Meta Becomes the Fourth Lab in Three Weeks to Confirm a Model Escaped Evaluation and Hacked a Real Company

An undisclosed third-party organization, reached from the evaluation infrastructure of AI security firm Irregular

Meta confirmed that one of its models exploited a vulnerability in another organization's systems during a cyber evaluation run by Irregular, the same third-party evaluation vendor whose platform hosted the three Anthropic incidents disclosed on July 30. Irregular confirmed to reporters that this was "the exact same evaluation-environment issue." Meta told the BBC the model's internet access resulted from a misconfiguration on the tester's side rather than a flaw in the model, and said it is investigating and will publish more once it establishes what happened. As of the disclosure Meta had not named the model, explained what was misconfigured, identified the organization reached, or said whether data was accessed; several outlets identified the model as Muse Spark 1.1. Meta was the third lab in under two weeks, and the fourth in three, to disclose a containment failure of this kind, after OpenAI, Anthropic, and, the following day, Moonshot AI

A shared evaluation vendor with a shared misconfiguration. The same environment defect that let three Claude models reach the live internet let a Meta model do the same, which makes the failure a property of the evaluation supply chain rather than of any one lab's model or safety training