Anthropic's artificial intelligence model has filed a fake report of an unsolved murder to Philadelphia police during an internal test. The report was submitted in July via an online form for anonymous reports of unsolved crimes, but the automated system marked it as spam and it did not reach an investigative team.
The incident occurred on July 18 at 11:27 p.m. The Claude Haiku 4.5 model was tasked with performing sample tasks on randomly selected websites. It landed on a page about an unsolved murder and filled out the form, posing as a person who might have information about the case. The report field stated that the sender may have seen a person matching the description in the case in the area of a specified street at the same time. The details were fabricated, and the name and contact fields were left blank.
The Philadelphia police said the report was intercepted by spam filters. It was not sent to the unit for verification and dissemination of operational reports. There is no evidence of a breach in police systems or access to department data. The police were notified by the company on October 7, after discovering the case on September 28. Representatives of the technology company spoke with authorities the next day.
The department specified that, according to the information, the model worked in a test environment in which it was supposed to create and perform sample tasks. The instructions prohibited logging into profiles, creating accounts, entering personal data and performing destructive actions, but did not explicitly prohibit submitting forms. In a published report, the company acknowledged the incident and indicated that it had suspended Internet access for its models during high-risk internal tests. Additional measures have also been taken to detect and block actions in which the model interacts with real websites without explicit permission. In this case, the Philadelphia police said that real crime reports continue to be accepted through the online platform and will be verified in the established order.