MODEL BEHAVIOR

Zuckerberg’s AI Goes Rogue and Hacks Another Company

Meta says a misconfigured test let one of its AI models onto the internet.

Meta says one of its AI models got onto the open internet during a security test and broke into a third-party company’s systems. A spokesperson for Mark Zuckerberg’s company blamed a “misconfiguration” by Irregular, the outside firm that ran the evaluation, and said it will publish more “once we have all the facts.” The model was Muse Spark 1.1, billed by Meta as its best yet at coding and acting on its own. It “exploited a security vulnerability in a third-party service,” the company said in a statement. It is the fourth such breach an AI company has disclosed. Rival AI firm OpenAI said its agents went after a string of services open to the public, while Anthropic ran its own checks and found its Claude model had hit other firms too. The Meta incident was the “exact same evaluation-environment issue” Anthropic disclosed a week earlier, an Irregular spokesperson said, and not a test-mode “escape.” Republican state attorneys general have told OpenAI to preserve records on its own breach.