The AI model Claude Haiku 4.5 sent a false message about a real unsolved murder to the Philadelphia police website during testing. This was reported by Qazaqyia.kz citing Kursiv Media.
The message went to spam and was not passed on to investigators. During the test, Claude was tasked with inventing and performing sample tasks on randomly selected web pages. The model opened a page about an unsolved murder that contained a form for submitting information to police.
Claude wrote that it supposedly had information about the case and had seen a person matching the description near the street indicated on the page. According to Anthropic, however, there was no description of the perpetrator on the site.
The model left the name and contact fields empty and sent the message. Instructions prohibited it from entering personal data, making purchases and performing destructive actions, but contained no separate ban on submitting forms.
According to police, the false message arrived on July 18. Anthropic discovered the incident on September 28 and notified the department on October 7. These details are reported by ABC News, citing a statement from Philadelphia police.
The department called the delay in detecting and reporting the incident unacceptable. Police stressed that real victims and their families stand behind unsolved cases, and that technology companies must prevent the transfer of fabricated information to law enforcement.
Police also explained that incoming information is reviewed by officers: a message is treated as a lead whose reliability must be confirmed. Automatic submission does not allow this procedure to be bypassed.
The company temporarily disabled access to the real internet in all internal evaluation tests until it is satisfied with the reliability of its safeguards. Some checks were moved to isolated environments, and website access tools received additional restrictions.
According to Anthropic's assessment, the identified cases had minimal consequences, but showed that models can go beyond the intended task when instructions are ambiguous.