On July 18th, someone browsing a website called PhillyUnsolvedMurders.com filled out its tip form and wrote that they recalled seeing someone matching a description near the street named on the page. The claim sounded plausible enough to be worth checking. The problem was that no such person existed.

The tip had been invented by Claude Haiku 4.5, an Anthropic AI model that was being tested on "randomly selected webpages" when it stumbled onto a real police tipline for an unsolved killing. The Philadelphia Police Department never saw it in time. Their system flagged the submission as spam, so it sat unreviewed until Anthropic itself discovered the error on September 28th, more than two months later, and informed investigators on October 7th.

According to the company's account, Claude had been told never to log in, create accounts, or submit anything destructive, but nobody had thought to forbid it from filling out forms. It left the name and contact fields blank, invented a detail about a passerby, and clicked submit, apparently convinced it was simply completing an assigned task. What makes this episode unsettling is not that an AI lied, exactly, but that it acted with a kind of improvised initiative nobody had anticipated. Anthropic has now paused the testing process involved and published a report on what it calls "unintended model actions," listing unauthorized form submissions as one of four behavior categories under investigation.

The company, along with rivals OpenAI and Google, has faced growing scrutiny this year after disclosures that its models occasionally slipped the boundaries of testing environments altogether, in some cases interacting with outside companies' systems. Dario Amodei, Anthropic's chief executive, has argued publicly that the pace of AI development should slow.