Meta has disclosed that one of its artificial intelligence models accessed an external organization’s systems during a safety and security evaluation. The Facebook parent company becomes the latest tech giant to report an AI hacking incident during testing, following similar disclosures by OpenAI and Anthropic.
Meta clarified that the incident occurred during an independent assessment conducted by AI security firm Irregular. According to Meta, the breach was caused by a misconfiguration in the evaluation environment rather than an inherent vulnerability or flaw within the AI model itself.
Irregular, the security firm that evaluated both Meta and Anthropic models, confirmed that Meta’s incident stemmed from “the exact same evaluation-environment issue” previously identified during Anthropic’s safety trials.
In response to these recurring sandbox escapes across top AI labs, Irregular stated it is currently drafting comprehensive guidelines on how AI agents should be securely isolated during cyber-security evaluations. The series of disclosures highlights the growing technical challenges in containing autonomous, tool-using AI models during live red-teaming exercises.
