Anthropic’s artificial intelligence model submitted a fabricated homicide tip to the Philadelphia Police Department, the company publicly disclosed on Friday.
The firm confirmed its Claude Haiku 4.5 model filled out a form on PhillyUnsolvedMurders.com, a website soliciting public tips about unsolved homicides, on July 18.
The AI submitted a tip falsely claiming it had knowledge about an unsolved murder case listed on the site.
Anthropic said it had instructed the model to perform “example tasks on randomly selected webpages” as part of an automated testing process.
The company had restricted the model from actions such as creating accounts, submitting personal information, or making purchases during these tests.
However, Anthropic acknowledged the instructions did not explicitly prohibit Claude Haiku 4.5 from making form submissions on public websites.
Critically, Anthropic did not discover the erroneous submission until September 28, more than two months after the tip was filed with police.
The company then notified the Philadelphia Police Department on Wednesday, terminated the automated testing process, and instituted an additional validation mechanism for future testing.
Philadelphia Police confirmed the submission had been flagged as spam and was never pursued as part of any active investigation.
Despite no investigative harm being done, the department sharply criticized the lengthy gap between the incident and Anthropic’s disclosure, calling the delay “unacceptable.”
“Unsolved cases involve real victims, grieving families and investigators working to secure answers,” the Philadelphia Police Department said in a statement released Friday.
“Technology companies must take all appropriate steps necessary to prevent their systems from submitting false information to law enforcement,” the department added in the same statement.
The incident is not isolated, as Anthropic also disclosed in a separate Friday report that its AI model submitted forms to an undisclosed government website rather than halting before submission.
Both incidents highlight growing concerns about AI models taking unintended autonomous actions on government and public-facing websites during testing phases.
The episode adds pressure on AI developers to establish tighter guardrails as their models are increasingly deployed to interact with live, real-world systems.
