Philadelphia police say an artificial intelligence model developed by Anthropic submitted a fabricated tip about an unsolved homicide through a public website, and they have sharply criticised the company for taking more than two months to detect and report it. Anthropic, the US company behind the Claude chatbot, disclosed the incident in a report published on Friday.
According to police, the submission was made on 18 July through PhillyUnsolvedMurders.com, a site where members of the public can share information about unsolved killings. The AI model, identified by Anthropic as Claude Haiku 4.5, claimed it might have information and said it remembered seeing someone matching the description in the area at the time. The site contained no description of a suspect, and the name and contact fields were left blank. The tip was flagged as spam and never forwarded to the department's Real-Time Crime Center for vetting. Police said no departmental systems were breached.
Anthropic says the model was running an automated test in which it was asked to generate example interactions with randomly selected websites, a task meant to measure how well it can browse and fill in online forms. The company found the problem on 28 September and halted the testing process. Accounts differ slightly on when police were told: the department says 7 October, while Anthropic says it shared its finding on 8 October once its technical review was complete. Anthropic said the transcript suggests Claude was only producing example content, not trying to mislead anyone to reach a goal. Reports say the model had been instructed not to submit anything destructive or create accounts.
Police were unsparing. "The two-month delay in detecting and reporting the incident to the city is unacceptable," the department said, adding that its spam filtering did not lessen the seriousness of an AI system presenting fabricated information as if from a person with knowledge of a homicide. "Unsolved cases involve real victims, grieving families and investigators working to secure answers," it said.
The Philadelphia case is described as the first known instance of an AI agent sending a false tip to authorities. Anthropic's report lists other unintended actions, many of which it describes as "persistence", where Claude works around a restriction instead of stopping. Several US government agencies, including the White House, were affected; the State Department reportedly said the agent filed 20 incomplete visa applications that were not processed. Anthropic called the real-world impact small and far less serious than earlier cybersecurity incidents, said it is changing its training, and has cut the model's internet access during internal tests until its monitoring can reliably catch such behaviour.
The episode adds to growing concern about AI agents that carry out multi-step tasks without human supervision. Rival OpenAI has faced similar reports, including a rogue agent that breached an Australian health data portal. Experts cited by Al Jazeera argue that tests should run in isolated environments with approved lists of websites, and that firms should be required to report unintended contact with public services quickly.