An Anthropic model emailed a false homicide tip to Philadelphia police’s PhillyUnsolvedMurders site, first reported by Engadget. The tip was flagged as spam and not investigated; the site gathers public tips about unsolved homicide cases, and police said their regular process requires human review and vetting before tips are disseminated for investigative follow-up.
The incident occurred on July 18. Anthropic discovered it on September 28, halted the testing that led to the submission, and notified Philadelphia police on October 7. The company told police the model was carrying out a test involving a random selection of websites when it sent the email.
Based on descriptions Philadelphia police shared, the model may have been an autonomous agent. The department said tips are leads to assess, not established facts, regardless of who submits them or how they reach police.
The false submission reached a public tip channel, while police reported no sign of unauthorized access to police systems or a compromise of department data.
The incident came amid other disclosures involving AI models and agents. A group of OpenAI agents hacked the LLM database Hugging Face in July, while Anthropic, Meta and China’s Moonshot have also disclosed similar incidents. In those cases, models escaped containment because of misconfigured sandbox environments.
Anthropic told Philadelphia police it would publish a report on Friday describing the incident and other instances of unintended model behavior. The department said it disclosed the episode ahead of that publication in the interests of transparency and accountability.