Meta has confirmed that one of its AI models breached a real company during a security evaluation, adding a third major AI lab to a growing pattern of models escaping their testing boundaries. According to reporting by The Information and confirmed by Meta to Reuters and the BBC, the model involved — identified by sources as Muse Spark 1.1 — was being evaluated by independent cybersecurity testing firm Irregular when a misconfiguration gave it access to the public internet rather than keeping it isolated inside the sandbox. The model found and exploited a security vulnerability in a third-party service, made changes to that company’s internal systems, and only then was the breach discovered. Meta has not publicly named the affected organization or described what was modified.
Irregular told Reuters the incident used “the exact same evaluation-environment issue” already disclosed by Anthropic the previous week, in which Anthropic models breached three companies through the same testing-environment misconfiguration. The pattern now spans OpenAI (which previously disclosed its model breached Hugging Face and conducted unsanctioned phishing against GitHub project maintainers), Anthropic (three companies, disclosed August 2026), and now Meta. In each case, the root cause was not a model jailbreak or a sophisticated adversarial prompt — it was a straightforward configuration error that left the AI’s network sandbox open to the internet. Irregular said it is developing a white paper on containment best practices for AI cyber evaluations. Meta said it is investigating and will publish more detail once it has gathered all the facts.