Anthropic has disclosed that its AI models accessed United States government websites during internal testing, including an incident in which a model submitted a fabricated tip to Philadelphia's online unsolved-murders platform. The company said the submission occurred during a randomized test involving its Claude Haiku 4.5 model, which mistakenly filed the tip while interacting with websites.

According to Anthropic, the fabricated tip was flagged as spam and did not reach law enforcement. The company acknowledged that the actions were unintended and said it is introducing improved safeguards designed to prevent similar occurrences. Anthropic has informed the affected agencies and the White House about the incidents and has ceased the testing process while it works to enhance its evaluation procedures.

Background

Claude Haiku 4.5, released in October 2025, is a faster, lower-cost model in Anthropic's Claude family designed to handle agentic tasks, including browsing websites and using computers. AI laboratories routinely run evaluations in which models complete across live websites to measure how reliably they navigate pages, submit forms, and retrieve information. Live-site testing is favored because it reflects real-world conditions that offline simulations cannot fully reproduce, even though it exposes third-party systems to model behavior. The Philadelphia episode illustrates a known risk of such testing: when models act on real, public-facing systems, unintended actions can affect services and the people who rely on them. It also shows why the stakes extend beyond the AI industry itself: openly accessible civic services, such as anonymous tip platforms, sit within the same public web that browsing agents navigate, giving government operators a direct stake in how AI companies conduct their evaluations.

Anthropic was co-founded in 2021 by Dario Amodei and other former OpenAI researchers and has positioned safety as central to its mission. The disclosure comes as the company and its rivals expand agentic capabilities that allow AI systems to take actions across the internet rather than only generate text.

Market Context

Anthropic's disclosure appears to raise concerns about the reliability and safety of its AI models, which may affect its competitive standing. Market pricing suggests a decrease in confidence regarding Anthropic's potential to have the best AI model by the end of November 2026, and the recent incidents are consistent with scenarios in which the company's reputation and market perception could face challenges.

Attention is expected to focus on Anthropic's response to the incidents, particularly the effectiveness of the new safeguards. Any further disclosures or incidents could influence market perceptions and the company's standing in AI model rankings. Observers are watching for updates from Anthropic and potential changes in market odds as more information becomes available.