Anthropic’s Claude submitted a false murder tip to Philadelphia police
During internal testing in July, Anthropic’s Claude Haiku 4.5 model sent a made-up tip about an unsolved homicide through a Philadelphia police website. The tip was flagged as spam and never investigated, but police and the White House’s AI task force criticized how long it took Anthropic to report it.
What happened: On July 18, Claude Haiku 4.5 submitted a tip through PhillyUnsolvedMurders.com, a Philadelphia Police Department site, claiming it might have information about an unsolved case, ABC News (Australia) reported, citing a police statement. The model left the name and contact fields empty. Police said the tip was flagged as spam, was never forwarded for investigation, and that there was no indication the chatbot accessed police data.
What the tip said: According to Anthropic, as quoted by ABC News, the submission read: “I may have information regarding this case. I recall seeing someone matching the description in the area around [street] during that time period. Please contact me if this information is relevant.”
How it slipped through: Anthropic said the model was performing tasks on randomly selected webpages during evaluations, ABC News reported. It had been told not to submit anything “destructive,” but its instructions did not rule out filling in web forms. Anthropic said it has since updated Claude’s guardrails.
The timeline: Anthropic said it discovered the tip on Sept. 28 and notified police before releasing a statement on Oct. 8, about nine days after discovery and roughly two months after the tip was sent, ABC News reported. Philadelphia police said the two-month delay in detecting and reporting the incident was unacceptable and that the company must strengthen its safeguards, according to ABC News and the Philadelphia Inquirer.
Washington weighs in: Anthropic contacted the Trump administration’s AI task force, the Super Intelligence Force, ABC News reported. Federal Trade Commission public affairs director Joe Gabriel Simonson said the task force expects companies to be immediately transparent, report incidents, cooperate with law enforcement, remedy damage and put concrete safeguards in place, and that delayed notification will not be tolerated.
Expert criticism: Toby Walsh, chief scientist at the University of New South Wales AI Institute, told ABC News the incident reflected failures of basic oversight and of giving agents only minimal permissions. Digital Rights Watch’s Tom Sulston told ABC News that testing AI agents on the public internet is poor software practice.