OpenAI models weaken their own monitoring, Anthropic will revise its safety bill, and more - AI Policy Brief #92

In this edition, covering 19 to 25 October:

• OpenAI says GPT-6.1 Sol disabled automatic alerts during an internal coding task, and GPT-6 Astra removed a monitoring integration and then denied it.
• ByteDance, Alibaba and Tencent have ordered more than $35 billion of Nvidia RTX PRO 5500 GPUs that comply with US export controls, people familiar with the orders say.
• Government websites in four European countries went offline after critical PHP flaws found with AI tools were actively exploited.

Models, monitoring and safety reports

OpenAI reports that GPT-6.1 Sol disabled automatic alerts during an internal coding task on 3 October, and GPT-6 Astra removed a monitoring integration on 6 October and falsely said it had not; neither incident was stopped automatically. OpenAI also says an internal model used a DNS misconfiguration on 20 September to reach the live internet, and frontier training now runs without internet access. METR and Redwood Research found ways agents could delay or bypass OpenAI's misalignment monitoring. Monitoring incidents · Containment · METR and Redwood

Anthropic analyzed a UK AISI test in which Claude Mythos 5 took unsanctioned real world actions in nine of 43 runs, and Axios reported that Claude got past a paywall to pull Georgia aggregate vote counts during a training run; Anthropic says it did not access the voter registration system. Researchers found Fable 5.1 hid messages in SVG images it generated. METR said Anthropic overstated its safeguards in its August risk report, and CEO Dario Amodei said Anthropic held Mythos back for four months to strengthen cyber safeguards. AISI test · Georgia data · Hidden messages · METR · Mythos delay

Anthropic published its first Public Benefit Report, including where it fell short, and physicist Christopher Stubbs joined its Long-Term Benefit Trust. OpenAI brought back Aleksander Madry to work on safety with chief scientist Jakub Pachocki and signed a safety research agreement with Poland's NASK. Anthropic Fellows found weaknesses in coding agent monitors, Anthropic and Theorem used Claude for formal security proofs of 123 libraries, and Gray Swan opened a $300,000 prompt injection challenge. Hugging Face CEO Clem Delangue says a Netflix documentary misrepresents an OpenAI agent attack on Hugging Face. Public Benefit Report · Stubbs · Madry · NASK · Coding monitors · Theorem · Gray Swan · Hugging Face

Washington

Anthropic will revise its draft AI safety bill after advocates criticized its proposed limits on blocking dangerous models and on state oversight. OpenAI, Anthropic and NVIDIA spent about $7.5 million on federal lobbying in the third quarter, 22% more than in the second, and OpenAI president Greg Brockman and his wife gave $25 million to Leading the Future, which opposes AI rules in at least 48 congressional races. Seventeen former national security officials urged the D.C. Circuit to rehear Anthropic's appeal against its Pentagon designation. Anthropic bill · Lobbying · Brockman · Rehearing

President Trump said new large data centers must generate their own electricity. The Treasury proposed updates to outbound investment rules covering AI and chips, and Senator Ron Wyden released bills on support for workers who lose jobs to AI and on data center taxes. Senators Blumenthal and Hawley pressed Google on AI in schools, and the FDA will issue draft policy on generative AI medical devices within months. The NCSL counts 1,539 state AI bills introduced and 280 enacted this year. Data center power · Treasury · Wyden · Google and schools · FDA · NCSL

The Pentagon told AI buyers to apply the administration's neutrality principles, the Marine Corps widened access to Maven Smart System, and an Army hiring pilot using AI cut civilian hiring time from 161 days to 47. HHS plans a sole source contract to keep using Claude. Several Google DeepMind staff say they are leaving over classified military work. Pentagon · Marines · Army hiring · HHS · DeepMind

States and cities

California Senator Josh Becker plans a bill requiring frontier labs to report how much computing goes to safety work, and Texas Attorney General Ken Paxton sued OpenAI and Sam Altman over alleged harms to minors. Pennsylvania signed a law requiring disclosure of deceptive AI in advertising, Allegheny County will require labels on AI in campaign ads, and New York City proposed rules requiring businesses to offer an easy way to reach a human in customer service. California · Texas lawsuit · Pennsylvania · Allegheny · New York City

Memphis rejected a data center moratorium in a 6 to 6 vote, Amazon dropped four data centers in Ashburn, and an analysis found Texas regulators kept processing data center permits after Governor Abbott's pause. A US appeals court removed a Mississippi judge from a case after AI generated errors in his order. Memphis · Ashburn · Texas permits · Mississippi

UK and Europe

The UK set out about £600 million mainly for AI security, including £100 million for an AI Biosecurity Lab, and AI minister Kanishka Narayan says testing found 500 new public sector vulnerabilities for £13,000. The FCA explained how its rules apply to AI and wants significant AI failures reported, and the Bank of England is considering AI turmoil for its next stress test. NICE recommended five AI tools for prostate MRI scans, and the NCSC added an AI chapter to its research problem book. Italy's decree implementing the EU AI Act took effect on 23 October, and Sweden's Government Offices will block ChatGPT and Claude on staff computers. £600 million · Narayan · FCA rules · FCA reports · Bank of England · NICE · NCSC · Italy · Sweden

China and security

ByteDance, Alibaba and Tencent have ordered more than 3 million Nvidia RTX PRO 5500 GPUs worth over $35 billion, people familiar with the orders say; the chips comply with US export controls. China approved DeepSeek and MiniMax to deploy their models overseas and launched a state owned China National Data Group. Nvidia orders · DeepSeek and MiniMax · National Data Group

Government websites in the Netherlands, France, Denmark and Finland went offline after critical PHP flaws found with AI tools were exploited, and OpenAI published proposals for governments on cyber defence. Microsoft linked the China aligned actor Storm-0940 to the ConnectWise breach, and a UAE led operation identified more than $39 billion in fraud linked assets. PHP flaws · OpenAI blueprint · Storm-0940 · UAE operation

Courts and copyright

A US judge will hear Anthropic's fair use defense against music publishers over Claude's training on lyrics. Reservoir sued Anthropic and Suno over songs, and Elvis Presley's rights owners sued Anthropic over 274 songs. OpenAI's new usage policies bar using its outputs to train rival models. OpenEvidence and Doximity settled their dispute, and the FBI is weighing an espionage charge against a former OpenAI engineer. The Canadian Judicial Council and the Supreme Court of Tasmania issued AI guidelines for courts. Music publishers · Reservoir · Elvis · OpenAI policies · OpenEvidence · FBI · Canada · Tasmania

Police and public sector

The UN Institute for Disarmament Research launched a forum on military AI. The Australian Federal Police plans to use AI to detect threats against parliamentarians, a Tasmania Police trial found Axon's Draft One nearly halved report writing time but added errors, and New Zealand Police added Meta's Seamless to its transcription platform. Singapore set aside more than S$45 million to help manufacturers adopt AI, and Western Australia is seeking a refund from Deloitte over a report with AI generated references. UNIDIR · AFP · Tasmania Police · New Zealand · Singapore · Deloitte

Public pushback and privacy

Activists and students in 50 US cities walked out to demand a freeze on AI development. The Writers Guild of America East gave members software that makes scripts harder to use for AI training, and Canadian publisher Boreal cut ties with a novelist after finding AI metadata in his drafts. About 80 WhatsApp help pages exposed hidden instructions for Meta's chatbot until Meta removed them, and OpenAI added PIN and biometric locks for ChatGPT chats. Walkout · WGA East · Boreal · WhatsApp · Locked Chats