WASHINGTON: Meta has confirmed that one of its artificial intelligence models accessed another company’s systems during a cybersecurity evaluation, adding to growing concerns over the ability to safely contain increasingly capable AI models.
The incident, disclosed by Meta on Wednesday, follows similar testing-related breaches involving AI systems developed by Anthropic and OpenAI, raising fresh questions about cybersecurity risks associated with advanced artificial intelligence.
According to Meta, the incident occurred after Irregular, an independent company conducting cybersecurity evaluations on its behalf, accidentally misconfigured the testing environment, giving the AI model access to the internet.
In a statement, Meta said the model “exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies.”
According to The Information, the model involved was Muse Spark 1.1, which Meta has described as its most advanced system for real-world coding and agentic tasks. The report said the model accessed an unidentified company’s systems and modified its internal environment.
Responding to the incident, an Irregular spokesperson told Reuters it was “the exact same evaluation-environment issue that was already disclosed by Anthropic last week” and did not involve a “sandbox escape or a sophisticated cyber action.”
“There are no current open issues. Irregular is developing a white paper to share best practices for containment and securely running cyber evaluations,” the spokesperson added.
The latest incidents have heightened concerns among US policymakers about whether increasingly powerful AI systems could be exploited to carry out or assist in cyberattacks.
Separately, a group of Republican state attorneys general has asked OpenAI to preserve documents related to its recent Hugging Face security breach. OpenAI has said it will comply with the request and publish a technical report on the incident.
Earlier this week, the White House hosted representatives from Meta, Anthropic, OpenAI and Google to discuss a newly finalised voluntary cybersecurity testing framework for advanced AI models.
According to Reuters, the Trump administration informed AI developers that open-weight models, including Meta’s Llama and Nvidia’s Nemotron, would not be included in the proposed voluntary AI safety testing framework.
