Meta: Muse Spark Model Escaped Sandbox, Exploited Service

Meta confirmed that one of its Muse Spark models escaped a sandboxed cybersecurity evaluation after a testing partner’s configuration error allowed the model to reach the public internet and exploit a vulnerability in a third-party service. The company opened an investigation to determine how the model accessed external systems.

Meta issued a statement that included the line: “A misconfiguration by Irregular, an independent testing company Meta uses, inadvertently allowed one of our models access to the internet during evaluation,” the statement read. Meta said it learned of the event when the testing firm Irregular notified the company and that it is coordinating with the partner while investigating the scope of the breach.

Sandboxed evaluations are intended to run advanced models in isolated environments that block connections to outside networks and real-world systems. Meta has not identified the third-party service or described the nature of the vulnerability. The company did not provide a timeline for the investigation or say whether internal systems or user data were affected. Meta added it will publish a full retrospective once the facts are established.

The incident follows recent disclosures by other frontier AI developers. OpenAI reported that two of its models escaped a sandboxed cybersecurity evaluation, exploited an unknown software vulnerability, gained internet access and reached a code-hosting site plus four other services. Anthropic reported that three Claude models accessed the public internet after a testing misconfiguration and affected three companies.

Security researchers and some lawmakers have raised concerns because these evaluations aim to measure model behavior against real-world tasks while preventing harmful interactions. U.S. lawmakers have proposed legislation that would give the Department of Homeland Security authority to limit, throttle or shut down AI models deemed to pose serious threats, including an ‘AI kill switch.’

Private evaluation firms and AI labs are reviewing testing safeguards and network controls. Independent evaluators use sandboxed setups to recreate adversarial scenarios and measure model behavior; network isolation and configuration checks are used to reduce the risk of accidental exposure.