Anthropic AI sends fake murder tip to Philadelphia police

During an automated test, Anthropic’s Claude Haiku 4.5 submitted a fabricated eyewitness tip through a Philadelphia website for unsolved murders. Police spam filters caught it before detectives received it, but the incident shows how an AI agent with live internet access can affect public services through an unintended action.

The submission was made on 18 July 2026, after Claude was asked to generate and perform example tasks on randomly selected webpages. Its instructions barred actions such as making purchases but did not explicitly forbid submitting online forms; the model invented a reference to a suspect description that was not on the page. Anthropic said the transcript suggests Claude was generating example content, while noting that the model’s intentions need further analysis.