artificial intelligence Claude sent to the Philadelphia Police a made-up clue about an unsolved murder in which he claimed having seen a possible suspect. Anthropicthe company that develops it, has recognized that the message came from its model Haiku 4.5.
The incident occurred on July 18 during an evaluation in which Claude had to propose and perform various tasks on randomly chosen pages. These tests allow us to check how an AI performs when interacting with real websites.
The model reached PhillyUnsolvedMurders.com, a Philadelphia Police website that collects citizen information on unsolved homicides, and found a form to provide clues. The instructions received by the AI prohibited logging in, creating accounts, entering personal data or making purchases, but they did not expressly prevent sending forms.
Claude filled out the form with a message that presented made-up information such as an observation made by a real person. ‘I may have information about this case. I remember seeing someone matching the description around (street) during that period. Please contact me if this information is relevant.’wrote.
The form allowed you to send notices without indicating a name or contact information and Claude left those fields blank. Furthermore, the website It did not even include a description of the perpetrator of the crime with which the alleged suspect could coincide.
According to Anthropic’s preliminary analysis, Claude appeared to be generating example content to accomplish the taskwithout trying to deceive anyone to achieve a goal. The notice was blocked as spam and never reached the center that reviews the clues before transferring them to investigators.
Anthropic discovered what happened on September 28 and notified police on October 7, according to the department. Agents located the notice in the records and They described the delay in detecting and reporting the incident as ‘unacceptable’.
The company has announced the suspension of direct internet access in all its internal evaluations until you verify that your security measures reliably detect these behaviors. It has also moved some tests to offline environments and added tools to block inappropriate actions.