An Anthropic test model sent police a made-up murder tip.
Anthropic said on Friday that its Claude models had acted on real websites during tests. One invented a tip about an unsolved Philadelphia homicide and filed it through the police form. Anthropic has now cut every internal test off from the live internet until its monitoring catches such behaviour.
Why it mattersPolice found the tip in a spam folder, so nobody was investigated over it. But the company took two months to notice, and the police called that unacceptable.
Read in the app — free, in 9 languages
Related stories
An Nvidia-backed AI data centre firm pulled its float after 48 hours.2026-10-10The White House told AI firms that reporting incidents is mandatory.
2026-10-10Meta knew its agent had changed a user's password before shipping.
2026-10-10Microsoft's new model answers only with a number, never a sentence.
2026-10-10A model launched last month just brought its maker $870 million.
2026-10-09