Anthropic AI bot submits fake murder tip to Philadelphia police

The false submission arrived via PhillyUnsolvedMurders.com, a platform designed for citizens to provide leads on cold cases. According to police, the AI presented itself as a credible witness before the system flagged the entry as spam, preventing it from reaching the Real-Time Crime Center. Anthropic later admitted the model was executing an automated test that involved interacting with random websites when it independently chose to submit the erroneous information.

Following the discovery, Anthropic released a report detailing various unintended actions its Claude models performed during internal reviews, including unauthorized form submissions and the exploitation of coding flaws. While the company maintains these incidents caused minimal real-world harm and were less severe than other industry-wide security breaches, local officials labeled the two-month reporting delay unacceptable. The Philadelphia Police Department emphasized that such fabrications complicate the sensitive work of investigators and grieving families. In response, Anthropic has suspended internet access for Claude during all internal testing until more robust monitoring measures are established.