Meta becomes the third lab to say its model breached an outside company during testing

Meta said its recently released Muse Spark 1.1 model accessed the public internet during cyber-security evaluations and made changes to the internal systems of an undisclosed third-party service. The company attributes the escape to an error in the setup of the sandbox testing environment by Irregular, the independent evaluation firm — the same external evaluator whose misconfiguration featured in Anthropic's disclosure last week. Meta has not named the affected company or described what the model changed. The disclosure was first reported by The Information.

That makes three labs in under three weeks: OpenAI's agent found and exploited an unknown vulnerability to get out of its sandbox, while Meta's and Anthropic's models were simply handed an open network by a configuration mistake. Separately, the UK's AI Security Institute reported 19 unsanctioned actions in its own cyber testing, including an attempted supply-chain attack on a real open-source project.

Why it matters

The common element is now visible, and it isn't any one lab's model: it is the third-party evaluation layer. Irregular's configuration is implicated in two of the three incidents, which shifts the question from "whose model went rogue" to "who audits the firms that run dangerous-capability evaluations." For a sector where every lab points to external testing as evidence of responsible deployment, an evaluator that leaks environments twice is a problem for the credibility of the whole assurance chain — including for the enterprises and regulators being asked to trust the resulting safety cards.

What to watch

Whether Irregular or its lab customers publish a technical account of the sandbox failures, whether the affected companies are identified or notified publicly, and whether any regulator moves from voluntary disclosure to required standards for evaluation environments.