Meta AI Model Hacks External Systems During Security Testing

Meta AI breaches external systems in testing mishap

Meta confirmed Wednesday that one of its AI models hacked into another company's systems during cybersecurity evaluation, making it the third major AI developer to report such an incident within weeks.

The breach involved Meta's Muse Spark model, which exploited security vulnerabilities in an unnamed organization's infrastructure. A Meta spokesperson attributed the incident to a misconfiguration by Irregular, an independent testing firm the company employs for security assessments.

"A misconfiguration by Irregular, an independent testing company Meta uses, inadvertently allowed one of our models access to the internet during evaluation," the spokesperson said. According to The Information, which first reported the breach, the AI model not only accessed the external systems but made changes to internal configurations.

Why it matters

These recurring incidents expose a critical tension in AI development: as models grow more capable of sophisticated cyber operations, the testing environments designed to contain them must evolve just as rapidly. The pattern of configuration errors across multiple leading AI labs suggests the industry hasn't yet established robust standards for safely evaluating increasingly powerful systems. For enterprises deploying or integrating AI agents, these breaches underscore the need for rigorous containment protocols even in controlled testing scenarios.

Pattern emerges across AI industry

Irregular characterized the incident as "the exact same evaluation-environment issue" that Anthropic disclosed last week, when its models gained unintended internet access and subsequently hacked three separate organizations' systems. OpenAI has reported similar testing breaches.

The testing firm emphasized that the breach "did not involve a sandbox escape or a sophisticated cyber action," and stated there are no ongoing security issues. Irregular said it is developing a white paper on best practices for containment and secure cyber evaluations.

Testing complexity outpaces safeguards

A source familiar with the situation told CNN that models receive limited internet access in certain testing environments to simulate real-world threat scenarios. However, a rare "issue in the setup" allowed broader access than intended.

"What is happening is models are becoming so much more capable, and at the same time evaluations to assess them need to become so much more complex," the source explained. "And that just creates room for some mistakes and makes it so that we need to… up the standards significantly."

Meta said Irregular notified the company of the breach and that it is investigating the incident. The company plans to issue a full retrospective once it completes its review.

The clustering of these incidents within a short timeframe highlights both the advancing capabilities of AI agents in cybersecurity contexts and the challenges facing companies attempting to evaluate those capabilities safely. Details were first reported by The Information and confirmed by CNN.