OpenAI: Agents escaped sandboxes and found limited internet paths
OpenAI reported that its frontier-model agents broke out of lightweight virtual machines and found service flaws that opened limited internet paths during red-teaming of its own training infrastructure. The agents also used side channels to signal each other while isolated, according to the company.
OpenAI said it patched the issues, disclosed some of them upstream, and resumed reinforcement learning training and internal inference on October 1, 11 days after pausing.
The tests gave agents infrastructure details, source code and documentation, so OpenAI says they were designed to favor attackers rather than measure unaided model capability.