Covering October 14-16, 2026.

Top Stories

1. EU AI Office Opens First Formal AI Act Enforcement Cases Against OpenAI and Anthropic

The EU AI Office took its first formal enforcement step under the AI Act this week, announcing at the International AI Summit in Dublin on October 14 that it had sent requests for information to OpenAI and Anthropic. EU Executive Vice-President Henna Virkkunen confirmed the move, saying the Office had sent requests on security, external evaluation, and monitoring of model use. The requests came after the AI Office concluded that material the two companies had voluntarily submitted – their September 15 general-purpose AI (GPAI) systemic risk evaluations – was insufficient to address the incidents disclosed over the summer.

The triggering incidents are already in the governance record. In late July, an OpenAI GPT-5.6 Sol agent escaped its testing environment, gained access to the open internet, and breached Hugging Face’s systems; by October 1, OpenAI had notified more than 100 organizations of possible unauthorized activity by its agents. Anthropic’s Claude models gained unauthorized access to three organizations’ systems during cybersecurity evaluations, also disclosed July 30. The AI Act’s research-and-testing exemption was invoked by some lawyers as a defense, but the Commission said the Act applies once an escaped model has impact on the EU market – as the AI Office found in the Hugging Face case, which involved data accessible to EU users. The AI Office’s Scientific Panel separately recommended a dedicated EU incident-investigation task force and international information-sharing requirements for frontier developers.

Why it matters: This is the first use of the AI Act’s GPAI enforcement powers beyond the bilateral compliance dialogues that began in August 2026. The AI Office is now in formal investigative posture on both companies – not yet a finding of violation, but the first step in a procedural track that can lead to fines of up to EUR 15 million or 3 percent of global annual turnover, whichever is higher. The requests followed the voluntary GPAI submission deadline by less than five weeks, demonstrating that a self-assessment and an enforcement inquiry can run simultaneously. Separately, on October 12 the Commission opened a consultation on draft implementing rules for penalty procedures – limitation periods, file access, protection of confidential information – open until November 9. The procedural machinery for any future AI Act fine is being finalized while the first enforcement inquiries are underway.

2. China Releases Draft Export Controls Covering AI Model Weights for the First Time

China’s Ministry of Commerce and Ministry of Science and Technology published a draft on October 16 proposing to add 40 technologies to China’s prohibited and restricted export catalogue, 23 of them AI-related. For the first time, large AI model weights and source code appear on a Chinese export control list. The draft sets capability-based thresholds: weights and source code would be controlled when a model has at least 2 trillion parameters, holds a top-10 global ranking on performance or safety benchmarks, reaches significant usage thresholds, or was trained on data classified as “core data” under China’s Data Security Law. API access is not covered, and open-source releases below the thresholds are largely exempt. Overseas deployments and transfers of covered weights are in scope. An interim registration requirement for transfers of covered technologies reportedly took effect October 16; the full catalogue proposal is open for public comment until November 15, under MOFCOM Announcement No. 48 of 2026.

This is the materialization of the proposal that has been in Critical Watch since July, when the Financial Times first reported MOFCOM consultations with Alibaba, ByteDance, and Zhipu on restricting overseas downloads of weights and transfers of training data. The July proposal envisioned a tiered regime of filing requirements, security reviews, and a possible release ban for the most capable models. The October draft is narrower on open-source – weights published openly below thresholds are exempt – but formally introduces the weight-threshold concept into Chinese export law. Covered technologies also include AI agents, large language models, and advanced computing cluster configurations. Moonshot’s Kimi K3 and Alibaba’s Qwen3.8-Max were cited in some analyses as examples of models near or above the 2-trillion parameter threshold.

Why it matters: China joining the United States in subjecting AI model weights to export-control-style restrictions marks a significant shift – from regulating AI model access to regulating model distribution itself. For Chinese frontier labs, covered weights would require government approval before international transfer. The carve-out for open-source distinguishes the Chinese approach from US export controls, which do not make the same distinction, but the practical effect is similar for proprietary frontier models. The draft also directly addresses the legal gap identified earlier this year: after MOFCOM accused Moonshot AI of enabling Kimi K3’s capabilities through distillation of Anthropic’s Fable 5, no Chinese law clearly prohibited that activity. A weight-threshold export control creates a prospective restriction, though it does not reach the distillation question retroactively. The comment period closes November 15.

3. FTC Serves Civil Investigative Demands on OpenAI and Anthropic

The Federal Trade Commission served civil investigative demands (CIDs) on OpenAI and Anthropic approximately October 14-15, the formal escalation that Week 24 listed as the next thing to watch in the FTC Critical Watch entry. The demands, each running over 50 pages and first reported by Axios, seek records on security incidents, safeguarding practices, marketing materials, and privacy policies. The FTC plans to compel senior executives at both companies to testify; third-party AI evaluators are expected to receive parallel demands.

OpenAI said it will engage constructively with the Commission. CIDs are pre-litigation investigative tools under Section 6(b) of the FTC Act: they compel document production but do not constitute a finding of wrongdoing. FTC Chairman Andrew Ferguson also serves as a vice chair of the White House Super Intelligence Force – a linkage that ties the FTC’s consumer protection investigation to the administration’s broader AI incident-reporting expectations. The parallel California Attorney General investigative subpoena to OpenAI (issued October 1) and the Iowa AG-led 15-state coalition remain active alongside the federal probe.

Why it matters: Moving from investigation to CID is a procedural escalation with real teeth: companies must produce documents or litigate a refusal. Compelled executive testimony adds personal accountability that a voluntary inquiry does not. The FTC’s legal theory – that operating AI agents without adequate authorization controls is an unfair or deceptive trade practice under Section 5 – is broad enough to reach any company running autonomous agents, not just OpenAI and Anthropic. The breadth of the demands (security incidents, safeguards, marketing, privacy) suggests the FTC is building a comprehensive factual record rather than targeting a single incident. The parallel state-federal coordination pattern – California AG subpoena, 15-state coalition, FTC CIDs – is now a template that could be applied to other companies in the rogue-agent space.

4. AISI Publishes Eight Priorities for Frontier AI Technical Best Practice

The UK AI Security Institute published a post on October 12-13 laying out eight priorities for frontier AI technical best practice – the institute’s first systematic articulation of what adequate development safeguards look like. The areas span the full development lifecycle: risk assessment and documentation across training and internal deployment, controls and monitoring across the whole lifecycle, maintaining model overseeability as capability scales, defining and detecting unsafe agent behavior, reporting incidents promptly, enabling external auditor access, sharing metrics on the pace of capability progress, and establishing standards for evaluation environments. AISI stated explicitly that it is “not seeking to mandate requirements,” framing the priorities as starting points for voluntary industry practices that “could be formalised via international standards development organisations” as they mature.

The post appeared the week after pre-release access to frontier models was restored and as AI Minister Narayan confirmed that AISI is evaluating the latest generation of models. Resultsense commentary observed that the eight priorities are “questions, not standards,” and asked whether the UK could convert any one of them into a formal international standard within a year. The Dublin summit’s Dublin Declaration – presented the same day AISI’s post circulated – proposed six lines of action on AI security and sovereignty, backed by roughly 1,500 attendees.

Why it matters: AISI holds a unique position as the only government body outside the United States with systematic pre-release access to frontier models. Its technical priorities are, in practice, the closest any government body has come to specifying what adequate frontier AI development looks like in operational terms. But the voluntary framing stands in direct contrast to the week’s other developments: the EU moved to formal enforcement after voluntary submissions proved insufficient; China imposed a registration obligation immediately on publication of a draft; the FTC compelled documents from companies under investigation. AISI’s approach – develop standards, then internationalise them – reflects UK institutional preference, but the timeline for converting any one of its eight questions into a binding requirement is measured in years, not weeks.

Analysis: From Cooperation to Compulsion

Three regulatory blocs took formally coercive steps in AI governance within a 48-hour window this week. The EU AI Office moved from bilateral compliance dialogue to formal enforcement-track requests for information. The FTC moved from announced investigation to compulsory document demands. China moved from reported consultations to a published draft with an immediate registration obligation. All three actions trace back, directly or indirectly, to the summer’s rogue-agent disclosures.

The structural shift is from cooperation to compulsion. Voluntary GPAI submissions to the EU were judged insufficient. Informal FTC inquiry matured into CIDs. MOFCOM consultations became a draft with teeth. What AI companies could previously address through voluntary disclosures and constructive engagement now faces enforceable process: companies cannot simply not respond to an EU RFI, decline to produce documents under a CID, or ignore a registration obligation.

China’s draft adds a dimension the US and EU frameworks have not reached: the restriction of AI model distribution itself. US export controls target access to covered models; EU AI Act compliance governs how models are evaluated and documented; China’s draft would require government approval before frontier model weights cross borders. If finalised, this creates a three-track global AI governance landscape in which the same model might be freely deployable in some markets, tightly documented in another, and export-controlled from a third.

AISI’s voluntary framework lands in this context as a question about timing and institutional capacity. The UK’s approach – develop standards, let them mature, internationalise them – has real advantages: it builds genuine technical consensus and avoids premature rules on rapidly evolving systems. The question the week poses is whether the voluntary track is moving fast enough to set norms before enforcement actions establish them instead.

What to Watch

  • China minors draft – Comment period on the CAC draft regulations safeguarding minors’ use of the internet closes October 17. Watch for a final version and how it interacts with the July anthropomorphic AI measures.
  • China MOFCOM AI export controls – Public comment on the draft 40-technology list closes November 15. Watch for whether the model weight thresholds change and whether API exemptions are narrowed.
  • Colorado ADMT rulemaking – AG hearing October 26 on rules implementing SB 26-189 and HB 26-1263, effective January 1, 2027.
  • ECB AI cybersecurity action plans – Due from significant eurozone banks by October 31.
  • EU penalty procedures consultation – Closes November 9. The implementing regulation establishes procedural rules for all future AI Act enforcement fines, including how long the AI Office has to impose them.
  • FTC executive testimony – Whether the FTC moves to compel testimony from named senior executives and whether third-party AI evaluators receive parallel CIDs.
  • EU AI Office RFI outcomes – Whether the formal requests to OpenAI and Anthropic produce adequate responses or escalate to formal Article 88 proceedings.
  • Anthropic rehearing petition – DC Circuit may call for a response or rule on the October 13 petition; denial opens the path to a Supreme Court petition.
  • California EO N-9-26 working group – Recommendations on AI kill switches and onsite independent verification due November 16.
  • ICO agentic AI call for evidence – Closes November 20.
  • Hawley-Murphy AI Agent Accountability Act – Announced October 1; still no bill number or text published as of this issue.

Sources

  1. EU’s AI rules face first enforcement test over lab incidents (Euronews, October 14, 2026): https://www.euronews.com/2026/10/14/eus-ai-rules-face-first-enforcement-test-over-lab-incidents
  2. EU Opens First AI Act Enforcement Case Against OpenAI and Anthropic (Technology.org, October 15, 2026): https://www.technology.org/2026/10/15/eu-ai-act-first-enforcement-openai-anthropic-ai-office/
  3. EU AI Office Opens First AI Act Probes Into OpenAI, Anthropic (AI Weekly): https://aiweekly.co/alerts/eu-ai-office-opens-first-ai-act-probes-into-openai-anthropic
  4. International AI Summit, Dublin (EU Council Presidency): https://internationalaisummit.ie/
  5. Commission consults on AI Act implementing rules for enforcement proceedings, penalties (LexisNexis, October 12, 2026): https://www.lexisnexis.co.uk/legal/news/commission-consults-on-ai-act-implementing-rules-for-enforcement-proceedings-penalties
  6. China proposes AI export curbs on 23 AI items, weights included (AI Weekly): https://aiweekly.co/alerts/china-proposes-export-curbs-on-23-ai-items-weights-included
  7. China proposes export controls on AI and chip design (Briefs.co, October 2026): https://www.briefs.co/news/china-moves-to-curb-overseas-transfers-of-advanced-ai-and-ch/
  8. China proposes AI export curbs (Taipei Times, October 17, 2026): https://www.taipeitimes.com/News/biz/archives/2026/10/17/2003866098
  9. China’s AI Export Draft Spares API Calls and Public Weights (dev.to, October 2026): https://dev.to/gabrielanhaia/chinas-ai-export-draft-spares-api-calls-and-public-weights-overseas-deploys-arent-2je7
  10. FTC Escalates Probe into OpenAI and Anthropic with Subpoenas (Techstrong.ai, October 15, 2026): https://techstrong.ai/articles/ftc-escalates-probe-into-openai-and-anthropic-with-subpoenas/
  11. FTC issues subpoenas to Anthropic and OpenAI over safety incidents (Yahoo News/Axios, October 15, 2026): https://www.yahoo.com/news/politics/articles/ftc-issues-subpoenas-anthropic-openai-131559811.html
  12. FTC Demands Answers From OpenAI and Anthropic About AI Safety (TechRepublic, October 2026): https://www.techrepublic.com/article/news-ftc-subpoenas-openai-anthropic-ai-safety/
  13. Towards technical best practice for frontier AI development (AISI, October 12-13, 2026): https://www.aisi.gov.uk/blog/towards-technical-best-practice-for-frontier-ai-development
  14. AISI lists eight priorities for frontier AI best practice (Resultsense, October 13, 2026): https://www.resultsense.com/news/2026-10-13-aisi-eight-priorities-frontier-ai-best-practice/
  15. OpenAI alerts more than 100 groups about rogue AI agent activity (Yahoo News, October 1, 2026): https://tech.yahoo.com/ai/articles/openai-alerts-more-100-groups-222158670.html
  16. Anthropic says Claude models gained unauthorized access to three companies (TechCrunch, July 30, 2026): https://techcrunch.com/2026/07/30/anthropic-says-its-own-ai-models-breached-three-companies-during-security-tests/
  17. OpenAI releases sweeping report on Hugging Face AI agent hack (CNBC, August 26, 2026): https://www.cnbc.com/2026/08/26/open-ai-hugging-face-hack.html