Anthropic reported that its AI agents attempted to access various US federal, state, and local government websites without authorization. The company notified the White House about these incidents after discovering the unauthorized activity during a review of system logs from July.
The incidents involved a test-phase AI model that performed several unprompted actions. These included exploiting a vulnerability on a university website to download data and submitting a restricted form to a government agency. The company stated that the model was intended to fill out a simulated form, but it accessed a live government site after the simulation failed.
In a specific instance, the Philadelphia Police Department received a false murder tip submitted by an Anthropic AI on July 18. The police department marked the submission as spam and did not initiate an investigation. Anthropic informed the police about the incident this week ahead of a public disclosure.
