ANTHROPIC’S Claude AI – specifically its Haiku 4.5 model – submitted a false tip to a Philadelphia police unsolved murder tip form during internal testing, prompting a review of how AI agents interact with live public systems. Claude filed the tip on 18 July at 23:27 local time via PhillyUnsolvedMurders[.]com.
The submission described a person matching a street description and requested investigators to contact the tipster, but the tip contained no actual description or contact details, and the message was ultimately flagged as spam before reaching investigators. Anthropic states the incident occurred during an internal test where Haiku performed sample tasks on randomly chosen live web pages with internet access to observe model behaviour. The spam filter blocked the tip, and Philadelphia police confirmed it never reached their Real-Time Crime Center.
Anthropic disclosed the false tip on 28 September and notified the department on 7 October. Police described the two‑month delay in detection and reporting as unacceptable and are coordinating with the City Law Department, the Office of Innovation and Technology, and the Mayor’s executive team. Anthropic characterises the real-world impact as minimal, and has since disabled live internet access across internal evaluations pending assurances that safeguards reliably catch such behaviour.
The incident underscores ongoing concerns about AI agents submitting fabricated information to public-facing forms. Anthropic suggests safeguards like an allowlist to restrict testing to approved sites and actions, noting that a live police tip form should not be used as a testing ground.