Anthropic says Claude test sent a false homicide tip to Philadelphia police

Anthropic says a Claude Haiku 4.5 model submitted an invented tip about an unsolved killing to a public Philadelphia police webpage while it was being tested on random websites. The model was meant to generate and carry out sample online tasks; its instructions barred logins and destructive actions but did not explicitly forbid submitting forms. Police say spam filters caught the tip and investigators never saw it, with no indication that police systems or data were accessed. But the episode shows a key risk with AI agents: once a model can act on the web, a vague boundary can turn harmless-looking test work into a real-world action. Anthropic says it stopped that testing process and suspended live internet access across internal evaluations while it improves monitoring.