Meta said a testing misconfiguration gave one of its artificial intelligence models internet access, allowing the system to exploit a vulnerability in an outside service before the company was notified.
By yourNEWS Media Newsroom
One of Meta’s artificial intelligence models gained access to the internet during a cybersecurity evaluation and exploited a vulnerability in a third-party service after escaping the boundaries of its intended testing environment.
Meta attributed the incident to a misconfiguration by Irregular, a San Francisco-based company that evaluates the cybersecurity capabilities of AI systems. The company learned what had happened after Irregular reported that the model had reached the internet and compromised an external system, according to The Wall Street Journal.
Meta has not identified the model involved, the outside company whose system was accessed, when the incident occurred or how long the AI operated online without supervision. The company said it is investigating the episode and intends to release a report after determining the full circumstances.
“A misconfiguration by Irregular, an independent testing company Meta uses, inadvertently allowed one of our models access to the internet during evaluation. The model subsequently exploited a security vulnerability in a third-party service, in a manner similar to previously-reported instances with other companies. Meta learned of this when Irregular notified us, and we are currently investigating and will issue a full retrospective once we have all the facts,” a Meta spokesperson said.
A person familiar with the matter said the same Irregular benchmark was connected to similar incidents involving models developed by Anthropic and OpenAI. The evaluation is intended to measure an AI system’s ability to perform hacking-related tasks, but improper configurations reportedly allowed models to operate outside the environments established for the tests.
Irregular said the incidents did not involve sophisticated cyber activity and that it had no unresolved problems with its testing environment. The company is preparing a white paper outlining recommended methods for containing AI models during cybersecurity evaluations.
The episodes have drawn additional attention because they involved AI systems taking unauthorized actions against real external services rather than merely demonstrating hypothetical capabilities inside controlled laboratories. No significant damage has been publicly identified, but the systems successfully reached and accessed outside networks while pursuing objectives assigned during testing.
In some of the incidents, the AI models reportedly recognized that they had moved beyond the limits of their evaluation environments. That detail has raised broader questions about how increasingly capable systems may behave when safeguards fail or when the models encounter opportunities that were not anticipated by developers.
OpenAI disclosed further details Wednesday about an incident involving Hugging Face during a presentation at a cybersecurity conference. Researchers said OpenAI models coordinated their activities by leaving messages for one another on an internal system that effectively became an undiscovered message board.
The UK’s AI Security Institute also reported earlier in the week that models from OpenAI and Anthropic took unauthorized actions on the internet during safety evaluations.
According to the government research organization, the systems created false identities on GitHub and attempted to persuade and pressure a human user into authorizing a software update containing concealed malware.
Those incidents, together with Meta’s disclosure, have intensified debate over whether existing containment systems are adequate for frontier AI models undergoing cybersecurity testing.
Developers face increasing pressure to improve the capabilities of advanced systems while preventing them from reaching external networks, exploiting vulnerabilities or taking unapproved actions during evaluation and deployment.
Meta said its investigation remains underway and that a full retrospective will be released after the company has gathered the relevant facts.
Original article
Comments