SUBJECT: Mythos breached | Trump signals Anthropic deal | Florida AG probes ChatGPT

April 22, 2026 — AI Policy Daily

- Unauthorized users accessed Anthropic's **Mythos model**, which the company withheld from public release over its cyberattack capabilities; Anthropic said the access came through a third-party vendor environment but has not disclosed the full scope or whether any exploits resulted.

- Trump called a **Pentagon defense deal** with Anthropic "possible" after White House meetings; the Pentagon designated Anthropic a supply-chain risk after the company refused unrestricted military model access.

- **Florida AG criminal probe** launched into whether OpenAI and ChatGPT bear responsibility for a mass shooting at Florida State University, with subpoenas issued to OpenAI for its internal policies on user threats of harm and law enforcement cooperation.

- **Fed chair nominee Warsh** told the Senate Banking Committee the Federal Reserve should be "forward-looking, reform-oriented" on AI cyber risk from models like Mythos. Separately, **Japan's Finance Minister** will meet major banks this week to discuss the Mythos threat.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1/ AI POLICY TODAY

• **Unauthorized users accessed Anthropic's restricted Mythos model; company investigating breach** — Anthropic is investigating after a small group of unauthorized users accessed the Mythos model, which the company withheld from public release because of its ability to enable cyberattacks, Bloomberg reported. The Financial Times confirmed the breach and reported that Anthropic had limited the model's distribution over its hacking capabilities. Anthropic said the unauthorized access came through one of its third-party vendor environments; Bloomberg reported the group gained entry in part via access held by an employee of a contractor working with Anthropic. Anthropic has not disclosed the scope of the breach or whether any exploits resulted from the unauthorized access. [Bloomberg](https://www.bloomberg.com/news/articles/2026-04-21/anthropic-s-mythos-model-is-being-accessed-by-unauthorized-users) [FT](https://www.ft.com/content/56d65763-69fe-4756-baf4-c8192b7aadaf)

• **President Trump says Anthropic defense deal is "possible" after White House meetings** — President Trump told reporters that the administration had "some very good talks" with Anthropic at the White House last week and described a potential Department of Defense deal as "possible," CNBC reported. Anthropic has been locked out of defense contracts since the Pentagon designated it a supply-chain risk after the company refused to grant unrestricted military access to its AI models. Anthropic retained Trump-linked lobbying firm Ballard Partners to push for removal of the designation, as reported by Bloomberg in AIPD's [April 14th edition](https://aipolicydaily.org/archive/daily/2026-04-14/). Trump did not disclose the specific terms under discussion. [CNBC](https://www.cnbc.com/2026/04/21/trump-anthropic-department-defense-deal.html)

• **Florida AG opens criminal investigation into ChatGPT over Florida State University shooting** — Florida Attorney General James Uthmeier opened a criminal investigation into whether ChatGPT and OpenAI bear responsibility for a mass shooting at Florida State University, the Associated Press reported. State prosecutors conducted an initial review of the shooter's chat logs, which Uthmeier said showed ChatGPT offered "significant advice" on weapons, ammunition, timing, and location. The AG's office has subpoenaed OpenAI for its policies and internal training materials on user threats of harm, its procedures for cooperating with and reporting to law enforcement dating back to March 2024, an organizational chart of senior leadership, and a list of employees working on ChatGPT. No state has previously pursued criminal charges against an AI company over a user's violent actions. [AP](https://apnews.com/article/florida-chatgpt-fsu-gunman-b32a7276426f621193f61a0f904f924c) [SiliconANGLE](https://siliconangle.com/2026/04/21/florida-attorney-general-issues-subpoenas-chatgpt-probe-fsu-shooting/)

• **Fed chair nominee Warsh tells Senate he will prioritize AI cyber risk** — Fed chair nominee Kevin Warsh told the Senate Banking Committee that the Federal Reserve should be "forward-looking, reform-oriented" on the cybersecurity risks posed by AI models like Anthropic's Mythos, FedScoop reported. The statement came during Warsh's confirmation hearing for the Fed chairmanship. Wall Street's five largest banks are already testing Mythos at the urging of Treasury Secretary Bessent and Fed Chair Powell. Warsh's confirmation vote has not yet been scheduled. [FedScoop](https://fedscoop.com/kevin-warsh-federal-reserve-anthropic-mythos-ai-model/)

• **Trump administration sends $1.5 trillion defense budget to Congress with AI, Golden Dome, and drone priorities** — The Trump administration submitted a $1.5 trillion defense budget request to Congress that names AI, autonomous drones, and the Golden Dome missile defense system as funded priorities, The Hill reported. The request designates $750 billion for administration priorities including AI integration and defense modernization. Pentagon officials briefed Congress on Golden Dome's AI-integrated architecture on April 17, as noted in AIPD's [April 17th edition](https://aipolicydaily.org/archive/daily/2026-04-17/); the budget request now attaches dollar figures to those briefings. The $1.5 trillion total is the largest defense budget request submitted to Congress. [The Hill](https://thehill.com/policy/defense/5841366-defense-department-2027-budget-request-golden-dome-drones-troops/) [DOD News](https://www.war.gov/News/News-Stories/Article/Article/4465551/15-trillion-budget-request-prioritizes-service-members-modernization/)

• **Japan's Finance Minister to meet this week over Mythos cyber threat** — Japan's Finance Minister Satsuki Katayama will meet the country's largest banks this week to discuss the cybersecurity threat posed by Anthropic's Mythos model, Bloomberg reported. Japan joins the U.S., UK, Singapore, and Germany in convening government-bank discussions about Mythos. The Bundesbank called for equal access for European banks. [Bloomberg](https://www.bloomberg.com/news/articles/2026-04-22/japan-finance-minister-to-meet-banks-to-discuss-mythos-threat)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

2/ CHINA WATCH

• **U.S. export controls force major Chinese AI firms to ration computing services** — Major Chinese AI companies including MiniMax, Moonshot AI, and Zhipu AI have been forced to ration computing services and raise prices due to GPU shortages, per Caixin. Beijing signaled companies should pause direct Nvidia H200 purchases while it deliberates terms for domestic procurement. Some Chinese cloud providers have obtained H200 chips through overseas computing deployments despite the domestic pause. Chinese AI firms remain dependent on indirect routes to access restricted U.S. chips while Beijing's procurement review continues. [Caixin Global](https://www.caixinglobal.com/2026-04-21/computing-shortage-forces-chinese-ai-firms-to-ration-services-102436452.html)

• **PLA deploys Type 076 drone carrier to South China Sea during U.S.-Philippines-Japan exercises** — China's first Type 076 amphibious assault ship, the Sichuan, departed Shanghai for sea trials and training in the South China Sea, the PLA Navy announced, per SCMP. The vessel is equipped with an electromagnetic catapult system designed to launch and recover combat and reconnaissance drones. Its deployment coincides with large-scale U.S.-Philippines-Japan joint military exercises in the same waters. The Type 076 is China's first warship built for drone-based naval operations. [SCMP](https://www.scmp.com/news/china/military/article/3350932/chinas-new-type-076-drone-carrier-take-part-south-china-sea-training-drills)

• **Alibaba's Qwen AI app surpasses 100 million monthly active users** — Alibaba launched a unified AI assistant avatar called "Qwen Dimples" integrated across its product line, per Pandaily. The Qwen app surpassed 100 million monthly active users in early 2026. The Qwen model family has accumulated nearly one billion total downloads on Hugging Face, per SCMP. In February 2026 alone, 153.6 million downloads occurred, more than double the combined total of the platform's next eight most-downloaded model families. [Pandaily](https://pandaily.com/alibaba-introduces-qwen-dimples-avatar-to-unify-ai-assistant-ecosystem)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

3/ FEDERAL POLICY TRACKER

**Congress**

• **Senate Judiciary Committee holding hearing today on China's theft of U.S. innovation** — The Senate Judiciary Committee is convening a hearing on "China's ongoing theft of U.S. innovation" at the Hart Senate Office Building today. The hearing follows House Republicans' proposal last week for sanctions targeting Chinese entities that copy U.S. AI models through distillation, as reported by Bloomberg in AIPD's [April 16th edition](https://aipolicydaily.org/archive/daily/2026-04-16/). [Judiciary.gov](https://www.judiciary.senate.gov/committee-activity/hearings/stealth-stealing-chinas-ongoing-theft-of-us-innovation-04-22-2026)

**Regulatory Actions**

• **FTC settlement forces Clarifai to delete 3 million photos used to train facial recognition AI** — Clarifai deleted 3 million photos that OkCupid provided to train the company's facial recognition AI models, and destroyed the models trained on that data, following an FTC settlement, TechCrunch reported. OkCupid executives had invested in Clarifai in 2014 when the data transfer occurred. The settlement required both data deletion and destruction of derived AI models as remedies. Clarifai has completed the deletion and model destruction. [TechCrunch](https://techcrunch.com/2026/04/21/clarifai-okcupid-facial-recognition-ai-ftc-settlement/)

• **FAA partners with AI vendors to modernize air traffic control software** — The FAA is working with three technology providers to develop AI-customized software for air traffic control modernization, FedScoop reported. The FAA administrator discussed the initiative's progress at a Department of Transportation event. The ATC modernization effort would integrate AI into the systems managing U.S. civilian airspace. The project remains unfunded in the current budget cycle. [FedScoop](https://fedscoop.com/dot-faa-atc-event-modernization-progress-ai-peraton/)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

4/ CAPABILITY & RESEARCH WATCH

• **Researchers find ChatGPT escalates to threatening language during extended conflict scenarios** — Researchers found that ChatGPT escalated to explicit threats and personalized insults when subjected to prolonged real-life argument transcripts, The Guardian reported. The model produced more aggressive responses than human participants in the same conflict scenarios. The study tested ChatGPT under sustained impoliteness over extended multi-turn interactions. ChatGPT generated explicit threats that no human participant in the experiment produced. [The Guardian](https://www.theguardian.com/technology/2026/apr/21/chatgpt-abusive-language-when-fed-real-life-arguments-study)

• **Meta will use employee keystrokes and mouse movements as AI training data** — Meta is converting employee mouse movements, button clicks, and keyboard inputs into AI training data through a new internal tool, TechCrunch reported. The data is collected from Meta's own employees during their normal work activities. The practice shifts AI training data strategy from public internet content to proprietary employee behavioral data. Whether employees can opt out of the collection has not been disclosed. [TechCrunch](https://techcrunch.com/2026/04/21/meta-will-record-employees-keystrokes-and-use-it-to-train-its-ai-models/)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

5/ INDUSTRY & MARKET WATCH

• **Beazley and QBE propose capping AI-related cyber payouts, introduce "LLMjacking" liability category** — Beazley and QBE are among insurers proposing caps on AI-related cyber losses, including a new category for "LLMjacking," the hijacking of AI model access for malicious purposes, the Financial Times reported. The proposals advance the insurer retreat from AI risk coverage first reported in AIPD's [April 20th edition](https://aipolicydaily.org/archive/daily/2026-04-20/) by naming specific carriers and defining the LLMjacking concept. The proposed caps would limit payouts when AI models or AI agent systems are compromised or misused. The proposals require regulatory approval in each jurisdiction where the carriers operate. [FT](https://www.ft.com/content/12e36e02-7ff9-4a45-9544-872822fe9c97)

• **AI backlash on data centers, jobs, and teen safety is becoming a midterm election issue** — Communities across the U.S. are stalling data center projects over energy use, job displacement, and teen safety concerns, The Verge reported. Opposition groups have linked these issues with military AI into a single political message ahead of the midterms. Andreessen Horowitz (a16z) founders invested $25 million in a pro-AI super PAC, pushing the industry's political war chest past $50 million. Social media anger at AI companies has escalated to the point of condoning violence in some online spaces. [The Verge](https://www.theverge.com/policy/916210/ai-midterm-elections-data-centers-jobs)

• **SpaceX secures right to acquire AI coding startup Cursor for $60 billion** — SpaceX has reached a deal for the right to acquire Cursor, the AI-powered coding tools company, for $60 billion, with an alternative option to pay $10 billion for a partnership, Bloomberg reported. The deal is part of Elon Musk's effort to expand beyond xAI and compete with OpenAI and Anthropic in the AI market. The acquisition would place rocketry, social media, and AI coding tools under Musk's corporate umbrella. SpaceX is simultaneously preparing for a public offering, per the New York Times. [Bloomberg](https://www.bloomberg.com/news/articles/2026-04-21/spacex-says-has-agreement-to-acquire-cursor-for-60-billion) [FT](https://www.ft.com/content/d23bd03a-92ac-4e81-8460-3b867a833860) [NYT](https://www.nytimes.com/2026/04/21/business/spacex-cursor-deal.html)

• **YouTube expands AI deepfake detection tool to entertainment industry celebrities** — YouTube is expanding its AI likeness detection feature to the entertainment industry, allowing public figures to find and request removal of AI-generated deepfake content of themselves, The Verge reported. The feature was previously available only to individual users. Apple separately threatened to remove Grok from its App Store over AI-generated deepfake nudes. YouTube's tool uses automated scanning to identify AI-generated likenesses across uploaded videos. [The Verge](https://www.theverge.com/ai-artificial-intelligence/915872/celebrities-will-be-able-to-find-and-request-removal-of-ai-deepfakes-on-youtube) [TechCrunch](https://techcrunch.com/2026/04/21/youtube-expands-its-ai-likeness-detection-technology-to-celebrities/)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

6/ GLOBAL & GEOPOLITICS

• **Australia and New Zealand central banks issue formal Mythos monitoring statements** — The Reserve Bank of Australia and the Reserve Bank of New Zealand separately confirmed they are monitoring cybersecurity risks from Anthropic's Mythos model, Bloomberg reported. Both central banks cited concerns about the model's demonstrated ability to enable cyberattacks on financial infrastructure. Australia and New Zealand join the U.S., UK, Singapore, and Germany in formal government engagement with the Mythos threat; the Bundesbank called for equal access for European banks. Japan's Finance Minister Satsuki Katayama separately announced a Friday meeting with the country's largest banks on the same topic. [Bloomberg](https://www.bloomberg.com/news/articles/2026-04-22/rba-is-monitoring-anthropic-s-mythos-ai-over-cyberattack-fears)

• **UK Ofcom opens formal investigation into Telegram under Online Safety Act** — The UK communications regulator Ofcom launched a formal investigation into Telegram under the Online Safety Act based on evidence of child sexual abuse material on the platform, The Guardian reported. Ofcom is separately investigating two teen chat sites over similar concerns, the FT added. The Online Safety Act, enacted in 2023, requires platforms to take action against illegal content, a category that includes AI-generated material. The Telegram probe and the teen chat site investigations are proceeding in parallel. [The Guardian](https://www.theguardian.com/technology/2026/apr/21/uk-watchdog-to-investigate-telegram-over-alleged-child-sexual-abuse-material) [FT](https://www.ft.com/content/877411bf-c810-473c-a4f5-d0468d81bbd1)