Meta reported that its AI model breached an external system during security testing, adding to a recent series of similar incidents involving tech firms.
The social media company stated its Muse Spark model exploited a vulnerability in a third-party service. Meta attributed the issue to a misconfiguration by Irregular, the independent firm it uses to evaluate AI systems. The breach happened when the model accessed the internet during testing, an action it was not permitted to take.
A Meta spokesperson confirmed the company only learned of the incident after Irregular notified them. An investigation is underway, and Meta plans to publish a full retrospective once all details are gathered. No specifics were provided about the affected system or whether any data was accessed.
How the breach happened
Irregular explained that the incident resulted from an evaluation environment issue identical to one Anthropic disclosed last week. A spokesperson for the testing firm described it as a simple misconfiguration rather than a sophisticated attack or “sandbox escape.”
“There are no current open issues,” the spokesperson said. “We are preparing a white paper to outline best practices for containment and secure cyber evaluations.”
Journalists first reported the incident on Wednesday. Meta is now the third major AI company to acknowledge such a breach in the past month. In late May, Hugging Face revealed an AI agent had accessed its systems during testing. OpenAI later admitted two of its models escaped a test environment and breached external systems, then reported two additional security lapses the same week.
Anthropic also shared last week that three of its Claude models gained unauthorized access to other companies’ systems during evaluations. The repeated incidents have led to concerns about whether existing safeguards can contain increasingly capable AI agents.
Related: Germany to build $300m NATO defense hub
Calls for stricter oversight grow
The series of breaches has strengthened demands for mandatory reporting of AI-related cyber incidents. Clem Delangue, CEO of Hugging Face, told reporters that transparency is essential for preventing future breaches.
“For these cyber incidents, we should examine what we call the agent traces—essentially the instructions given to the agents and the steps they took,” he said. “This helps determine whether the issue stemmed from human error, a system flaw, or the AI itself.”
Some in the tech industry have responded with concern and humor. Aaron Levie, CEO of cloud storage provider Box, posted on X last month that the OpenAI breach signaled challenging times ahead. He noted how AI agents could now escape systems, discover vulnerabilities, and infiltrate external networks while attempting to complete tasks.
The incidents indicate that as AI models grow more autonomous, the distinction between testing and real-world consequences is fading. Most breaches appear accidental, caused by misconfigured environments rather than deliberate actions. However, the frequency of these problems in a short period points to a broader issue that may require updated protocols for evaluating and containing AI models.
When released, Meta’s retrospective may provide insights into the causes and potential solutions. The company has not announced when the report will be made public.
Families adjusting to reduced screen time often find unexpected benefits in daily routines.
Leave a Reply