Claude sent Philadelphia police a false murder tip — Ars Technica
During internal testing by Anthropic, the artificial intelligence model Claude Haiku 4.5 sent the police in Philadelphia, United States, a false tip regarding an unsolved murder. As Ars Technica reports, the form on the police website was submitted on July 18. Police said the email generated after the tip was submitted went to the spam folder, so no person reviewed it.
Error during testing
Anthropic explained that Claude Haiku 4.5 had been instructed to fill out various web forms up to the final submission stage, but not to submit them. According to the company, in several cases the model mistakenly submitted forms while expecting an additional confirmation page.
During one test, Claude opened a page about an unsolved murder that contained a form for providing information to police. The model stated that it had allegedly seen a person matching the description near a street mentioned in the case. The name and contact information fields were left blank.
Police response and Anthropic's actions
According to Philadelphia police, Anthropic discovered the incident on September 28, stopped the relevant automated test, and introduced an additional verification mechanism. On October 8, the company informed police, after which the department found the submission in the website's records and confirmed that the email associated with it had remained in spam.
More current news is available on the UA.News Telegram channel Telegram.
Police called the two-month delay in discovering and reporting the incident unacceptable. They also noted that the standard procedure involves human review of tips before they are forwarded for further investigative action, and that these safeguards limited the impact of the case.
Anthropic also reported other incidents involving U.S. government websites. According to The New York Times, citing anonymous sources, Anthropic agents submitted 20 incomplete visa applications through a U.S. State Department form, which were not processed.
The company halted some testing, moved some of it to an offline environment, or redesigned it so that tasks would not interact with real websites. Anthropic also decided to extend the shutdown of internet access to all internal evaluations until it confirms the reliability of its safety and monitoring measures. The company said that new tools blocked all the described unwanted actions during retesting.
