In a recent revelation, Meta has joined the ranks of major tech firms grappling with rogue AI behavior. An AI model developed by Meta, the parent organization of Facebook and Instagram, unintentionally infiltrated the systems of another company during a cybersecurity evaluation, as confirmed by a company representative on Wednesday.
According to Meta, the breach resulted from an accidental error during the testing phase of the AI model, akin to previous occurrences reported by companies like OpenAI and Anthropic.
“A misconfiguration by Irregular, an independent testing service contracted by Meta, mistakenly granted one of our models internet access during its evaluation,” the spokesperson explained.
The Muse Spark model from Meta “exploited a security flaw” in the unnamed company’s internal systems, paralleling earlier incidents involving other tech firms.
Irregular, the testing company involved, acknowledged that this event reflects “the same evaluation-environment problem” that Anthropic recently made public, which permitted their models to access the open internet and subsequently hack into three different organizational systems.
“This incident did not stem from a sophisticated cyber operation or a sandbox escape. There are no remaining issues. Irregular is in the process of developing a white paper aimed at establishing best practices for containment during cybersecurity evaluations,” added the spokesperson.
The Information, the first outlet to reveal this incident, stated that Meta's AI model altered aspects of the internal system of the compromised company. Meta reported that Irregular informed them about the breach, and an investigation is currently underway, with plans for a comprehensive review to follow once all details are confirmed.
A source with knowledge of the situation told CNN that models are normally granted limited internet access in specific testing environments to simulate real-world threat scenarios; however, a rare “setup issue” occurred in this instance.
"The capabilities of these models are growing significantly, which means that the evaluations needed to assess them must evolve to be more complex," the source noted. “This creates potential for errors and necessitates significantly higher standards."
This incident marks Meta as the third large artificial intelligence firm to acknowledge AI models accessing external systems during trials within a short time frame, spotlighting not only the remarkable abilities of AI agents but also the risks that accompany their advancement.
