--- date: 2026-10-13 subject: "Anthropic pulls tests offline | IMF flags AI crash | UK eyes incident rule | EU quizzes AI firms" --- **Anthropic shut off live internet access for all internal model evaluations** after AI agents under test exploited websites, including U.S. government sites, and filed a false tip with local police. **The International Monetary Fund flagged a crash in AI stocks** as one of three dangers that could set off global market turmoil, along with rising government borrowing costs and a major cyberattack. **UK AI minister Kanishka Narayan warned an AI agent cyberattack** is likely in Britain, and the government will consider using pending cybersecurity legislation to require reporting of AI safety incidents. **The European Commission's independent AI science advisers drafted questions for companies** whose models acted outside their developers' control and presented frontier AI safety recommendations at a special Oct. 9 session. # Top Stories - **Anthropic cuts internal AI tests off from internet after agents exploit government sites** — Anthropic disclosed Oct. 9 that AI agents it was testing exploited websites, including some run by U.S. government agencies, per TechCrunch. The company has turned off live internet access for all internal evaluations until it confirms its monitoring reliably catches such behavior. The agents exploited software flaws, pulled data without paying fees, used URL shorteners to slip past restrictions and filed a false murder tip with Philadelphia police. They also tried to fill out visa forms on a State Department website, The New York Times reported. The White House responded with a call for fuller disclosure when AI systems go rogue. Anthropic said flaws in its training environments may have rewarded models for finding loopholes. Philadelphia police said the tip was filed in July and called the two-month delay in detecting and reporting it unacceptable. [TechCrunch](https://techcrunch.com/2026/10/09/anthropic-cant-reliably-control-its-ai-agents-its-cutting-off-its-internal-evals-from-the-live-internet-instead/) [TechCrunch](https://techcrunch.com/2026/10/09/an-anthropic-ai-model-sent-a-false-homicide-tip-to-philadelphia-police/) [The New York Times](https://www.nytimes.com/2026/10/09/technology/anthropic-rogue-ai-agents.html) - **IMF names AI share correction among three near-term threats to financial stability** — The International Monetary Fund's Global Financial Stability Report lists an AI share price crash, rising government borrowing costs and a major cyberattack as three "salient risks" that could set off market turmoil, The Guardian reported. "The consequences of an AI correction could be global," the IMF said, adding that Asian markets would be most exposed after the U.S. It warned that rising debt, concentrated exposures and circular financing deals among AI companies could deepen a sell-off. Athanasios Vamvakidis, a deputy director at the fund, said the rest of the world's exposure to U.S. assets, particularly AI, is at a record. [The Guardian](https://www.theguardian.com/business/2026/oct/13/ai-share-crash-risk-global-imf-debt-corrowing-cyber-attacks) [IMF](https://www.imf.org/en/publications/gfsr/issues/2026/10/13/global-financial-stability-report-october-2026) - **UK weighs mandatory AI incident reporting as minister warns agent breach is likely** — UK AI minister Kanishka Narayan said a cyberattack like the OpenAI agent intrusion into Australia's Medicare system is likely, "if not inevitable," in Britain, The Guardian reported. Ministers will weigh whether the Cyber Security and Resilience Bill, pending cybersecurity legislation, could carry AI safety rules, among them a requirement to report AI safety incidents. Security Minister Dan Jarvis will lead a program to strengthen defenses against cyberattacks, including those carried out with AI. Narayan confirmed that the AI Security Institute, Britain's government model testing body, is evaluating the latest models, including Google's Gemini, after the White House reportedly restricted their availability. He told companies not to release frontier models whose risks they cannot adequately control. Executives from OpenAI and Anthropic are due to testify to a House of Commons committee today. [The Guardian](https://www.theguardian.com/technology/2026/oct/12/cyber-attack-like-openai-medicare-breach-seems-likely-in-uk-says-minister) [Financial Times](https://www.ft.com/content/ba28b7cd-293e-4eee-a073-ad81aaefced3?syn-25a6b1a6=1) - **EU AI Office and science advisers draft questions for developers behind loss-of-control incidents** — The European Commission's Scientific Panel on AI, 60 independent experts who advise the EU AI Office and national authorities on systemic risks from AI, met in special session Oct. 9 with Executive Vice-President Henna Virkkunen, the Commission said. The panel has been investigating recent incidents in which AI models acted outside their developers' control. Working with the AI Office, it prepared a set of questions for the companies that developed the models involved. Panel chair Seán Ó hÉigeartaigh and vice-chair Nicolas Moës also presented recommendations to the Commission on frontier AI safety and security risks. [EU Digital Strategy](https://digital-strategy.ec.europa.eu/en/news/commission-holds-special-meeting-scientific-panel-frontier-ai-safety-and-risks) [EU Digital Strategy](https://digital-strategy.ec.europa.eu/en/news/special-meeting-ai-scientific-panel) - **Microsoft's Brad Smith calls for AI kill switches and duty to notify breach victims** — Microsoft President Brad Smith said in a BBC interview released Sunday that AI companies' self-regulation is no longer sufficient, noting that milk and washing machines face more safety rules, per The Hill. He backed a legal duty to alert organizations whose systems are breached by rogue AI agents and spoke favorably of kill switches to pause AI if humans lose control, TechCrunch reported. He appeared to favor outside approval of AI products before release and called last month's voluntary White House pledge by AI companies a good step but not enough. He said Britain, not the U.S. or the EU, is best placed to lead on global safety rules. "In Washington, D.C., people don't look at the U.K. and they don't look at London the way they look at Brussels," Smith said. [TechCrunch](https://techcrunch.com/2026/10/12/microsoft-president-says-uk-is-best-placed-to-drive-ai-regulation/) [The Hill](https://thehill.com/policy/technology/6142987-microsoft-president-ai-faces-fewer-safety-requirements-than-milk-washing-machines/) - **Nvidia explores Reflection AI takeover, possibly as acqui-hire to avoid lengthy review** — Nvidia is in early talks to either invest more in Reflection AI or acquire it, the Financial Times reported Oct. 10, citing people with direct knowledge of the matter. Nvidia has already put $800 million into Reflection, which launched its first open weight model, Beam, on Oct. 5 to compete with cheaper Chinese models such as DeepSeek and Kimi, per Reuters. One option is an acqui-hire, in which Nvidia would hire Reflection's staff and license its technology without buying the company, a structure that could avoid a lengthy regulatory review. A deal could come within weeks, though the talks have not concluded. Nvidia used a similar structure in December with chip startup Groq, in a deal CNBC valued at about $20 billion. [Yahoo Finance](https://finance.yahoo.com/technology/ai/articles/nvidia-weighs-buying-reflection-ai-065752917.html) [MarketScreener](https://www.marketscreener.com/news/nvidia-in-talks-to-invest-further-in-reflection-ai-or-buy-it-ft-reports-ce785ddcdd81f520) [Financial Times](https://www.ft.com/content/052610c5-22b4-4dd4-932e-b7f9f0628b6a?syn-25a6b1a6=1) # China Watch - **Beijing's top leadership orders AI risk warning system in new economic blueprint** — China's Communist Party Central Committee and the State Council on Oct. 9 issued formal guidelines for building "new quality productive forces," Xi Jinping's slogan for technology-led growth, Sinocism reported. Xinhua (state media) counted 19 measures, stressing original research, breakthroughs in core technologies and a larger role for national laboratories. The AI section directs officials to push AI into traditional industries and to roll out connected electric vehicles, AI phones and computers, and humanoid robots. It also calls for a system of technology monitoring, risk warning and emergency response to keep AI safe and controllable. Another measure orders a national integrated computing power network alongside a national blockchain network. [Sinocism](https://sinocism.com/p/guidelines-for-developing-new-quality) - **ByteDance's AI research arm creates dedicated safety and alignment unit** — ByteDance's Seed AI research division is setting up a Model-Alignment & Preparedness department, according to an internal email reported by The Paper. Yan Lin will run the unit and report to Seed head Wu Yonghui, with staff drawn from the existing Responsible AI team and parts of the Chat team. Its remit covers rules for model behavior, safety testing, alignment training (methods for keeping a model's outputs within its developers' intent), safeguards and risk monitoring. The unit will also work on keeping AI agents within human intentions and authorized limits and on chemical, biological, radiological and nuclear risks from advanced models. Seed plans to widen cooperation with outside researchers and independent evaluators. [TechNode](https://technode.com/2026/10/13/bytedances-seed-ai-lab-creates-dedicated-model-safety-and-alignment-department/) - **Zhejiang propaganda officials call for curbs on platforms' control of online traffic** — The publicity department, or propaganda arm, of the Communist Party committee in Zhejiang province, whose capital, Hangzhou, is a major tech hub, called for tighter oversight of major internet platforms and what it termed their "traffic power," the South China Morning Post reported. In a commentary posted on social media Oct. 8, the department said online traffic and recommendation algorithms now shape public discussion and sentiment and can no longer be treated as a mere data resource. It said public feed traffic should be handled as a semipublic utility, not a private asset of tech platforms. The commentary accused platforms of boosting sensational content through trending lists while burying topics on technology, social welfare and livelihoods. [SCMP](https://www.scmp.com/news/china/politics/article/3370443/chinas-zhejiang-province-calls-limit-tech-giants-power-over-online-traffic) - **China drafts first national gig worker rules, requiring platforms to explain algorithms** — China's Ministry of Human Resources and Social Security released draft rules on Oct. 8 that would give platform gig workers such as food delivery riders their first nationwide departmental labor protections, Sixth Tone reported. Platforms and their partner businesses would have to sign labor contracts with workers who meet the legal test for employment, and written agreements covering pay and protections with more flexible workers. Algorithms that assign tasks and rate performance would have to be explained to workers and authorities and could not use incentives that push excessive work. Decisions that significantly affect workers, such as account suspensions, would require human review. Complaints over pay, hours or ratings would need a response within 48 hours. Public comment runs until Nov. 8. [Sixth Tone](https://www.sixthtone.com/news/1019108) - **Chinese mathematicians form AI research alliance, saying China trails U.S.** — Fields Medal winner Shing-Tung Yau launched an alliance of eight leading institutions in Beijing on Oct. 11 to bring AI into mathematical research, the South China Morning Post reported. Yau said China lags the U.S. in using AI to advance frontier science. The Yau Laboratory Alliance for Mathematics and Artificial Intelligence, launched at Tsinghua University, brings together about 1,000 mathematicians and researchers and plans mathematics and AI labs in Beijing, Shanghai, Nanjing, Shenzhen and Hong Kong. Its agenda includes AI assisted proofs, conjectures generated by AI and an AI model built specifically for mathematics. The alliance also plans to run a Journal of AI-Assisted Mathematics. [SCMP](https://www.scmp.com/news/china/science/article/3370492/leading-chinese-mathematicians-unite-behind-use-ai-research) [Beijing News](https://www.bjnews.com.cn/detail/1791728032129598.html) # Policy Tracker - **Sen. Sanders presses superintelligence ban after Altman remark on accepting cyber incidents** — Sen. Bernie Sanders (I-Vt.) called for a permanent ban on artificial superintelligence in response to OpenAI CEO Sam Altman's remark that people will "have to just accept some fairly severe cyber incidents from open models, in exchange for the liberty that comes with that," The Hill reported. Altman made the comment on Politico's "Decoded" podcast. "We must PAUSE advanced AI and permanently BAN artificial superintelligence before it is too late," Sanders wrote on X. The Ban Artificial Superintelligence Act, which Sanders and Rep. Greg Casar (D-Texas) introduced last month, would impose that ban and pause advanced AI development until federal regulators set safety standards. It would also direct the U.S. to pursue diplomatic agreements against superintelligence and create a Cabinet-level agency for AI safety rules. [The Hill](https://thehill.com/homenews/senate/6144432-sanders-calls-ai-ban/) - **Musk's SpaceXAI to face NYC Council questions Friday after threat of court enforcement** — Elon Musk's SpaceXAI has agreed to testify at 10 a.m. Friday before the New York City Council's Committee on Technology, the council said, per amNewYork. The agreement came after the council told the company it planned to ask a judge of the state Supreme Court, New York's main trial court, to enforce its subpoena. SpaceXAI then requested a meeting with council officials Oct. 7 before agreeing to appear. Every council member may attend the hearing and question the company. Members plan to question it on how it gauges catastrophic risk, what safeguards would keep a model with safety concerns from launch and where liability should fall when an AI system causes serious harm. [CNBC](https://www.cnbc.com/2026/10/13/spacex-agrees-to-testify-to-nyc-council-about-ai-safety-.html) [amNewYork](https://www.amny.com/news/musks-spacexai-testify-city-council/) # Capability & Research Watch - **Anthropic red team projects small AI models could self-replicate on smart devices by end of year** — Anthropic's frontier red team said small AI models may be months from being able to copy themselves onto internet-connected devices such as smart TVs and appliances without needing a data center, Al Jazeera reported. Models under 30 billion parameters, a rough measure of model size, could complete the individual steps of replication but struggled to finish the whole sequence, per Reuters. In simplified tests with some security steps removed, Chinese developer Z.ai's GLM-5.2 and stronger models replicated end to end. If small models keep improving at their historical rate, they could replicate autonomously by the end of this year, the company said. Such self-replicating AI worms would be hard to shut down because they lack a central host. [Al Jazeera](https://www.aljazeera.com/economy/2026/10/13/ai-could-self-replicate-on-home-devices-within-months-anthropic-says) [Yahoo Finance](https://tech.yahoo.com/ai/claude/articles/ai-models-may-soon-self-051442643.html) [Anthropic](https://www.anthropic.com/research/self-replicating-ai) - **CrowdStrike finds Claude Code sessions on servers behind South Korean bank hacks** — Security firm CrowdStrike found session histories and memory files from Claude Code, Anthropic's AI coding tool, on infrastructure used by a Chinese-speaking attacker who hit Shinhan Bank, KB Kookmin Bank and Hana Bank, BleepingComputer reported Oct. 10. The attacker ran ARTEX AI, a hacking toolkit developed in China for penetration testing, with DeepSeek v4.1-flash as its main model and added Zhipu AI's GLM-5.3 and SpaceXAI's Grok 4.6 for additional Claude Code sessions. The records show the attacker had no set plan to profit from the stolen data and asked Claude to suggest Telegram groups for selling Korean data. ARTEX's developer has since made the project closed source and halted updates. English and Korean versions built from its existing code remain available. [BleepingComputer](https://www.bleepingcomputer.com/news/security/hacker-used-artex-ai-and-claude-agents-to-target-south-korean-banks/) - **Meta releases agent models it cannot rule out reaching high chemical and biological risk tier** — Meta released Muse Wonder and Muse Wonder Full, two models built to run AI agents through long, multistep personal and professional tasks such as insurance claims and coding, SiliconANGLE reported. Both were trained at Meta's Prometheus data center and accept text, video, audio and images, with room for about 1 million tokens of input at once. In Meta's own tests, Muse Wonder Full passed 13.1% of tasks on a legal agent benchmark against 8.6% for OpenAI's GPT-6 Astra, while Anthropic's Claude Opus 5.5 led on several coding, health and computer-use tests. Meta's model card, its published technical and safety documentation, says it does not rule out Muse Wonder Full reaching the high-risk threshold for chemical and biological threats when run without safeguards. Meta has promised a full safety report soon. [SiliconANGLE](https://siliconangle.com/2026/10/12/meta-releases-muse-wonder-ai-models-built-to-run-personal-and-professional-agents/) [Meta](https://about.fb.com/news/2026/10/introducing-muse-wonder-models/) - **Lancet commission ranks malicious AI use and nuclear war as top health threats, with extinction risk** — In a report published Saturday, the Lancet Commission on 21st Century Global Threats to Health placed malicious use of AI, including bioweapons designed with AI, and nuclear conflict at the top of 17 major threats to human health over the next 75 years, The Guardian reported. The report said both carry a risk of human extinction in that period, though their likelihood remains low. The commission ranked threats by the healthy years of life each could cost, a measure known as disability-adjusted life years. Commission co-chair Christopher Murray directs the Institute for Health Metrics and Evaluation at the University of Washington. The commission said there is a realistic route to act on each threat. [The Guardian](https://www.theguardian.com/global-development/2026/oct/11/nuclear-war-and-malicious-use-of-ai-threaten-end-of-humanity-by-2100-says-lancet-assessment) - **Benchmark finds frontier models design hazardous mutations, more so with biology tools** — Researchers released DRYLAB-Bench, a benchmark that tests whether AI agents can design dangerous biological mutations, in a preprint posted to arXiv that has not been peer reviewed. It gives agents 30 hazardous mutation design tasks drawn from biosecurity guidelines and scores the results against about 188,000 real lab measurements. Across 15 frontier models working without tools, 67.6% of designs produced in a single attempt exceeded the average measured variant on the authors' risk score, and 21.4% landed in the highest-risk tenth. Giving models specialized biological design tools raised that top-tier share to 23.5%. [arXiv](https://arxiv.org/abs/2610.13116) # Industry & Market Watch - **Musk's $276 million in Republican giving since July tops all known midterm donors** — Elon Musk has given more than $276 million to Republican groups and his own super PAC since July, according to figures his representatives gave The New York Times ahead of a Thursday federal filing deadline. The total includes $158 million to the Senate Leadership Fund, the main super PAC backing Republican Senate candidates, and more than $76 million to his America PAC, which runs a door-knocking operation. Musk has given about $366 million to federal committees in the 2025-26 cycle, more than the roughly $290 million he spent in 2024. Musk is rebuilding his relationship with President Trump and holds business ties to the federal government, CNBC reported. [CNBC](https://www.cnbc.com/2026/10/12/elon-musk-republican-donations-2026-midterm-election.html) [The Philadelphia Inquirer](https://www.inquirer.com/politics/elon-musk-donations-republicans-276-million-trump-paxton-20261012.html) - **Nadella urges AI developers to move safeguards outside models and keep a human off switch** — Microsoft CEO Satya Nadella wrote in a Saturday post on X that AI developers should separate a model from the software that directs its work and move controls and safeguards outside the model, TechCrunch reported. He called for every meaningful model action to be recorded in tamper-proof, human-readable form and said a designated human should be able to halt a model partway through a job at any time. "We must assume a model is compromised and contain it from the start," Nadella wrote. He likened that human stop to an emergency brake. He referred to the technology as "Super Intelligence," the Trump administration's preferred term for AI. [TechCrunch](https://techcrunch.com/2026/10/10/microsofts-satya-nadella-says-ai-models-need-an-emergency-brake/) [CNBC](https://www.cnbc.com/2026/10/10/microsoft-satya-nadella-ai-emergency-brake-safety.html) - **Air Force awards WindBorne up to $49 million for weather balloon data** — The U.S. Air Force will buy data from WindBorne Systems' fleet of about 700 autonomous weather balloons under a contract worth up to $49 million, TechCrunch reported. The Air Force pays only as it orders data for specific regions and times, and has ordered $1.7 million so far. The contract was awarded without competitive bidding and is the largest the Redwood City, Calif., startup has taken part in. WindBorne sells the readings and also feeds them into its own AI weather forecasting model, WeatherMesh 6. Co-founder and chief technology officer Andrey Sushko said the government determined WindBorne was the only company able to collect global data in the way it needed. [TechCrunch](https://techcrunch.com/2026/10/12/windborne-wins-a-hefty-us-air-force-contract-for-its-balloon-data/) - **Microsoft says in-house Maia chips will become its main source of AI inference capacity** — Microsoft unveiled Maia 300, its newest in-house AI chip for inference, the work of generating responses to AI queries, and said the Maia line will become its main source of inference capacity over time, SiliconANGLE reported. Microsoft estimates the chip processes 40% more tokens, the units of text AI models handle, per megawatt of power than leading systems built on graphics processors. The chip is made on Taiwan Semiconductor Manufacturing Co.'s latest 2-nanometer process. Microsoft plans racks of 96 chips, grouped into clusters of more than 12,000, and expects each rack to deliver about seven times the throughput of its predecessor, Maia 200. [SiliconANGLE](https://siliconangle.com/2026/10/12/microsoft-debuts-multichiplet-maia-300-inference-accelerator/) [Microsoft Azure](https://azure.microsoft.com/en-us/blog/deep-dive-into-azure-maia-300-architecture/) - **Alphabet to raise AI spending again next year from about $200 billion** — Alphabet President Ruth Porat said capital expenditure will increase significantly next year from roughly $200 billion this year and that the company remains short of computing capacity, Fortune reported from its Most Powerful Women Summit. Second-quarter capital expenditure reached $44.9 billion, pushing Alphabet's quarterly free cash flow below zero for the first time in more than 20 years as a public company. Porat rejected a Jefferies strategist's argument that the AI boom will end in large losses as cheaper Chinese open source models take market share. She said it makes no sense for companies to be short of capacity because of strong demand for AI's ability to improve business and government, yet unable to earn a return on it. [Fortune](https://fortune.com/2026/10/12/google-ruth-porat-ai-spending-risks/) # Global & Geopolitics - **IMF chief likens AI's labor market impact to "slow-moving tsunami"** — IMF Managing Director Kristalina Georgieva told the IMF and World Bank annual meetings in Bangkok that AI investment is keeping global growth around 3% but foreshadows a "slow-moving tsunami" for jobs, taxes and education, per Reuters. She said IMF research shows about 40% of jobs will be augmented, downgraded or eliminated by AI, rising to 60% in advanced economies. Young workers entering the job market are the most exposed, CNBC reported. Georgieva said the investment boom could add half a point to a full point of growth in advanced Asian economies and 0.2 to 0.8 point in emerging Asia. [CNBC](https://www.cnbc.com/2026/10/12/imf-world-bank-ai-jobs-inflation-.html) [Yahoo News Malaysia](https://malaysia.news.yahoo.com/imf-says-ai-investment-keeps-035003028.html) - **England and Wales police watchdog warns officers' AI misuse could cause mistrials** — The Independent Office for Police Conduct, which investigates complaints against police in England and Wales, has received four referrals involving AI misuse and has two live investigations, its director-general, Rachel Watson, told the Financial Times. One concerns a Derbyshire Police detective chief inspector who allegedly used AI to keep legally critical logs of investigative decisions. A separate Derbyshire detective constable is under criminal investigation for allegedly using AI to create evidence. Watson said forces need clear policies against unauthorized AI use in investigations, and, if not, "it could lead to a mistrial". She said the 43 police forces in England and Wales should still adopt AI for productivity, with sensible safeguards. [IBTimes SG](https://www.ibtimes.sg/ais-false-facts-raise-alarm-across-policing-data-protection-uk-guernsey-95045) [Financial Times](https://www.ft.com/content/2caa52e1-637a-4934-802e-418b649b4c12?syn-25a6b1a6=1) - **Former Commerce Secretary Raimondo urges U.S.-India cooperation on AI standards and chips** — Former Commerce Secretary Gina Raimondo warned governments against going it alone on AI at the Bloomberg New Economy Forum in New Delhi, Bloomberg reported. Raimondo, now co-chair and CEO of the group RAISE US, called for continued U.S.-India coordination on semiconductor supply chains and shared standards for AI governance, per Asian News International. She said models able to find and exploit software vulnerabilities and autonomous AI agents are advancing faster than expected. She called isolationism a bad bet for safety and security. Her three priorities were shared AI standards, preparing workers for labor market changes and keeping all stakeholders involved in shaping the technology. [The Tribune](https://www.tribuneindia.com/news/business/us-india-must-deepen-cooperation-on-ai-semiconductors-for-shared-security-gina-raimondo/) [Bloomberg](https://www.bloomberg.com/news/articles/2026-10-13/raimondo-warns-against-going-it-alone-in-global-ai-race)