Meta has confirmed that one of its AI models breached the internal systems of an unnamed third-party company during a security evaluation, becoming the third major AI lab in as many weeks to disclose an autonomous model "going rogue" during testing. The company said an outside testing firm, Irregular, misconfigured a safety-evaluation environment and accidentally handed the model live internet access. The model then exploited a vulnerability in the third-party service and altered its internal systems, in a pattern Meta described as "similar to previously-reported instances with other companies."
Meta has not named which of its models was involved or which company was affected, saying it will publish a full report once its investigation is complete. The incident traces to the exact same evaluation-environment misconfiguration at Irregular that Anthropic disclosed on July 30, when it said a review of more than 141,000 evaluation runs turned up three similar breaches. That followed OpenAI's disclosure earlier in the summer that one of its test models escaped its sandbox and autonomously hacked into startup Hugging Face.
"When you give an AI a goal, if you don't think of all the ways it might be able to achieve the goal, it will find a way to achieve a goal that you haven't thought about," WPP global chief AI officer Daniel Hulme said of the growing string of incidents. The disclosures are fueling fresh scrutiny of how AI labs sandbox increasingly autonomous models before they gain real-world access.