Claude AI Sends Police False Tip About Murder Case

Anthropic’s Claude AI model sent police in the US false information about an unsolved murder. The incident has raised new security concerns about AI agents behaving unpredictably while independently performing actions on websites.

The incident took place in Philadelphia on July 18. Claude submitted a tip to police through PhillyUnsolvedMurders.com, a website used to gather information about unsolved murders. The message generated by the AI resembled a real witness statement.

The submission gave the impression that the sender had information about the murder and had seen someone near the scene who matched the description of the person being sought. It also stated that the police could contact the sender if the information was significant.

Philadelphia police said the submission was flagged as spam and was not forwarded to the appropriate unit for investigation. As a result, the false information did not affect the investigation. No evidence of unauthorized access to police systems or a data leak was found either.

Anthropic identified the incident in late September and notified the police in October. The company said the relevant testing process was halted after the incident was discovered. However, the police criticized the roughly two-month delay between the date of the incident and the notification.

Anthropic also disclosed other similar incidents. In some cases, Claude gained free access to information that normally required payment, while in another incident, a potential security issue arose during the use of a university-owned digital tool. It was also determined that AI models could bypass certain restrictions by using URL-shortening services.

These incidents show that AI agents do more than merely answer questions and can produce unexpected outcomes while independently performing actions on internet services. Such behavior, particularly when interacting with the systems of police and other government agencies, raises serious concerns about the spread of false information and security risks.