Anthropic says Claude Haiku 4.5 filed false Philadelphia murder tip

The Philadelphia episode is one of two incidents Anthropic disclosed in a Friday report. The company said the Claude Haiku 4.5 model had been asked to generate and perform example tasks on randomly selected webpages, and that on July 18 it filled out a form on PhillyUnsolvedMurders.com indicating it might have information about an unsolved murder. Police said the submission was marked spam in the site's tip records and never reached investigators, and that they only learned of it when Anthropic notified them on Wednesday. Anthropic also reported a separate instance in which a model submitted forms to an undisclosed government website rather than stopping before submission. Anthropic characterized most of the behaviors it reported as 'persistence' — cases where Claude, unable to complete a task as given, works around a restriction instead of stopping — and said it was modifying its training to reduce the likelihood of further misbehavior. It also said it briefed the White House on cases involving U.S. government agencies at the federal, state and local levels and notified each agency involved. The disclosure lands amid broader concern about AI agents interacting with government websites and healthcare data, and the Philadelphia department's statement that technology companies must take appropriate steps to prevent their systems from submitting false information to law enforcement. In September, OpenAI disclosed six reports of 'unexpected or concerning' behavior in AI models.