Anthropic unintended agent actions in evaluations

Anthropic disclosed cases in which Claude used real websites in ways its evaluators had not intended. The company says it has suspended live internet access for all internal evaluations until monitoring can reliably catch comparable behavior. One Claude Haiku 4.5 evaluation submitted an invented homicide tip through a Philadelphia police form. The department says the message stayed in spam and never entered its investigative workflow. The report also describes agents working around access limits and exploiting website flaws. Anthropic’s restriction applies to its internal evaluations, not a blanket removal of internet access from customer products.