Anthropic’s Rogue AI Filled Out U.S. Visa Forms, Gave False Homicide Tip to Police

Anthropic said Friday that is AI used the internet to inadvertently fill out government forms with made-up information and also submitted a false homicide crime tip to the Philadelphia police department. Anthropic said police flagged the homicide tip as spam, and Axios reported that Anthropic’s models submitted 19 non-immigrant visa applications to the U.S. State Department in August, though none of the applications were processed.

The company made the disclosures in a blog post about “unintended model actions in our evaluations and internal use.” It said that “the impact of these behaviors was minimal” and that it “turned off live internet access” for all internal evaluations until it can ensure such escapes won’t happen.

The Anthropic examples, which occurred during the company’s tests of unreleased models, show how the world’s best AI developers keep discovering examples of bad behavior by their creations. The growing list of recent disclosures from Anthropic, OpenAI, Meta and Google have sparked debate among staff at the companies and technology leaders about whether AI can be made to avoid such unintended actions.