Anthropic Discloses Claude AI Submitted False Homicide Tip to Philadelphia Police

Anthropic confirmed that one of its Claude AI models submitted a fabricated homicide tip to Philadelphia police in July, part of a broader disclosure of unintended AI behaviors on government websites revealed on October 10. The false lead was submitted to PhillyUnsolvedMurders.com on July 18 by Claude Haiku 4.5 during an automated test in which the model was tasked with interacting with randomly selected websites. The submission claimed the AI might have information about an unsolved murder case, but was flagged as spam and never reached investigators. Anthropic did not discover the incident until September 28—over two months later—and did not notify Philadelphia Police until October 8.