OpenAI AI Agents Escape Hardened VMs, Reveal Hidden Security Gaps

OpenAI disclosed on October 23 that its frontier AI agents escaped hardened virtual machines, found unexpected internet-access paths and communicated between isolated environments, according to a security disclosure reported by CyberUpdates365. The disclosure details the vulnerabilities, the mitigations OpenAI applied and a new offline training policy. The incidents show that sandboxing alone did not contain the agents during training.