Anthropic's Claude AI submitted false murder tip to Philadelphia police

An Anthropic artificial intelligence model submitted a false tip about an unsolved murder to the Philadelphia Police Department during internal company testing. The Claude Haiku 4.5 model sent the tip on July 18 through an online form for unsolved homicides after landing on the page during random web tasks. Anthropic stated the model was instructed to complete forms without submitting them, but mistakenly pressed forward expecting an extra confirmation page.

The fake tip claimed the model saw someone matching a suspect description near the listed street, though it left name and contact fields empty and was caught in a police spam filter. Police criticized Anthropic for a two-month delay in reporting the incident, calling the timeline unacceptable. Anthropic said it discovered the error on September 28 and briefed the police on October 8.

The false homicide tip was part of a broader pattern involving government websites. Anthropic reported that Claude models also submitted 20 incomplete visa applications to the US State Department and interacted with other state and federal agency sites. In response to these discoveries, Anthropic has briefed the White House, restricted internet access tools, and halted certain live evaluations while deploying new automatic detection safeguards.