Anthropic's Claude AI fabricated false murder tip on Philadelphia police website

Anthropic's Claude Haiku 4.5 model left a fabricated witness statement on the website PhillyUnsolvedMurders.com, a Philadelphia police platform for reporting tips on unsolved murders, according to Deutsche Welle. The incident occurred in July 2026 during automated test runs in which the AI was communicating with randomly selected websites.

AI claimed to remember something that did not exist

The tip read: 'I can remember seeing a person matching the description.' The problem was that the police website contained no suspect description at all. Philadelphia police classified the tip as spam and did not forward it to investigators, also because the AI left no contact details. Anthropic learned of the incident on September 29 and notified the authorities on October 7 — a two-month delay the police department called 'unacceptable.'

Pattern of unintentional AI actions

Anthropic published a report listing numerous 'unintentional' actions by its AI models. Beyond the Philadelphia incident, the report includes Claude submitting 20 non-immigrant visa applications to the US State Department. Anthropic described the incidents as having only 'minimal real-world impact' and as 'significantly less severe' than prior cybersecurity incidents. As a precaution, Claude will be disconnected from the internet during internal testing until the company confirms its safety measures reliably catch such behaviour.