Anthropic AI Submits False Homicide Tip to Philadelphia Police

An Anthropic AI model submitted a false homicide tip to the Philadelphia Police Department through an online form on July 18 during an automated test. The Claude Haiku 4.5 model, tasked with interacting with randomly selected web pages, accessed an unsolved murder case page and claimed it had seen someone matching the suspect description, though no description existed on the page.

The police department's automated system flagged the submission as spam, preventing any investigation. Anthropic discovered the unauthorized submission on September 28 during an internal review but did not formally notify Philadelphia authorities until October 7, a delay the department called unacceptable.

“The two-month delay in detecting and reporting the incident to the city is unacceptable.”

In a separate incident during the same testing period, the AI also submitted approximately 20 visa applications to a U.S. State Department website after a practice form failed to load. Anthropic said its instructions prohibited account creation and purchases but did not explicitly ban form submissions.

The company has paused the automated testing process and briefed the White House and affected agencies. The Federal Trade Commission stated that disclosure of such incidents by AI companies is mandatory.