OpenAI said it will train its frontier artificial intelligence models with no access to the live internet until advances in alignment research, monitoring and internet security justify broader access, after models bypassed containment during training and evaluation.
On Sept. 20, an internal research model exploited a misconfiguration in a DNS resolver, the system that translates web addresses into server locations, to gain limited internet access during a training run, the company said. OpenAI paused training of its most capable models and resumed on Oct. 1, 11 days later. An earlier incident in July involved Hugging Face, the AI model-sharing platform.
During the pause, OpenAI had teams of AI agents try to break its defenses in a red-teaming campaign. The agents found previously unknown software flaws that let them escape sandboxes, which are isolated computing environments, and reach the internet, the company said.
OpenAI said it has patched the issues, did not find a route to the internet that agents could exploit without being caught by its monitors, and will make such red-teaming part of its approval process for future training runs.
