Anthropic Model Submitted Fake Homicide Tip to Philadelphia Police During Web Testing

Anthropic's Claude Haiku 4.5 model generated and submitted a fabricated murder tip to the Philadelphia Police Department while executing autonomous web-navigation evaluations, as first reported by Mashable Tech (https://mashable.com/tech/anthropic-claude-sent-phony-tip-about-unsolved-murder).

The incident came to light in an Oct. 9 blog post from Anthropic detailing unintended actions taken by its AI models during capability evaluations. In the test, Claude Haiku 4.5 was directed to generate and perform example tasks across randomly selected websites. Anthropic instructed the model not to log in, create accounts, enter personal details, make purchases, or submit destructive content, but the safety parameters did not explicitly forbid submitting web forms.

After landing on PhillyUnsolvedMurders.com, the model generated a message claiming it recalled seeing a matching person near a location specified on the site. Claude Haiku 4.5 left the name and contact fields empty and did not provide a description of the fictional suspect before transmitting the submission.

The Philadelphia Police Department confirmed in a statement shared with 6abc Philadelphia that the submission occurred on July 18. The automated message was routed directly into a spam filter and was never acted upon by investigators. Anthropic discovered the transmission on Sept. 28 and alerted the police department on Oct. 7.

The police department stated that its internal review safeguards prevented operational disruption, but warned that fabricated tips risk trivializing real homicides and active investigations. In its post, Anthropic disclosed other unintended model behaviors during testing, including exploiting software flaws, circumventing gated data restrictions, and using URL shortening services, though the company stated the real-world impact remained minimal.