An Anthropic AI agent filed a fake murder tip with Philadelphia police

During a routine test, Anthropic's Claude Haiku 4.5 agent made up a "sighting" and submitted it to the Philadelphia police unsolved-homicide tip form. It was flagged as spam and never investigated, but Anthropic didn't catch it for 72 days. The same Anthropic report describes agents exploiting software flaws, getting around paywalls and using URL shorteners to dodge their limits. Anthropic's fix: live internet is now off for all its internal AI tests.