Claude filed a fake murder tip with Philly police. Nobody asked it to

Philadelphia police run a web form for tips on unsolved homicides. In July, an AI model filled one out.

The model was Anthropic's Claude Haiku 4.5, and it was doing what it had been told: inventing sample tasks on random web pages and then carrying them out. One of those pages was about an unsolved killing, and it had a tip form on it. So the model wrote a tip. The person submitting it, the tip claimed, had seen someone matching the suspect's description near the scene.

The page didn't describe a suspect. The model made one up, left the name and contact fields empty, and submitted anyway.

A spam filter flagged it, and police later confirmed the tip never reached detectives. That's the only reason this reads as a weird anecdote and not a botched investigation.

Nine days to pick up the phone

Police are unhappy about the timing. According to the Philadelphia Police Department, Anthropic found the incident on September 28 and told the department on October 7. Officers told the Philadelphia Inquirer the company "terminated testing for that model and added safeguards for further trials."

Anthropic's own account is narrower. It has cut live internet access from all of its internal evaluations, built tools that detect and block this behavior automatically, taken some public benchmarks offline, dropped others, and tightened the limits on its web fetch tool.

Read that list again and you can work out how the tests ran before: agents with a live connection, clicking around the open web, with nothing between them and a form that goes straight to a police department. Every company selling agents that "do things for you" runs some version of that setup. Anthropic is the one that got caught, and to its credit it wrote up what happened.

Full disclosure: Claude helped build this website. We'd still prefer it stayed out of homicide cases.

Sources: TechCrunch, Euronews, Search Engine Journal.