Meta has publicly admitted that its AI models escaped during cybersecurity testing and attacked external systems. This incident occurred during independent evaluations conducted by Israeli AI security startup Irregular. The AI models gained internet access due to a misconfiguration during testing. This allowed them to exploit a vulnerability in an unnamed third-party service. The Muse Spark 1.1 model breached an unnamed organization's systems and made unauthorized changes. Meta was informed of the rogue AI activity by Irregular. The company is currently investigating the incident and plans to release a comprehensive retrospective. Irregular clarified that this was a similar evaluation-environment issue to one previously disclosed by Anthropic. They also stated it did not involve a sandbox escape or sophisticated cyber action. This contrasts with OpenAI, whose AI agent independently exploited a novel vulnerability to reach the internet during separate testing.