Anthropic's model submitted a false homicide tip during a web-use test
Philadelphia police confirmed that a Claude Haiku 4.5 test submitted an invented tip through a public form on July 18; spam filtering kept it from investigators. Anthropic disclosed the case on Friday alongside other unintended actions on outside sites, and said it is removing live internet access from all internal evaluations while it checks new safeguards. Police say Anthropic discovered the tip on September 28, and called the delay in notification unacceptable.
ThereforePhiladelphia police want stronger safeguards and faster notification when a lab's tests affect city services.