEC or FINRA retention requirements. Payment card information creates another concern. If cardholder data could enter the workflow, the organization needs to verify the vendor’s current PCI DSS status before allowing that use case.</p><p>Privacy reviews need to go beyond whether customer information is used to train the model. Anthropic’s commercial terms may provide protections around the use of customer data for model training, but that is only one part of the review. Organizations still need to understand retention, subprocessors, incident response obligations, data residency, and how information moves through any newly introduced integrations.</p><p>Those answers should come from the current contract, Data Processing Addendum, security documentation, and subprocessor information, not simply from a product announcement or marketing page.</p><p>None of this means that an AI plugin somehow bypasses GDPR, CCPA, HIPAA, NIST CSF, or other security and privacy requirements. The real issue is that these rapidly changing AI platforms can expose weaknesses in traditional vendor risk management.</p><p>Most vendor assessments are performed at a point in time. The vendor is reviewed, the contract is signed, the risk is documented, and the platform is approved. Six months later, however, the same platform may have capabilities and integrations that did not exist when the assessment was completed.</p><p>That is where frameworks such as NIST CSF, ISO 27001, and CIS Controls become important. Supply chain and third-party risk management should not end when the contract is signed. Organizations need a way to identify meaningful changes in what their vendors can access and what they can do with organizational data.</p><p>There is also a vendor concentration issue that deserves more attention.</p><p>When the company providing the underlying AI model also begins offering specialized applications and integrations on top of that model, organizations may have more dependency on a single provider than their risk register shows. Two applications that appear to be separate vendors or services may ultimately depend on the same foundation model or infrastructure.</p><p>That changes the concentration risk.</p><p>The practical response is straightforward, although implementing it can be difficult: treat significant new capabilities from an existing AI vendor as a change-management trigger, not simply as a product update.</p><p>When an AI vendor introduces a new plugin, connector, integration, or major capability, reopen the assessment. Review the data flow. Determine what information the new capability can access. Map that access against applicable regulatory and contractual requirements. Review the vendor’s current security and privacy documentation.</p><p>Most importantly, make sure the approval reflects what the platform does today, not what it was capable of doing when the organization originally approved it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dab6362a14a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding Quiet AI Failures and Governance]]></title>
            <link>https://medium.com/@jorge.castaneda13/why-do-we-have-ai-governance-frameworks-378148394383?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/378148394383</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[governance-framework]]></category>
            <category><![CDATA[aigovernanceframework]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 16:58:28 GMT</pubDate>
            <atom:updated>2026-08-29T17:01:49.388Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ydJu0Xv7wOZ3XPqRJtaBSA.jpeg" /></figure><p>Let&#39;s be honest, what do you think of when you think of AI failure? Most people picture AI failure as something dramatic like a chatbot saying something offensive, a self-driving car missing a stop sign. In practice, the failures that do the most damage are quiet. A model keeps running, keeps producing answers that look reasonable, and slowly drifts away from reality or starts reinforcing a bias it was never supposed to have. Nobody notices until the pattern shows up in a lawsuit, or a very public news story.</p><p>This is the problem an AI governance framework is actually built to solve. Not the flashy failures. The slow ones. Understanding how these systems break, and where human oversight has to sit in the process, is useful even if you never touch a line of code. Legal, compliance, and risk teams are the ones who end up explaining these failures to a client, a board, or a regulator, so it helps to know what’s actually going on under the hood. I’m doing my best to keep up with AI, and I’ve done some online training from several platforms, and much of what I’ve learned and understood is why I decided to write this post.</p><p><strong>Why AI Models Fail Quietly, Not Loudly</strong></p><p>An AI model isn’t programmed step by step the way traditional software is. It learns patterns from data, which means its behavior is only as good as what it was trained on and how well it holds up once it meets the real world.</p><p>A few things go wrong repeatedly. Models can be brittle, meaning they handle the situations they were trained on fine but fall apart the moment something unusual shows up. A predictive maintenance model, for example, might be trained almost entirely on data from equipment running normally. When a rare failure mode actually occurs, the model has never seen anything like it and misses the warning signs completely.</p><p>Training data itself is often the root cause. If the data reflects old biases, contains errors, or just doesn’t represent the population the model will actually serve, the model inherits those blind spots. Sometimes the problem isn’t the data at all but the goal the model was optimized for. A chatbot built to maximize subscription conversions can end up using manipulative prompts or pressure tactics to hit that number, technically succeeding while working against the user’s actual interest.</p><p>Then there’s automation bias, which is a human problem more than a technical one. People get used to trusting a model’s output and start rubber-stamping its recommendations, even when something looks off. Combine that with teams that only test against clean benchmark data instead of stress-testing under adversarial or unusual conditions, and you get systems that pass every internal review and still fail once they’re live.</p><p><strong>The Bias Loop Nobody Notices Until It’s Too Late</strong></p><p>One of the more troubling patterns is the self-reinforcing feedback loop, where a model’s own output changes the world in a way that then gets fed back into the model as new training data.</p><p>For example, a model trained on historically biased data makes a prediction that reflects that bias. Someone acts on that prediction, which changes real-world behavior. That changed behavior gets logged as new data. The model gets retrained on that new data and treats it as objective proof that its original prediction was correct, so it makes the same call again, only with more confidence.</p><p>Predictive policing is the textbook example. If arrest records already reflect uneven policing in certain neighborhoods, a model trained on that data will predict more crime in those same areas and send more officers there. More officers on the ground naturally means more arrests, and those arrests get logged right back into the system as confirmation that the original prediction was accurate. The cycle repeats and gets stronger each time, even though nothing about actual crime rates necessarily changed.</p><p>Content recommendation systems run into a version of this too. When an algorithm is optimized purely for engagement, it learns that emotionally charged or extreme content keeps people scrolling. It shows more of that content, users interact with it more, and the system doubles down. The result is an echo chamber the algorithm built almost by accident, just by chasing the metric it was told to optimize.</p><p><strong>When the Ground Shifts Under a Model</strong></p><p>Even a well-built model degrades over time simply because the world keeps moving and the model doesn’t, unless someone retrains it. This is called model drift, and it has a few different forms.</p><p>Covariate drift happens when the type of data coming into the model changes, even though the underlying relationship between that data and the correct answer hasn’t. Picture a retail model trained on the buying habits of one customer group. If a completely different demographic starts using the app, the inputs look very different from what the model trained on, even though the basic psychology of what makes someone buy something hasn’t actually changed.</p><p>Concept drift is different and often more dangerous. Here the inputs can look exactly the same as before, but what they actually mean has changed. Fraud detection is the classic case. A transaction can have the exact same technical fingerprint as a legitimate one from years ago, but fraud tactics evolve, and that same fingerprint might now indicate an actual scam. The rules of the game changed even though nothing about the raw data did.</p><p>There’s also label drift, where the overall rate of an outcome shifts across the board, like a broad economic downturn suddenly pushing loan default rates up across an entire portfolio, independent of anything about individual borrowers. And a plainer example that ties all of this together: a customer service chatbot trained on pre-pandemic language patterns can start misunderstanding customers within a year or two, simply because the way people talk and what they’re asking about has moved on without it.</p><p>Left unmonitored, drift acts like a slow leak. The model passed every test before launch and still quietly gets worse every month it runs.</p><p><strong>Keeping a Human in Charge</strong></p><p>Because models can fail in ways that are hard to catch from the outside, governance frameworks build in specific levels of human oversight, and the right level depends on how much is at stake.</p><p>Human-in-the-loop, usually shortened to HITL, means a person has to sign off before the system’s recommendation becomes a real decision. This is the standard for high-stakes calls like credit underwriting, hiring screens, or medical diagnostics, where a specialist reviews the model’s suggestion before anything happens.</p><p>Human-on-the-loop, or HOTL, is looser. The system runs on its own and makes decisions in real time, but a person is actively watching the outputs and has the authority to step in, override a result, or shut the system down if something looks wrong. This shows up in high-volume environments like content moderation or logistics routing, where waiting for a human to approve every single action isn’t realistic.</p><p>A third layer, human-in-command, sits above both. This is where the big, structural calls live: setting the thresholds that trigger an incident review, deciding when to pull a system offline entirely, owning the rollback plan. It’s not about approving individual outputs; it’s about who has final authority over the system as a whole.</p><p><strong>The Risk Categories Governance Is Actually Built For</strong></p><p>When people talk about “AI risk” as one big blur, it’s worth breaking it into the categories governance frameworks actually track, because each one has a different owner and a different failure mode.</p><p>Operational risk is what happens when a business leans too hard on a model that then fails or behaves unpredictably, like an inventory system glitch causing real supply shortages. Legal and regulatory risk comes from laws like the <a href="https://artificialintelligenceact.eu/high-level-summary/">EU AI Act</a> (NOTE: the link provides a high-level summary of the EU AI Act) and how they treat personal data, licensing, and individual rights; getting this wrong can mean fines and forced recalls. Financial risk covers everything from the compute costs of retraining a model going over budget to an unsupervised trading algorithm misreading the market and losing real money fast.</p><p>Cybersecurity risk is its own category with AI-specific threats: prompt injection, where someone manipulates a model’s input to make it do something it shouldn’t; model poisoning, where corrupted data gets slipped into training; and model inversion, where an attacker reverse-engineers a model to extract the data it was trained on. Supply chain risk shows up because most organizations don’t build their own models from scratch; they use third-party or open-source ones, which means someone else’s data breach or licensing mistake becomes your legal exposure too. And reputational risk is the one everyone remembers: a discriminatory hiring model or an unfair credit denial makes headlines, and the backlash lands on the organization that deployed it, not just the vendor who built it.</p><p><strong>Why a Model Has to Be Able to Explain Itself</strong></p><p>A lot of AI governance comes down to one basic requirement: if a model makes a decision that affects someone, that person needs to be able to understand why, and ideally, contest it.</p><p>It helps to separate three terms people often use interchangeably. Transparency is an organization’s overall ability to document and audit how its systems work. Interpretability is more technical, referring to whether developers can actually trace how the model got from input to output. Explainability is the practical version of all this: can the organization give a person a plain-language reason for the specific decision that affected them?</p><p>This matters most with deep neural networks, which are often described as black boxes because their internal logic isn’t something a human can just read off. In high-stakes areas like loan approvals or tenant screening, that opacity becomes a real legal problem, not just a technical inconvenience. Governance frameworks push teams toward tools like SHAP and LIME, which score how much each input feature actually mattered to a specific output, or in the highest-risk cases, toward simpler and more interpretable models entirely, like decision trees, instead of a neural network no one can fully explain.</p><p>None of this works without knowing where the data came from in the first place. Data provenance and lineage tracking let a team trace a bad output back to the specific data or transformation step that caused it. And on the front end, users need to be told plainly when they’re talking to an AI system, what it’s for, and how to reach an actual person if they need to.</p><p><strong>Getting a Model Ready to Go Live</strong></p><p>Moving a model from development into production isn’t a single step, it’s a checklist that touches nearly every department in an organization.</p><p>It usually starts with a model card, a document that lays out what the model is for, how it performed in testing, where its training data came from, and what its known limitations are. From there, the model goes through stress testing, not just against clean benchmarks but against adversarial inputs and edge cases designed to break it on purpose. Legal and privacy teams run an AI Impact Assessment or a Data Protection Impact Assessment to flag potential harms like bias or privacy exposure before launch, and a separate audit confirms the training data was actually licensed and collected lawfully.</p><p>Depending on where the model will run, engineers may need to shrink it down for edge devices through techniques like quantization or distillation, or add retrieval-augmented generation to cut down on made-up answers. Oversight controls get built in at this stage too, whether that’s HITL approval gates, HOTL monitoring, or a human-in-command escalation path for emergencies.</p><p>Before anything goes live, there’s a formal change management submission and a rollback plan, so the organization can revert to the last safe version if something goes wrong. Rather than releasing to everyone at once, most teams use a staged or canary rollout, sending the model to a small slice of users first and watching the numbers closely. The last gate is a multidisciplinary sign-off: legal, engineering, business, cybersecurity, privacy, and governance all have to agree the model is ready before it reaches full production.</p><p>None of these steps exist to slow a project down for its own sake. They exist because the failures described earlier, the bias loops, the drift, the opacity, are hard to see from the inside once a system is live. The purpose of a governance framework is to catch that.</p><p>I hope you find this post helpful and informative. Thanks for stopping by</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=378148394383" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What is Data Loss Prevention?]]></title>
            <link>https://medium.com/@jorge.castaneda13/what-is-data-loss-prevention-c8bb1325d76b?source=rss-fbd41c6aa29f------2</link>
            <guid isPermaLink="false">https://medium.com/p/c8bb1325d76b</guid>
            <category><![CDATA[data-classification]]></category>
            <category><![CDATA[data-privacy]]></category>
            <category><![CDATA[dlp]]></category>
            <category><![CDATA[data-loss-prevention]]></category>
            <dc:creator><![CDATA[Jorge Castaneda, D.Eng (c), MLS]]></dc:creator>
            <pubDate>Sun, 23 Aug 2026 14:15:09 GMT</pubDate>
            <atom:updated>2026-08-23T14:15:09.610Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*EzR4mpvujE2wJOymjkpbvQ.png" /></figure><p>Picture your company’s sensitive information as water, and your company as the container holding it. Now picture that container with a few small cracks nobody has noticed yet: a shared folder left open to anyone with the link, an email sent to the wrong “John” in someone’s contacts, a lost or stolen laptop. Every one of those cracks lets information leak out, and once it’s out, there’s no calling it back.</p><p>That’s the problem data loss prevention, or DLP, tries to solve. DLP isn’t a single piece of software or some rules; it’s a company-wide approach to figuring out where sensitive information lives, understanding how it could get out, and putting the right safeguards in place before it does. I’ll do my best to briefly explain what DLP means, why a program is important for companies, and the basic building blocks: data types, classification levels, and sensitive information types, that make a DLP program work.</p><p><strong>The meaning of Data Loss Prevention</strong></p><p>Strip away the acronym and DLP is a simple idea: know what information matters, know where it lives, and stop it from ending up somewhere it shouldn’t.</p><p>Information tends to leak out in one of two ways.</p><p><strong>By accident</strong>, which is the far more common scenario. Someone attaches the wrong file to an email, an employee sets up a new file-sharing link and forgets to restrict who can open it, or a contractor copies a project folder onto a personal laptop and never deletes it.</p><p>Nobody meant any harm in those cases, but the information still got out.</p><p><strong>On purpose</strong>, when the loss is intentional. A departing employee copies a client list before their last day, someone with a grudge sends internal documents to a competitor, or an outsider who has stolen a password uses it to quietly pull files out of the company’s systems.</p><p>The intent is different in each case, but the outcome looks the same from the outside: sensitive information sitting somewhere it was never supposed to be.</p><p>A DLP program is built to catch both kinds of leaks.</p><p>It also helps to be clear about what DLP is not. It isn’t one piece of software you install and forget about, and it isn’t a single locked-down rule that blocks everything and hopes for the best. A real DLP program combines people (who need to know what counts as sensitive and why), process (how information is supposed to be handled day to day), and technology (systems that watch for the risky moments and step in). Leave out any one of those and the others end up doing all the work alone.</p><p><strong>Why every company needs a DLP Program</strong></p><p>It’s tempting to think DLP is only for hospitals, banks, and giant corporations with something obviously valuable to protect. In practice, nearly every company holds information that would cause real damage if it got into the wrong hands: customer names and addresses, employee records, vendor contracts, pricing details, product plans that haven’t launched yet.</p><p><strong>A DLP program is important for a few plain reasons.</strong></p><p><strong>Trust.</strong> Customers and partners share information with a company because they expect it to be handled carefully. A leak, even a small one, damages that trust in a way that’s hard to earn back.</p><p><strong>Obligation.</strong> Many companies are legally required to protect certain categories of information, health details, payment information, and other personal data among them. A DLP program is part of how a company demonstrates it’s meeting those obligations, rather than hoping nothing goes wrong.</p><p><strong>Continuity.</strong> A serious data loss incident doesn’t just cause embarrassment. It pulls people away from their actual jobs to investigate, notify affected customers, and clean up the aftermath. A DLP program is what keeps a bad afternoon from turning into a bad quarter.</p><p>Think about what actually happens after a leak, even a small one. Someone has to figure out exactly what got out and who it affects; customers or partners may need to be told directly, and legal or communications teams often get pulled in.</p><p>Meanwhile, the work that was supposed to happen that week doesn’t. A DLP program exists to make that scenario the exception rather than something the company is quietly waiting to happen.</p><p>None of this requires guessing where the risk sits. It starts with understanding the two basic shapes that company data comes in.</p><p><strong>Structured vs. Unstructured Data</strong></p><p>Before anyone can protect information, they need to understand what form it actually takes. Nearly everything a company stores falls into one of two categories.</p><p><strong>Structured data</strong> is information that lives in a tidy, organized system, the kind of data that fits neatly into rows and columns. A customer database, a payroll system, an inventory spreadsheet, and a financial ledger are all structured data. It has a predictable shape and usually lives in one well-defined place, which makes it comparatively easier to locate and protect, because everyone generally knows where it is.</p><p><strong>Unstructured data</strong> is everything else, and in most companies, it’s the majority of what gets created day to day. Emails, chat messages, Word documents, PDFs, meeting notes, scanned contracts, and voice memos are all unstructured data. It doesn’t sit in one central system; it’s scattered across inboxes, laptops, chat threads, and personal cloud drives. That’s exactly why unstructured data is where most accidental leaks happen: nobody set out to lose it; it simply lives in too many places to watch closely without help.</p><p>A few quick examples make the difference concrete:</p><ul><li><strong>Structured:</strong> a spreadsheet of customer account numbers, a CRM record of a sales deal, a database table of employee salaries</li><li><strong>Unstructured:</strong> an email thread discussing a client’s account, a shared document listing new-hire details, a screenshot of a contract pasted into a chat message</li></ul><p>This distinction is important to understand because structured and unstructured data need different kinds of attention. Structured data can often be protected by locking down one system and controlling who has access to it. Unstructured data requires watching activity itself, what gets emailed, uploaded, copied, or shared, because the information isn’t sitting still in one place to begin with.</p><p>Understanding which type of data a company is dealing with is the first step. The next is deciding how sensitive that information actually is, and that’s where classification comes in.</p><p><strong>Data Classifications</strong></p><p>Not all company information deserves the same level of protection, and treating everything as equally sensitive makes it harder, not easier, to protect what actually matters. Most companies sort information into three tiers.</p><p><strong>Public data</strong> is information that’s meant to be seen by anyone. Marketing materials, published blog posts, press releases, and job postings all fall here. There’s no downside to this information being widely seen, and that’s the point of publishing it.</p><p><strong>Internal data</strong> is meant to stay inside the company, even though nothing catastrophic happens if it slips out. Internal memos, meeting notes, project timelines, and general business communications live in this tier. A competitor seeing an internal meeting agenda is awkward, not damaging.</p><p><strong>Confidential data</strong> is the tier that needs the tightest protection, the information that would cause real, lasting harm if it ended up in the wrong hands. Customer personal details, health records, financial account information, employee social security numbers, trade secrets, and unreleased product plans all belong here.</p><p>Classification only works when it’s practical, not theoretical. A useful way to find the confidential tier quickly is to ask straightforwardly: what specific piece of information would hurt this company most if it leaked tomorrow? The answer to that question is almost always confidential data, and it’s the right place to focus protection first.</p><p>Classification tells a company what deserves the strictest handling. The next piece is making sure that intention actually gets enforced automatically, rather than depending on someone remembering to be careful.</p><p><strong>Sensitive Information Types</strong></p><p>Here’s the practical problem with classification alone: a company can decide that “confidential data” needs strict protection, but nobody has time to read every document, email, and spreadsheet by hand to check what’s inside it. This is where sensitive information types come in.</p><p>A sensitive information type, often shortened to SIT, is a reusable definition that lets a system automatically recognize a specific kind of sensitive information, without a person having to read the document first. Think of it as a template for “what this kind of sensitive data looks like,” so a system can scan a file or an outgoing email and flag it the moment it spots a match.</p><p>Common examples of sensitive information types include:</p><ul><li>A social security number</li><li>A credit card or bank account number</li><li>A passport or driver’s license number</li><li>A company’s own custom identifiers, such as employee ID numbers or internal project code names</li></ul><p>Sensitive information types are what turn classification from a good intention into something that actually gets caught. “Protect confidential data” is a policy. “Automatically flag any outgoing email that contains what looks like a social security number” is something a system can actually act on. That’s the real value of building out sensitive information types, and they’re the difference between a DLP program that exists on paper and one that catches a leak before it happens.</p><p>Getting the definitions right also matters for a quieter reason: a definition that’s too broad ends up flagging harmless, everyday work, and employees quickly learn to tune out or work around a system that constantly cries wolf. A well-built sensitive information type finds the real thing without burying it in false alarms, which is what keeps a DLP program useful instead of becoming background noise nobody trusts.</p><p><strong>Putting the Pieces Together</strong></p><p>Data loss prevention isn’t one tool or one rule; it’s the combination of everything covered here: understanding how information leaks, knowing whether it’s structured or unstructured, classifying it as public, internal, or confidential, and using sensitive information types to make that classification enforceable. None of these pieces does much on its own.</p><p>A classification scheme with nothing to enforce it is just a document nobody reads. A detection system with no classification behind it is watching for everything and nothing at once. Together, the pieces are what keep a company’s most important information where it belongs.</p><p>Start with one question: what specific piece of information would hurt your business most if it got out? Find that answer, figure out where it lives, decide which tier it belongs in, and you’ve already identified the first thing your DLP program needs to protect.</p><p>I hope you find this post helpful and informative. Thanks for stopping by!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c8bb1325d76b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>