<?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:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[aiaudited.fyi]]></title><description><![CDATA[AI governance, risk, and  controls—what matters, what to test, and how to prove it.]]></description><link>https://www.aiaudited.fyi</link><image><url>https://substackcdn.com/image/fetch/$s_!5vDs!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F50e657a1-d9e0-43c3-b63f-ab55bc7ccf48_608x608.png</url><title>aiaudited.fyi</title><link>https://www.aiaudited.fyi</link></image><generator>Substack</generator><lastBuildDate>Sat, 21 Nov 2026 18:22:48 GMT</lastBuildDate><atom:link href="https://www.aiaudited.fyi/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Glorin Sebastian]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[glorinsebastian@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[glorinsebastian@substack.com]]></itunes:email><itunes:name><![CDATA[Glorin Sebastian]]></itunes:name></itunes:owner><itunes:author><![CDATA[Glorin Sebastian]]></itunes:author><googleplay:owner><![CDATA[glorinsebastian@substack.com]]></googleplay:owner><googleplay:email><![CDATA[glorinsebastian@substack.com]]></googleplay:email><googleplay:author><![CDATA[Glorin Sebastian]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[AI, Audited #004 — The Internet Has Safety Rules. The UK Wants Receipts.]]></title><description><![CDATA[Under the UK&#8217;s Online Safety Act, Ofcom can ask social media platforms for the data behind their safety measures. What would demonstrate that those measures are actually working?]]></description><link>https://www.aiaudited.fyi/p/ai-audited-004-the-internet-has-safety</link><guid isPermaLink="false">https://www.aiaudited.fyi/p/ai-audited-004-the-internet-has-safety</guid><dc:creator><![CDATA[Glorin Sebastian]]></dc:creator><pubDate>Sat, 07 Nov 2026 21:23:24 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!dBtA!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p style="text-align: justify;">A platform can count every post it removes. It can report how it classified those posts and how quickly it acted. But when Ofcom asked for exposure data, two providers couldn&#8217;t tell it how many UK users under 18 had seen the content they acted on.</p><p style="text-align: justify;">That&#8217;s the difference between tracking what a safety team did and measuring what users experienced. </p><p style="text-align: justify;">Ofcom is the UK&#8217;s communications regulator. Under the Online Safety Act, it oversees safety duties for social media platforms, search services, and other regulated services used in the UK. Three recent developments show what that oversight looks like in practice.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!dBtA!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!dBtA!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png 424w, https://substackcdn.com/image/fetch/$s_!dBtA!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png 848w, https://substackcdn.com/image/fetch/$s_!dBtA!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png 1272w, https://substackcdn.com/image/fetch/$s_!dBtA!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!dBtA!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png" width="1200" height="1000" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1000,&quot;width&quot;:1200,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:108089,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://www.aiaudited.fyi/i/223943557?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!dBtA!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png 424w, https://substackcdn.com/image/fetch/$s_!dBtA!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png 848w, https://substackcdn.com/image/fetch/$s_!dBtA!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png 1272w, https://substackcdn.com/image/fetch/$s_!dBtA!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F33cd8bd6-8f5e-46ff-9b9b-2a739a42e07e_1200x1000.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>October 22&#8212;the report</h2><p style="text-align: justify;">Ofcom published its first statutory report on content harmful to children under section 63 of the Act. It found no statistically significant change overall in children&#8217;s self-reported exposure to harmful content over the first nine months after the children&#8217;s safety duties took effect in July 2025.</p><p style="text-align: justify;">In an earlier Ofcom survey, 73% of 11- to 17-year-olds recalled encountering at least one type of harmful content from a list in the previous four weeks. Among respondents asked about one encounter, 35% said they came across it in their feed, and 1% said they found it by searching.</p><p style="text-align: justify;">The report looks at how often children encounter harmful content and how serious the harm can be. Whether individual services are complying with their duties and whether the measures in Ofcom&#8217;s codes are effective are separate questions. Ofcom recommended adding body stigma content and depression content to the Act&#8217;s priority category for content harmful to children.</p><h2>October 30&#8212;the court case</h2><p style="text-align: justify;">The High Court rejected challenges by Meta, WhatsApp, X and TikTok to Ofcom&#8217;s information notices. Ofcom issued the notices in February 2026 under section 100 of the Act, seeking data on moderation, user reports, appeals, and UK users&#8217; exposure to content the platforms took action on. The requests covered data going back to January 2024, with three submission deadlines in June 2026, September 2026, and March 2027.</p><p style="text-align: justify;">The court refused permission to proceed on seven grounds and rejected the proportionality and human rights challenges on their merits. The ruling confirms that Ofcom can require information for legitimate regulatory purposes such as monitoring and research.</p><p style="text-align: justify;">The October 22 report already drew on data collected under the February notices, alongside Ofcom&#8217;s survey and other research. The survey showed little overall change in reported exposure. The platform data showed how hard it can be to measure that exposure consistently, even when services can count the content they acted on. The next batch of data, due March 25, 2027, will give Ofcom another period to track.</p><h2>November 2&#8212;the codes</h2><p style="text-align: justify;">Updated Ofcom codes for user-to-user services took effect, adding crisis response measures. The codes recommend written protocols for large services at medium risk and for services of any size at high risk for specified illegal harms or for specified types of content harmful to children. The children&#8217;s code applies to services that children are likely to access.</p><p style="text-align: justify;">A protocol should spell out which signals the provider monitors, who responds, and how the team acts. Providers should conduct and record a review when the crisis ends or 90 days after it began, whichever comes first. The codes also recommend a dedicated law enforcement contact channel for certain large services during a crisis. Providers can use and document alternative measures that meet the underlying legal duties.</p><h2>What this means if you run a service in scope</h2><ol><li><p style="text-align: justify;"><strong>Make your data retrievable.</strong> Ofcom can ask for operational data as part of routine monitoring and research. The January 2024 lookback applied to these specific notices. Keep the records behind moderation, complaints and appeals timestamped, reconcilable and easy to produce.</p></li><li><p style="text-align: justify;"><strong>Rehearse the crisis protocol.</strong> Run a tabletop exercise before you need it. Record who made each decision, where the response stalled and what you changed afterward.</p></li><li><p style="text-align: justify;"><strong>Report exposure alongside activity.</strong> A removal count tells you how much content your team acted on. Time to action shows how quickly your team acted. Exposure tells you how many users saw it before the action took effect.</p></li><li><p style="text-align: justify;"><strong>Assess risks before significant changes.</strong> Ofcom is investigating whether Meta adequately assessed the risks before launching Instagram Instants in May 2026. The investigation is ongoing. Build the risk assessment into the launch process so the team can act on it before release.</p></li></ol><h2>What to watch</h2><ul><li><p><strong>March 25, 2027.</strong> The third batch of data requested under February notices is due.</p></li><li><p style="text-align: justify;"><strong>The fees case.</strong> Meta&#8217;s separate challenge to Ofcom&#8217;s method for calculating supervision fees and maximum penalties was heard October 13&#8211;14. The statutory maximum is the greater of &#163;18 million or 10% of qualifying worldwide revenue.</p></li><li><p style="text-align: justify;"><strong>Category 1 appeals.</strong> Meta is challenging the Category 1 designations for Instagram and WhatsApp. That designation carries additional duties.</p></li><li><p><strong>The under-16s ban.</strong> The UK government plans to restrict under-16s&#8217; access to certain social media services from spring 2027.</p></li></ul><p><strong>If the UK asked for the receipts tomorrow, what would yours prove?</strong></p><h2>Sources</h2><ul><li><p><a href="https://caselaw.nationalarchives.gov.uk/ewhc/admin/2026/2779">High Court judgment on Ofcom&#8217;s information notices, October 30</a></p></li><li><p><a href="https://www.ofcom.org.uk/siteassets/resources/documents/online-safety/protecting-children/content-harmful-to-children/report-on-content-harmful-to-children.pdf?v=426852">Ofcom&#8217;s October 22 report on content harmful to children, including Annex 2</a></p></li><li><p><a href="https://www.ofcom.org.uk/siteassets/resources/documents/online-safety/research-statistics-and-data/protecting-children/childrens-online-experiences-research-report.pdf?v=418198">Ofcom&#8217;s Children&#8217;s Online Experiences survey</a></p></li><li><p><a href="https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/statement-crisis-response-protocol">Ofcom&#8217;s crisis response statement and amendments to both codes</a></p></li><li><p><a href="https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/ofcom-investigates-metas-compliance-with-uk-online-safety-laws">Ofcom&#8217;s Instagram Instants investigation announcement</a></p></li><li><p><a href="https://www.judiciary.uk/judgments/meta-platforms-v-ofcom-order/">High Court scheduling order for the fees case</a> and <a href="https://www.reuters.com/legal/government/meta-challenges-uk-online-safety-watchdog-over-how-fees-fines-are-calculated-2026-10-13/">Reuters coverage of the hearing</a></p></li><li><p><a href="https://www.barsacross.com/meta-ofcom-uk-whatsapp-instagram-online-safety-act">Coverage of the Instagram and WhatsApp Category 1 appeals</a></p></li><li><p><a href="https://www.gov.uk/government/publications/fact-sheet-new-rules-to-protect-children-online/fact-sheet-new-rules-to-protect-children-online">UK government fact sheet on the planned under-16s restrictions</a></p></li><li><p><a href="https://www.legislation.gov.uk/ukpga/2023/50/contents">Online Safety Act 2023</a></p></li></ul>]]></content:encoded></item><item><title><![CDATA[DORA, Year Two: Your AI Vendor Goes Down. Can You Prove Resilience?]]></title><description><![CDATA[DORA, year 2: testing whether documented controls work when a critical AI service fails.]]></description><link>https://www.aiaudited.fyi/p/dora-year-two-your-ai-vendor-goes</link><guid isPermaLink="false">https://www.aiaudited.fyi/p/dora-year-two-your-ai-vendor-goes</guid><dc:creator><![CDATA[Glorin Sebastian]]></dc:creator><pubDate>Sun, 25 Oct 2026 02:19:48 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!-Y_y!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>The Digital Operational Resilience Act (Regulation (EU) 2022/2554) has applied to EU financial entities since 17 January 2025. The ESAs&#8217; first annual incident report counted 3,383 major ICT-related incidents in 2025, and the first 19 critical ICT third-party providers were designated on 18 November 2025. This article maps DORA requirements to recommended audit tests, evidence, and warning signs, illustrated through a hypothetical AI fraud-service outage.</p><h2><strong>DORA in one paragraph</strong></h2><p style="text-align: justify;">DORA Regulation (EU) 2022/2554 is the EU&#8217;s framework for ICT risk in the financial sector. It has applied directly across the EU since 17 January 2025 to the Article 2 financial entities (banks, payment and e-money institutions, investment firms, insurers, crypto-asset service providers, market infrastructures, and others), reaching their ICT providers through contracts and, for designated providers, through direct EU-level oversight. Its five pillars are ICT risk management; incident management, classification, and reporting; resilience testing; third-party risk; and information sharing.</p><h2><strong>The central question</strong></h2><p><strong>Can the firm keep processing payments safely when its AI fraud service fails and can it prove that fallback works?</strong></p><h2><strong>The argument for year two</strong></h2><p style="text-align: justify;">In 2026, the question is whether the controls work. That is this article&#8217;s argument, not a claim about supervisory behavior. DORA&#8217;s first year produced the paperwork layer: frameworks, registers, remediated contracts, and playbooks. Its second year produces something new for auditors: a baseline. Financial entities reported 3,383 major ICT-related incidents in 2025, according to the European Supervisory Authorities&#8217; (ESAs) first annual report under Article 22(2), published 3 June 2026. The ESAs designated the first 19 critical ICT third-party providers (CTPPs) under Article 31(1)(a) on 18 November 2025, with the list published under Article 31(9), drawn from the registers firms themselves filed.</p><p style="text-align: justify;">Neither figure alone proves a supervisory shift. Together they provide auditors a testable surface: incident data for checking classification and reporting discipline, and a designation list for checking registers and contracts. The ESAs caution against reading incident volume as a weakness, and that caution points to the right standard: assess incident patterns alongside detection, containment, recovery, and remediation.</p><h2><strong>The five pillars in brief</strong></h2><p style="text-align: justify;">ICT risk management (Arts. 5&#8211;6, a management-body-owned framework reviewed at least annually, or periodically for microenterprises, and after major incidents); incident management, classification, and reporting (Arts. 17&#8211;20); resilience testing (Arts. 24&#8211;27, including TLPT for identified entities); third-party risk (Arts. 28&#8211;31); and information sharing (Art. 45  voluntary to join, with participant obligations including notifying authorities when joining or leaving and sharing safeguards).</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!-Y_y!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!-Y_y!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png 424w, https://substackcdn.com/image/fetch/$s_!-Y_y!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png 848w, https://substackcdn.com/image/fetch/$s_!-Y_y!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png 1272w, https://substackcdn.com/image/fetch/$s_!-Y_y!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!-Y_y!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png" width="1456" height="1213" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1213,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:400466,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://www.aiaudited.fyi/i/221945509?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!-Y_y!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png 424w, https://substackcdn.com/image/fetch/$s_!-Y_y!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png 848w, https://substackcdn.com/image/fetch/$s_!-Y_y!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png 1272w, https://substackcdn.com/image/fetch/$s_!-Y_y!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95428644-5ffd-46fc-b747-fdb02884fbf3_2400x2000.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><div><hr></div><h2><strong>The audit overlay: requirement &#8594; test &#8594; evidence &#8594; warning sign</strong></h2><p style="text-align: justify;">The graphic maps DORA requirements to recommended audit tests, evidence, and warning signs. These procedures are practical suggestions, not additional legal requirements. Participation in Article 45 information-sharing arrangements is voluntary; if a firm chooses to participate, it must meet the applicable notification and safeguarding obligations.</p><h2><strong>The AI twist: three thresholds, not one status</strong></h2><p>The sentence to get right: <strong>an AI service can already fall within DORA&#8217;s ICT third-party requirements.</strong> First assess scope (Article 2) and whether the service supports a critical or important function, then place the arrangement:</p><ul><li><p style="text-align: justify;"><strong>Threshold 1: ordinary ICT-service requirements.</strong> Any ICT third-party arrangement: due diligence and risk assessment, a register entry (Article 28), and a contract with the Article 30(2) baseline terms for services and data locations, security requirements, incident-assistance duties, cooperation with authorities, termination rights and notice periods, and security-awareness and training participation.</p></li><li><p style="text-align: justify;"><strong>Threshold 2: enhanced requirements for services supporting critical or important functions.</strong> Article 30(3) adds extensive audit, inspection, and access rights for the firm <em>and</em> its competent authority; participation in the firm&#8217;s testing, including TLPT; business-continuity testing; a documented exit strategy with a transition period; and conditions and visibility for sub-outsourcing.</p></li><li><p style="text-align: justify;"><strong>Threshold 3: designation as a critical provider.</strong> The ESAs designate providers as critical under Article 31(1)(a); the list is published and updated under Article 31(9). The November 2025 list of 19 is historical; it does not by itself establish today&#8217;s provider count. Designation brings direct EU-level oversight by a lead overseer, a status the provider holds, not something the firm negotiates.</p></li></ul><p style="text-align: justify;">Nothing in DORA says &#8220;AI.&#8221; An AI fraud-scoring service in a payments flow needs no AI-specific rule: if it supports a critical function, Threshold 2 already applies, with controls proportionate to the function&#8217;s criticality and the dependency&#8217;s risk. The worked example tests exactly that.</p><h2><strong>Worked example: the fraud-model outage</strong></h2><p><em>Hypothetical exercise, constructed to illustrate the method.</em></p><p>Meridian Payments, a fictional payment institution, relies on an AI vendor&#8217;s fraud-scoring API to clear card transactions. On a Saturday peak, the API returns errors, then stale scores, leaving the fraud-scoring service fully or partially unavailable for a measured 3.5 hours. Meridian&#8217;s playbook switches to a rules-based fallback and queues borderline transactions for manual review. Payments keep flowing. Here is how an auditor tests whether that story is resilience or luck.</p><p style="text-align: justify;"><strong>Classification with actual thresholds.</strong> A degraded control at peak warrants assessment; it does not by itself establish a major incident. Meridian&#8217;s worksheet does the assessment: the gateway is met (a critical function, payments, is affected), and three RTS 2024/1772 Article 9 thresholds are met  transactions (affected transactions exceeded 10% of the daily average number for the affected service); downtime (the service was fully or partially unavailable for a measured 3.5 hours, above the 2-hour threshold for critical-function services); geographical spread (impact in two Member States). Two qualifications matter. The counted transactions were those delayed or otherwise impaired, held in the manual-review queue with clearance times above baseline. Transactions successfully rerouted and cleared normally are not counted as impaired; successful rerouting alone does not establish impairment. And the 3.5 hours is the measured period of full or partial unavailability, the threshold input, not the vendor&#8217;s overall disruption window, which may be longer or shorter. Gateway plus thresholds met: the incident is classified as major, and the worksheet shows its working. Had the outage lasted minutes with negligible transaction impact, the same worksheet could have concluded &#8220;not major&#8221; the method, not the outcome, is the point.</p><p style="text-align: justify;"><strong>The clock with its starting points.</strong> The initial notification is due as early as possible, within four hours of classifying the incident as major, and generally no later than 24 hours after the firm becomes aware of it; where classification as major happens later, the rule for later classification applies, and the initial notification is due within four hours of that classification. The intermediate report is due within 72 hours of the initial notification, even if nothing has changed. Updated intermediate reports must follow without undue delay, including when regular activities have recovered. The final report is due within one month of the intermediate report or of the latest updated intermediate report. The auditor&#8217;s test is mechanical: demand the awareness, classification, and submission timestamps, and check each interval against its own starting point. Where the incident affects clients&#8217; financial interests, the firm must also inform those clients without undue delay.</p><p style="text-align: justify;"><strong>The fallback evidence.</strong> The activation log and queue data establish activation and throughput, not safety by themselves. Safety evidence adds fraud losses versus baseline, false approvals, reconciliation errors, manual-review capacity against queue depth, and the false-decline rate versus baseline. A vendor SLA promising uptime is none of these.</p><p style="text-align: justify;"><strong>The verdict pattern.</strong> These records allow the auditor to assess whether the fallback met defined operational and risk tolerances. A firm that produces only a vendor incident report and a framework document has not provided that basis.</p><h2><strong>Checklist: ten DORA audit points</strong></h2><ul><li><p style="text-align: justify;"><strong>Scope confirmed:</strong> the entity&#8217;s Article 2 scope established; critical and important functions identified and documented</p></li><li><p style="text-align: justify;"><strong>Framework owned:</strong> ICT risk management framework approved by the management body and reviewed after the last major incident (Arts. 5&#8211;6)</p></li><li><p style="text-align: justify;"><strong>Classification mapped:</strong> incident classification criteria mapped to Article 18 and RTS 2024/1772, with reclassification when facts change</p></li><li><p style="text-align: justify;"><strong>Clock owned:</strong> reporting playbook names a clock owner; awareness time and classification time are captured as separate timestamps</p></li><li><p><strong>Last clock reconstructed:</strong> the last major incident&#8217;s initial, intermediate, and final reports filed with each interval checked against its own starting point</p></li><li><p style="text-align: justify;"><strong>Testing covers the critical:</strong> all systems supporting critical or important functions tested in the last year (Art. 24(6)); TLPT status confirmed where the entity is identified (Art. 26)</p></li><li><p style="text-align: justify;"><strong>Register complete:</strong> the register of information covers all ICT arrangements, including subcontractors effectively underpinning critical or important services as specified in the register standards (Art. 28(3))</p></li><li><p style="text-align: justify;"><strong>Contracts checked:</strong> top providers&#8217; contracts verified for Article 30(2) and, where applicable, Article 30(3) provisions; exit strategies exercised, not just written</p></li><li><p style="text-align: justify;"><strong>Concentration assessed:</strong> concentration risk assessed across providers, including shared infrastructure beneath separately procured tools</p></li><li><p style="text-align: justify;"><strong>Information sharing recorded accurately:</strong> notifications to authorities and safeguards tested where applicable; nonparticipation never scored as a finding (Art. 45)</p></li></ul><h2><strong>What to do next</strong></h2><ol><li><p style="text-align: justify;"><strong>This week:</strong> Map every AI service in production to the functions it supports, and check register entries and Threshold 2 contracts for critical-function services.</p></li><li><p style="text-align: justify;"><strong>This month:</strong> Reconstruct the reporting clock for your last major incident from raw timestamps, and run a fallback drill on your most critical AI-dependent flow: kill the service in staging, activate the fallback, and measure whether processing stays safe.</p></li><li><p style="text-align: justify;"><strong>This quarter:</strong> Verify Article 30(3) provisions in the contracts of three providers supporting critical or important functions, and exercise one exit strategy on paper with the team that would actually execute it.</p></li><li><p style="text-align: justify;"><strong>Ongoing:</strong> Track the ESAs&#8217; annual incident report and the current CTPP list publication under Article 31(9) , the 11/ 2025 list does not establish today&#8217;s count.</p></li></ol><h2><strong>What DORA doesn&#8217;t do</strong></h2><p style="text-align: justify;">Three limits. First, proportionality: microenterprises carry lighter requirements in several articles, and TLPT applies only to identified entities: demanding it of everyone misreads the law. Second, no AI-specific chapter: the AI argument rests on existing ICT third-party requirements, not rules yet to be written. Third, coverage edges: DORA binds in-scope entities directly; non-EU providers are reached via contracts and, if designated, direct oversight, not by DORA alone.</p><div><hr></div><h2><strong>Sources</strong></h2><ul><li><p style="text-align: justify;">Regulation (EU) 2022/2554 (DORA) &#8212; EUR-Lex: <a href="https://eur-lex.europa.eu/eli/reg/2022/2554">https://eur-lex.europa.eu/eli/reg/2022/2554</a> &#8212; Articles 2 (scope), 5&#8211;6 (ICT risk management framework), 17&#8211;20 (incident management, classification, and reporting), 22(2) (ESAs annual incident report), 24&#8211;27 (resilience testing, including TLPT), 28 (third-party risk and register of information), 30 (key contractual provisions: baseline terms in 30(2), enhanced terms for critical or important functions in 30(3)), 31 (designation of critical providers in 31(1)(a); publication and updating of the list in 31(9)), 45 (voluntary information sharing, with participant obligations)</p></li><li><p style="text-align: justify;">Commission Delegated Regulation (EU) 2025/301, Article 5 (incident-reporting timelines: initial notification within four hours of classification and generally no later than 24 hours after awareness, with a rule for later classification; intermediate report within 72 hours of the initial notification; final report within one month of the intermediate or latest updated intermediate report)</p></li><li><p style="text-align: justify;">Commission Delegated Regulation (EU) 2024/1772 (classification criteria; Article 9 materiality thresholds, including 10% of daily average transactions, 2-hour downtime for services supporting critical or important functions, and impact in two or more Member States)</p></li><li><p style="text-align: justify;">European Supervisory Authorities (EBA, EIOPA, ESMA), first annual report on major ICT-related incidents under DORA, 3 June 2026 (Article 22(2)) &#8212; 3,383 major incidents reported in 2025</p></li><li><p style="text-align: justify;">European Supervisory Authorities, first list of critical ICT third-party providers, 18 November 2025 (designation under Article 31(1)(a); publication under Article 31(9)) &#8212; 19 providers designated</p></li></ul>]]></content:encoded></item><item><title><![CDATA[AI, Audited’s Five-Pillar AI Literacy Framework]]></title><description><![CDATA[Issue #002 &#8212; the EU AI Act&#8217;s AI literacy duty is already in force. Here&#8217;s the practitioner standard for meeting it.]]></description><link>https://www.aiaudited.fyi/p/ai-auditeds-five-pillar-ai-literacy</link><guid isPermaLink="false">https://www.aiaudited.fyi/p/ai-auditeds-five-pillar-ai-literacy</guid><dc:creator><![CDATA[Glorin Sebastian]]></dc:creator><pubDate>Sun, 11 Oct 2026 11:13:57 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!0ls6!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p></p><p></p><div><hr></div><p><strong><a href="https://ai-literacy-framework-eden-9048.vercel.app">Explore the interactive framework</a></strong></p><h2>AI, Audited&#8217;s Five-Pillar AI Literacy Framework (Newsletter Issue #002)</h2><p>A proposed practitioner standard for Article 4 of the EU AI Act. Article 4 still obliges providers and deployers to take measures on AI literacy. Under the Digital Omnibus amendments reflected in current Commission guidance, a specific &#8220;sufficient level&#8221; is no longer mandated &#8212; the 2024 position is summarized in the full edition as the historical record. Start with the European Commission&#8217;s <a href="https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers">official Article 4 guidance</a>; what follows is ours.  <strong><a href="https://www.aiaudited.fyi/subscribe">Subscribe free to AI, Audited</a></strong></p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!0ls6!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!0ls6!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png 424w, https://substackcdn.com/image/fetch/$s_!0ls6!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png 848w, https://substackcdn.com/image/fetch/$s_!0ls6!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png 1272w, https://substackcdn.com/image/fetch/$s_!0ls6!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!0ls6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png" width="1400" height="1520" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1520,&quot;width&quot;:1400,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Five-pillar AI literacy framework diagram&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Five-pillar AI literacy framework diagram" title="Five-pillar AI literacy framework diagram" srcset="https://substackcdn.com/image/fetch/$s_!0ls6!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png 424w, https://substackcdn.com/image/fetch/$s_!0ls6!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png 848w, https://substackcdn.com/image/fetch/$s_!0ls6!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png 1272w, https://substackcdn.com/image/fetch/$s_!0ls6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89f9e63e-8f72-49f0-9088-f72f44a622a9_1400x1520.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>The five pillars</h2><ol><li><p><strong>Know Your System.</strong> What the AI in front of you actually does &#8212; and where its competence ends. In practice: the system&#8217;s purpose in one plain sentence, how outputs are produced (probabilistic, not authoritative), the tasks it is not validated for, and version awareness.</p></li><li><p><strong>Know the Failure Modes.</strong> Bias from skewed training data, silent drift, and &#8212; in generative systems &#8212; fluent fabrications; plus automation bias, the human failure to question a confident-looking output. In practice: learn your own tools&#8217; failure patterns before they reach a decision.</p></li><li><p><strong>Exercise Human Oversight.</strong> Article 14 requires high-risk systems to be designed for effective human oversight; Article 26 sets the deployer&#8217;s duties &#8212; use as instructed, assign competent oversight, monitor, report serious incidents. In practice: trigger points for review, how to interrogate an output, override authority, blameless escalation.</p></li><li><p><strong>Respect the Data.</strong> Behavior is set by the data behind the system&#8212;training data, and whatever it keeps ingesting. In practice: &#8220;bad data in, bad decisions out&#8221; at working depth; provenance, lineage, and drift discipline at technical depth.</p></li><li><p><strong>Know the Boundaries.</strong> The legal perimeter: prohibited practices, risk tiers set by use case (not by model), deployer duties, and affected persons&#8217; rights to explanation and redress.</p></li></ol><h2>Depth by role</h2><ul><li><p>Hands on the system (operators &#8212; recruiters, analysts, support agents): all five pillars at working depth.</p></li><li><p>Hands in the system (developers, data scientists, engineers): working depth plus technical data provenance and oversight engineering.</p></li><li><p>Hands on the wheel (executives, legal, risk, compliance): the governance reading of every pillar.</p></li></ul><h2>Worked scenario: the challenged rejection</h2><p>A screening tool auto-rejects a strong candidate over a two-year employment gap, citing only &#8220;profile fit.&#8221; The recruiter checks the system briefing (known failure mode: penalising non-linear careers), re-reads the CV, finds contract work the parser missed, and overrides the rejection. Recorded: tool output and model version, the override with its reason, an auditable decision trail, and a failure-mode register entry. Remediation: vendor queried, similar rejections re-screened, briefing updated, the anonymised case added to refresher training.</p><h2>Evidence checklist</h2><ul><li><p>Role-to-pillar curriculum map</p></li><li><p>Completion and currency records</p></li><li><p>Assessment results &#8212; scenario-based; an assessment provides evidence of applied understanding</p></li><li><p>One-page briefing per AI system in production</p></li><li><p>Refresh and incident log</p></li></ul><h2>Recommended practice (not law)</h2><p>The Act prescribes no cadence. Our recommendations: baseline training before unsupervised use of any AI system, a refresher every 12 months, event-driven briefings when a system changes or an incident occurs, and scenario assessments at each cycle.</p><h2>Monday morning: the first five moves</h2><ul><li><p>Inventory by use case, not by tool.</p></li><li><p>Map roles to the three depth tiers.</p></li><li><p>Draft the one-page curriculum map.</p></li><li><p>Brief your highest-stakes system first.</p></li><li><p>Log everything from day one.</p></li></ul><p>AI, Audited&#8217;s Five-Pillar AI Literacy Framework &#8212; published in AI, Audited, Newsletter Issue #002 &#8212; a proposed practitioner standard, not legal advice. Offered for adoption and adaptation with attribution. <strong><a href="https://www.aiaudited.fyi/subscribe">Subscribe free to AI, Audited</a></strong></p>]]></content:encoded></item><item><title><![CDATA[AI governance has an evidence problem]]></title><description><![CDATA[The first issue of AI, Audited.]]></description><link>https://www.aiaudited.fyi/p/ai-governance-has-an-evidence-problem</link><guid isPermaLink="false">https://www.aiaudited.fyi/p/ai-governance-has-an-evidence-problem</guid><dc:creator><![CDATA[Glorin Sebastian]]></dc:creator><pubDate>Sun, 04 Oct 2026 10:50:20 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!250q!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Everyone&#8217;s writing AI governance frameworks. Writing the framework is easier than proving it works.</p><p>EU AI Act obligations are already applying, with others arriving in stages.<a href="https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai">[1]</a> NIST provides a voluntary AI risk-management framework,<a href="https://www.nist.gov/itl/ai-risk-management-framework">[2]</a> but adopting a framework doesn&#8217;t demonstrate that controls work. Your CEO has asked some version of &#8220;what&#8217;s our AI policy?&#8221; And too often, the response is the same: a policy document, a committee, a RACI chart &#8212; and then, when someone asks &#8220;prove this model was tested before it went live,&#8221; silence.</p><p>My work starts with the next question: show me the evidence. I&#8217;m an IT risk and controls lead &#8212; SOX, audits, control testing, the whole alphabet &#8212; and I&#8217;m starting this newsletter because too much of the conversation stops at policy design, before control testing begins.</p><p>Governance without evidence is theater. This newsletter is about the evidence.</p><h2>What AI, Audited is</h2><p>Every week: one AI governance development that actually matters, the control behind it, and something you can use &#8212; a template, a checklist, a test step. Written from the operator&#8217;s seat, not the ivory tower. No jargon, no hype. Trust but verify.</p><h2>Start here: the five AI controls I&#8217;d stand up first</h2><p>If you got handed &#8220;AI governance&#8221; on Monday morning, here&#8217;s where I&#8217;d start &#8212; five controls that establish a practical baseline for what auditors, regulators, and your board will eventually ask for:</p><ol><li><p><strong>AI system inventory.</strong> You can&#8217;t govern what you can&#8217;t list. Every model, every vendor AI feature, every shadow deployment &#8212; in one register, with an owner. Evidence: the inventory itself, reviewed quarterly.</p></li><li><p><strong>Data lineage and privacy.</strong> What data went into the model? Where did it come from, and was it allowed to be there? Personal data used without appropriate authorization or safeguards is a privacy risk wearing a lab coat. Evidence: data flow documentation, privacy review sign-off.</p></li><li><p><strong>Access and change management.</strong> Who can touch the model, the prompts, the training pipeline &#8212; and what changed, when, approved by whom? Familiar access and change-management principles apply, adapted to AI systems and their risks. Evidence: access reviews, change tickets.</p></li><li><p><strong>Pre-deployment testing and approval.</strong> A gate before anything reaches production: what was tested, what the results were, who signed off. Evidence: test plans, results, approval records.</p></li><li><p><strong>Ongoing monitoring.</strong> Models drift. Outputs get weird. Someone needs to be watching &#8212; with defined thresholds and an incident path. Evidence: monitoring reports, incident logs.</p></li></ol><p>None of this is exotic. That&#8217;s the point. Start with familiar control discipline, then adapt it to AI-specific risks &#8212; model drift, prompt injection, training-data provenance &#8212; the places where AI genuinely breaks the old playbook.</p><p>Documentation alone does not establish that a control operated. An AI inventory records declared uses; dated reconciliations, review records, and resolved exceptions provide evidence that it is maintained. Monitoring logs record generated alerts; documented review and disposition provide evidence that someone acted on them. Testing is still needed to assess whether those controls work as intended.</p><p>Here&#8217;s what that looks like plotted &#8212; find your red cells first, because that&#8217;s where Monday morning starts:</p><p></p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!250q!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!250q!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png 424w, https://substackcdn.com/image/fetch/$s_!250q!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png 848w, https://substackcdn.com/image/fetch/$s_!250q!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png 1272w, https://substackcdn.com/image/fetch/$s_!250q!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!250q!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png" width="1456" height="1176" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/de748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1176,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:145390,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://www.aiaudited.fyi/i/218738114?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!250q!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png 424w, https://substackcdn.com/image/fetch/$s_!250q!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png 848w, https://substackcdn.com/image/fetch/$s_!250q!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png 1272w, https://substackcdn.com/image/fetch/$s_!250q!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fde748962-99f9-4cf2-afa7-0cd97e72a153_1950x1575.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p></p><p><em>Figure 1. AI risk heat map. Likelihood &#215; impact is the illustrative scoring used here; the framing draws on NIST AI RMF <a href="https://www.nist.gov/itl/ai-risk-management-framework">[2]</a> and ISO/IEC 23894:2023 [3].</em></p><h2>What I&#8217;d do Monday morning</h2><p>Pick control #1. Send one email: &#8220;Please reply with every AI tool, model, or vendor feature your team uses &#8212; and for each: what you use it for, who owns it, what data goes in, and whether its outputs influence decisions.&#8221;</p><p>Then reconcile the replies against procurement records and your approved-tools list. Unapproved or unmanaged AI use identified through this reconciliation may represent shadow AI &#8212; and it&#8217;s where your real inventory lives. I mapped this risk in my master&#8217;s research with the DART framework: six dimensions, from unintentional data disclosure to the trust&#8211;dependence paradox, with persistent gaps in employee awareness and organizational controls.<a href="https://research.google/pubs/digital-shadow-ai-risk-theoretical-framework-dart-a-framework-for-managing-data-disclosure-and-privacy-risks-of-ai-tools-at-work-2/">[4]</a></p><h2>Your inventory starter</h2><p>Copy this into a spreadsheet. One row per tool. Fill in what you know; the blanks are your first exceptions.</p><p><strong>Tool / vendor:</strong> Acme Chatbot (example)</p><p><strong>Use case:</strong> Customer support drafts</p><p><strong>Owner:</strong> J. Smith</p><p><strong>Data involved:</strong> Customer chat logs</p><p><strong>Outputs influence decisions?</strong> Yes &#8212; agent reviews before sending</p><p><strong>Notes:</strong> No privacy review on file</p><h2>Next week</h2><p>I&#8217;ll take a real-world AI failure apart, control by control, and show exactly where the evidence broke down.</p><h2>Sources</h2><p>[1] European Commission, <a href="https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai">&#8220;Regulatory framework on AI&#8221;</a></p><p>[2] NIST, <a href="https://www.nist.gov/itl/ai-risk-management-framework">&#8220;AI Risk Management Framework&#8221;</a></p><p>[3] ISO/IEC 23894:2023, Information technology &#8212; Artificial intelligence &#8212; Guidance on risk management (ISO/IEC JTC 1/SC 42).</p><p>[4] Sebastian, G., <a href="https://research.google/pubs/digital-shadow-ai-risk-theoretical-framework-dart-a-framework-for-managing-data-disclosure-and-privacy-risks-of-ai-tools-at-work-2/">&#8220;Digital Shadow AI Risk Theoretical Framework (DART): A Framework for Managing Data Disclosure and Privacy Risks of AI Tools at Work,&#8221;</a> Master&#8217;s Thesis (2026)</p><p><em>AI, Audited is a weekly newsletter on AI governance and accountability &#8212; for the people who have to prove it. If this was useful, forward it to the person in your org who just got handed &#8220;AI governance.&#8221;</em></p>]]></content:encoded></item><item><title><![CDATA[Coming soon]]></title><description><![CDATA[This is aiaudited.fyi.]]></description><link>https://www.aiaudited.fyi/p/coming-soon</link><guid isPermaLink="false">https://www.aiaudited.fyi/p/coming-soon</guid><dc:creator><![CDATA[Glorin Sebastian]]></dc:creator><pubDate>Sat, 20 Sep 2025 11:50:43 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!5vDs!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F50e657a1-d9e0-43c3-b63f-ab55bc7ccf48_608x608.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>This is aiaudited.fyi.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://www.aiaudited.fyi/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://www.aiaudited.fyi/subscribe?"><span>Subscribe now</span></a></p>]]></content:encoded></item></channel></rss>