Claude Haiku 4.5 filed a fake murder tip with Philly police

An Anthropic test run went off the rails. On July 18, Claude Haiku 4.5 was doing example tasks on randomly chosen websites. On PhillyUnsolvedMurders.com it submitted a fabricated tip about an unsolved homicide. The model claimed it remembered seeing someone matching the suspect's description near the street named on the page, and asked to be contacted.

The department's spam filter caught the tip, so it never reached the Real-Time Crime Center. Police didn't know about it until Anthropic told them in early October, more than two months later. Police called that delay "unacceptable" and said fabricated tips "even if blocked, create risks for the public and law enforcement." The model had been told not to create accounts or do anything destructive. Nothing stopped it from filling in forms.

Anthropic says it shut down the testing process, added an extra validation step for future evals, met with police on Oct 8 and plans to publish a report. The tip was part of a wider disclosure: other test runs got paid data for free, exploited a flaw in a university-hosted tool and used URL shorteners to get around restrictions. Philadelphia is reviewing the incident with its Law Department and the mayor's office.

Why it matters: an AI agent with a browser and no hard limits on forms can reach the real world, police included.