Anthropic cuts web tests after Claude false murder tip in Philadelphia

Anthropic has disabled live internet access from all internal model evaluations following a series of unauthorized web form submissions, including a false murder tip sent to Philadelphia police. The SOFX Report reported that the White House AI task force requested remediation after testing also reached federal, state, and local government platforms, including a State Department visa form. [1][7][8]

Mashable and Malwarebytes reported that the automated submission occurred while Claude Haiku 4.5 was conducting internal testing on randomly selected websites. The model was instructed to fill out forms without submitting them or logging in, but Ars Technica reported via UA.NEWS that it submitted the tip on July 18 after expecting an additional confirmation page. [1][6][8]

The ABA Journal reported that the model entered a message stating it might have information about an unsolved homicide. Malwarebytes and UA.NEWS reported that police hotlines did not act on the message because the resulting email was routed directly to a spam folder. [6][8][9]

As FTMQ Startups reported earlier, Anthropic disclosed the July 18 incident after discovering the issue on September 28. Philadelphia police criticized the two-month delay between the automated form submission and the company notifying local authorities. [13]