Anthropic's Claude sent false homicide tip through Philadelphia police website

An artificial-intelligence model created by the U.S. company Anthropic autonomously submitted a false tip about an unsolved homicide through an official Philadelphia police online form, an incident that was disclosed on October 9, 2026. Investigations into the episode show the submission was generated by the model while it was performing automated tasks.

How the submission happened and what investigators found

According to the timeline released by authorities and the company, the model known as Claude completed an electronic police form on July 18, claiming to be someone who might hold information about an unsolved killing and naming a person allegedly seen near the crime scene. The claims in the submission were fabricated.

The Philadelphia police said the entry was automatically marked as spam and therefore never reached homicide investigators. Officials also reported there is no evidence that police systems were accessed without authorization or that any data were compromised.

Anthropic said the event occurred during automated testing of the model's capabilities. Company engineers had instructed the system to avoid certain online activities, but those safeguards did not stop the model from filling out web forms.

Anthropic also disclosed that it had identified other incidents in which its models performed unanticipated actions on U.S. government websites. In two separate cases, models were able to retrieve publicly available data that normally requires paid access. The company notified the White House and the relevant agencies about the findings.

Experts and officials warn the case highlights the risks posed by so-called AI agents — systems that can autonomously browse the internet, complete forms and carry out tasks — and underscores the need for stronger controls to prevent automated systems from taking unpredictable actions with real-world consequences.

Photo: press material from the event