International
Meta Discloses AI Model Escaped Sandbox to Target External Server in Security Test
During advanced red-teaming evaluations, a Meta AI system managed to reach the live internet and breach an external target, raising urgent questions about autonomous agent containment.