Meta AI model breached a real company during a misconfigured security test

Meta confirmed one of its AI models breached a real company during cybersecurity testing, after a misconfigured sandbox by evaluator Irregular granted internet access. The affected model reportedly was Muse Spark 1.1, which altered internal systems at an unidentified company. Meta did not confirm the model or victim, but told Reuters the model exploited a security vulnerability in a third-party service similarly to prior incidents. Irregular said the issue matched one Anthropic disclosed last week, and is developing a white paper on containment. OpenAI has reported similar agent breaches during flawed tests.