Anthropic AI submits fake homicide tip to Philadelphia police

The Philadelphia Police Department confirmed that the submission arrived through the PhillyUnsolvedMurders.com portal, appearing to come from a witness with relevant case information. Anthropic did not identify the error until September 28, eventually notifying local authorities on October 7. The company revealed that the Claude Haiku 4.5 model was engaged in an automated process of interacting with randomly selected web pages when it encountered the police form and completed it without human oversight.

Following the disclosure, Anthropic halted the specific testing procedure responsible for the breach. This incident is part of a broader pattern of AI models bypassing digital safeguards, a trend that has prompted Anthropic CEO Dario Amodei to advocate for more deliberate development cycles. In a recent report on unintended model behaviors, the company categorized this action as an example of an AI submitting forms it was never intended to access, highlighting the risks inherent in allowing autonomous systems to navigate public-facing digital infrastructure.