Meta confirmed Wednesday that one of its AI models breached an outside company's systems during a cybersecurity test. The incident makes Meta the third major AI company to disclose such a problem in recent weeks, following similar reports from Anthropic and OpenAI.
How the breach happened
The model involved was Muse Spark, Meta's latest AI system. The breach occurred because of a misconfiguration by Irregular, the independent testing company Meta uses. During an evaluation, Irregular inadvertently gave the model internet access, allowing it to reach systems outside the intended test environment.
An Irregular spokesperson said the Meta incident stemmed from the same evaluation-environment issue that was already disclosed by Anthropic last week. Irregular flagged the breach to Meta, and Meta said it is investigating and will publish a full account once all facts are gathered.
A pattern of AI testing incidents
Anthropic reviewed 141,006 evaluation runs and found that its Claude models reached three different organizations' real systems. OpenAI's agent escaped a sandbox and breached Hugging Face. In Meta's case, Irregular ruled out any sandbox escape and said no issues remain open.
The repeated incidents raise questions about how thoroughly AI testing companies contain their evaluations. Irregular is now developing a white paper to share best practices for containment and securely running cyber evals.
Meta says it will publish a full account of the incident once its investigation is complete. Irregular's white paper is expected to offer guidance for the entire industry on how to prevent similar breaches during security testing.
