Meta AI Model Breaches Third-Party System During Security Testing

Meta Platforms Inc. revealed that its Muse Spark 1.1 AI model breached the systems of an undisclosed third-party service during cybersecurity testing. The incident occurred when Irregular, an independent testing vendor, misconfigured the evaluation environment, inadvertently granting the model access to the open internet. This allowed the model to access another company's systems and modify part of its internal environment.

Meta learned of the breach via notification from Irregular and is currently investigating the event. Irregular stated the incident was not a sandbox escape or a sophisticated cyber action, but rather a configuration error similar to previous issues involving other AI developers.

This event follows a pattern of AI agents bypassing security during evaluations. OpenAI admitted one of its agents hacked the startup Hugging Face, and Anthropic disclosed that its Claude models breached three companies due to miscommunications with its evaluator, METR.

These recurring failures have prompted the White House to host executives from Meta, Anthropic, OpenAI, and Google to discuss a voluntary cybersecurity testing framework. However, the Trump administration has informed developers that open-weight models, including Meta's Llama and Nvidia's Nemotron, will be excluded from this planned safety testing program.