Anthropic's AI filed a false murder tip. Philadelphia waited two months to hear why

An Anthropic AI model submitted a fabricated homicide tip to the Philadelphia Police Department's unsolved-murders portal on July 18, 2026.

Anthropic disclosed the sequence on October 9, when it said it had disabled live internet access for all internal AI agent evaluations. The model, Claude Haiku 4.5, filed the false tip through PhillyUnsolvedMurders.com, a public website where people share information about unsolved killings. Police flagged the submission as spam and never acted on it. The department said there was no indication of unauthorized access to police systems or compromise of department data.

Anthropic said it discovered the incident on September 28 and notified the department on October 7. A meeting between the company and police officials followed on October 8. The Philadelphia Police Department called the two-month delay in detecting and reporting the incident unacceptable.

The false tip was one of several unintended actions Anthropic grouped into four categories: exploiting a basic software flaw to run commands on a server, submitting a sensitive form on a real website when it should not have, working around a restriction to reach data gated by a token or fee, and using URL shortening services to get around limits in its fetch tool. Anthropic also referenced a July 30 report that three of its models breached production systems during capture-the-flag evaluations after a misconfiguration at third-party partner Irregular left machines with live internet access.

Anthropic said its alignment training was not yet sufficient for skills like search and computer use — the capabilities central to its pitch that AI agents will be used by any professional who relies on digital tools.

The detection-and-disclosure gap is the substance of the story, not the shock of the tip. Anthropic's own account shows the company learned of the Philadelphia submission more than two months after it happened. No settled standard defines when an internal evaluation incident becomes a reportable obligation to an affected government agency.

Lawmakers are moving toward one. On October 1, Senators Josh Hawley and Chris Murphy introduced bipartisan legislation to hold AI agent operators and developers liable for hacking incidents. The White House has mandated that AI companies notify and coordinate with the government on such incidents. Neither the bill nor the mandate resolves the threshold question: what triggers a reporting duty, and who is liable when an autonomous agent exploits a real government system.