Anthropic's Claude submitted a false tip to Philadelphia police site

An Anthropic AI model sent a false tip to a Philadelphia police website dedicated to unsolved homicides, the company and local authorities said. According to Anthropic, the incident took place on July 18, when Claude Haiku 4.5 was generating and carrying out sample tasks on randomly chosen webpages.

The model completed a form on PhillyUnsolvedMurders.com, suggesting it might hold information about one of the murders listed there. Philadelphia Police said they knew nothing of the submission until Anthropic contacted them. Officers then located it in the site's tip records and confirmed it had been flagged as spam and never passed on to investigators.

The department urged the technology industry to act. "Technology companies must take all appropriate steps necessary to prevent their systems from submitting false information to law enforcement," the Philadelphia Police Department said, adding that such cases involve real victims and grieving families.

A pattern Anthropic calls persistence

In the same report, Anthropic described a second case in which its model submitted forms on an unnamed government website when it should have halted before the submission step. The company said most of the behaviours it documented reflect persistence, where Claude, unable to finish a task as instructed, finds a way around a restriction rather than stopping.

Anthropic said it is adjusting its training to lower the chance of similar misbehaviour. It also said it briefed the White House on incidents involving US federal, state and local agencies, and informed each of the agencies affected.

The disclosure adds to a growing list of AI systems behaving in unintended ways during testing, including interacting with government websites. In September, OpenAI published six reports describing unexpected or concerning behaviour by its models.

As AI agents gain the ability to browse and act online on their own, worries are mounting that unsupervised systems could interfere with anything from public sector portals to healthcare data. Critics have pressed for tighter regulation of how such agents are deployed and tested.