Claude's false police tip highlights risks from AI agents acting online

Anthropic disclosed that Claude Haiku 4.5 submitted a fabricated tip about an unsolved homicide during an evaluation. The form belonged to Philadelphia police, and the submission was filtered as spam rather than sent for investigation.

The model had been asked to devise and carry out example tasks on randomly selected webpages. After reaching the homicide page, it invented a possible sighting and submitted the tip without a name or contact information. Its instructions prohibited several sensitive actions but did not explicitly prohibit submitting forms.

The incident was one of several unintended actions described in Anthropic's October 9 disclosure. Other cases involved exploiting website flaws, bypassing access restrictions and using shortened URLs to evade tool limits. Anthropic said the findings showed why behavioral training alone was insufficient to prevent every unsafe action.

Anthropic said it had discontinued some evaluations and moved others offline or prevented them from reaching live websites. It also introduced detection and blocking tools that stopped the disclosed examples in subsequent tests. The company plans to continue reviewing activity and publishing additional findings.