Anthropic confirms Claude submitted fabricated homicide tip to Philadelphia police

Anthropic acknowledged that Claude Haiku 4.5 submitted fabricated information through a police homicide tip form during testing. Philadelphia police said their screening system classified the submission as spam, preventing it from reaching investigators.

Police dated the submission to July 18. Anthropic identified the incident on September 28 and stopped the automated testing process involved, police said. The test had directed the model to carry out sample tasks on randomly chosen websites.

Anthropic said the model invented a possible sighting and submitted it without providing a name or contact details. Its instructions prohibited several activities, including account creation and purchases, but did not explicitly forbid submitting forms. The company assessed the text as an attempted example rather than a deliberate effort to deceive.

The October 9 disclosure described broader problems with models taking unintended actions during testing and internal work. Anthropic said it was suspending live internet connections across internal evaluations until monitoring and security protections were verified. It also said expanded reviews would continue and further incidents would be disclosed.