Meta's AI model hacked another company during security testing
What's the story
Meta has confirmed that one of its artificial intelligence (AI) models compromised another company's systems during a cybersecurity evaluation. The incident involved Meta's Muse Spark 1.1, an advanced model designed for real-world coding and agentic tasks. A configuration error allowed the AI to access the public internet during testing, leading to unauthorized access and modification of internal systems at an unnamed company.
Incident details
A series of breaches by AI models
The breach was traced back to a misconfiguration in the "sandbox" testing environment, which is usually meant to keep AI systems isolated during security evaluations.
This incident adds to a string of recent cases where advanced AI agents from top developers have accessed external systems during controlled tests.
Last week, Anthropic revealed that some of its models breached three companies during cybersecurity evaluations.
Exploitation details
Misconfiguration by Irregular led to breach
The evaluation was conducted in partnership with an external testing firm, Irregular.
A spokesperson for Meta said that the issue stemmed from a configuration error by Irregular, which allowed the AI model to exploit a vulnerability in another third-party service.
"The model exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies," Meta said.
Response
Irregular clarifies the breach was not a 'sandbox escape'
An Irregular spokesperson told Reuters that this was the "exact same evaluation-environment issue that was already disclosed by Anthropic last week" and stressed it didn't involve a "sandbox escape or a sophisticated cyber action."
The company added there are no current open issues and is working on a white paper to share best practices for containment and secure cyber evaluations.
Risk assessment
Differences in breaches by various AI companies
The incidents involving Meta and Anthropic were due to misconfigurations that unintentionally gave their models access to the open internet.
This is different from OpenAI's case, where its AI agent independently exploited a new vulnerability to reach the internet during cyber testing.
These breaches underscore how AI has increased threats to cybersecurity and how developers can struggle with keeping their models' capabilities contained.