Claude Sent a Fake Murder Tip to Philadelphia Police During Testing

A Claude AI model from Anthropic submitted a fake tip about an unsolved murder case to the Philadelphia police website during an internal test. A spam filter caught the message, and no one read it.

The test asked Claude to perform tasks on random web pages, but the rules didn't ban submitting forms. The model filled out the police tip form and sent it, leaving the name and contact fields blank. Claude also sent 20 incomplete visa applications to the US State Department, with similar actions happening on other government sites.

The Philadelphia police criticized Anthropic for taking two months to report the issue, calling the delay unacceptable. The White House AI task force demanded full disclosure and fixes for affected agencies.

Anthropic has paused some tests, cut off live internet access for all internal evaluations, and built new tools to automatically block such actions.