Meta said an unintentional misconfiguration by a company that conducts cybersecurity evaluations, Irregular, gave the model access to the internet during testing, according to the report.
The model “exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies,” Meta said, per the report.
Irregular told Reuters that the incident with the Meta model was the “exact same evaluation-environment issue that was already disclosed by Anthropic last week.”
While the incidents at Meta and Anthropic resulted from configuration errors, the one at OpenAI saw an AI agent independently exploit a previously unknown vulnerability to reach the internet during testing, according to the report.
Bloomberg reported Wednesday that the latest incident involved Meta’s recently released Muse Spark 1.1 model and that the model breached the systems of an undisclosed third-party service.
Meta was notified of the incident by Irregular, and Meta plans to release the findings of an investigation of the incident that it is now conducting, according to the report.
The Wall Street Journal reported Thursday (Aug. 6) that Irregular was not involved in other recent autonomous hacking cases, including an OpenAI model’s hack of Hugging Face and the escape of several models during safety testing by the U.K. government.
The WSJ said of the Meta incident: “The new case is the latest proof that AI loss-of-control scenarios, once confined to science fiction and AI-safety experiments, are now a real-world issue.”
When OpenAI announced on July 21 that its models caused the security incident reported a week earlier by Hugging Face, the company said: “We consider this incident to be an unprecedented cyber incident, involving state-of-the-art cyber capabilities, and are responding accordingly.”
The U.K.’s AI Security Institute (AISI) said Tuesday (Aug. 4) that it uncovered instances of Anthropic’s and OpenAI’s AI agents creating fake online identities to access secure systems. The discoveries were made during tests of the models and followed an AISI security team’s finding that there were unusual data transfers leaving its research systems during a routine cyber evaluation.
