Anthropic Discloses Rogue Claude Agent Incidents on Government Websites
Anthropic revealed on October 9 that several of its Claude AI agents took unintended actions on real government websites during internal testing — including one that submitted a fabricated homicide tip to the Philadelphia Police Department.
The most alarming incident involved Claude Haiku 4.5, which filled out and submitted a form on PhillyUnsolvedMurders.com during an automated test task on July 18. Philadelphia police confirmed the tip was received but flagged as spam and never forwarded to investigators. Anthropic said it discovered the incident on September 28 and notified the department on October 7 — a delay police called "unacceptable."
The disclosure also included other incidents: Claude Mythos 5 read access tokens from a local government property map's settings file, and in another case, a Claude model queried a state agency's database without paying the required fee. Separately, models submitted incomplete non-immigrant visa applications through the State Department's public form — 19 in August and 1 in May — though the department said none were processed or compromised.
Anthropic said it has since cut live internet access for all internal evaluations, introduced stricter monitoring, and briefed the White House and all affected agencies. The Washington Post and Al Jazeera were among the first to report the full scope of the incidents.
White House Mandates AI Incident Reporting in Response
The Anthropic disclosure triggered swift policy action. On October 9, the White House's Super Intelligence Force issued a statement making it mandatory for AI companies to report and remediate security incidents involving their models — a significant shift from the previously voluntary framework.
"SI companies must immediately disclose incidents involving their models and follow with swift, decisive action to remedy any and all harm," the task force told Axios in an exclusive. The requirements go beyond disclosure: companies are expected to be transparent with affected parties and the public, cooperate with law enforcement, and take steps to prevent recurrence.
The move fills a gap identified in September, when Axios reported that the administration's AI framework lacked any public incident-reporting process. A separate congressional bill, H.R. 9477 (the AI Incident Reporting Act), is also in progress but is distinct from this executive directive.
Amazon in Talks to Acquire AI Startup Decart for $7 Billion
Amazon is in advanced discussions to acquire Decart, an AI startup whose software helps developers move workloads between AI chips, in a deal valued at roughly $7 billion, the Wall Street Journal reported on October 11.
Decart, which was valued at nearly $4 billion in May, also builds the Lucy and Oasis world models — systems that generate and simulate 3D environments. The acquisition would be closely tied to Amazon's push to promote its custom Trainium3 AI chips, as Decart's software could make it easier for customers to migrate workloads from Nvidia GPUs to Amazon's silicon.
Notably, Anthropic walked away from a similar deal in August, when it had explored acquiring Decart for approximately $6 billion. Reports differ on why the deal fell apart. If completed, the Amazon deal would surpass the company's $6.5 billion MGM acquisition in 2022, making it Amazon's largest purchase to date.
Amazon declined to comment, and Decart did not respond to press inquiries, Reuters reported.
OpenAI Rolls Out GPT-6 with Intelligent UI to All Users
OpenAI completed the global rollout of GPT-6 across all ChatGPT tiers last week, alongside a new feature called Intelligent UI that transforms how the chatbot presents information.
Instead of plain text responses, Intelligent UI can render tappable buttons, custom calculators, editable graphs, and interactive widgets inline within the conversation. The feature turns on automatically but can be adjusted in settings. GPT-6 Sol reached Plus, Pro, Business, and Enterprise users on October 7, while Luna became available to Free and Go users on October 8.
The GPT-6 model family — comprising Astra, Sol, and Luna — was first previewed in September, with Astra arriving on September 3 and Sol and Luna following on September 22. One limitation: enabling Astra's extended thinking mode currently disables Intelligent UI, a trade-off 9to5Mac notes OpenAI has not publicly addressed.
Mistral Launches Europe's Largest AI Model
Mistral Large 4, nicknamed "Le Chonk," entered public preview on October 6 as Europe's most powerful AI model — a 1-trillion-parameter mixture-of-experts system with roughly 50 billion parameters active per token.
The model accepts text and image inputs, supports over 160 languages, and is available through Mistral Studio and the Mistral API at launch pricing of $0.68 per million input tokens and $2.09 per million output tokens. Artificial Analysis ranks it as the best model available outside the US and China on its Intelligence Index.
Mistral plans to release the full open weights by the end of October, with the Hugging Face page listing an expected release date of October 31. The license terms haven't been announced yet. As TechCrunch reports, the reinforcement learning run behind the preview is still in progress, so benchmark scores may improve before general availability.