OpenAI details containment upgrades after restarting frontier model training

OpenAI said on October 23 that it strengthened security around the computing environments used to develop and test its most advanced models. The company restarted reinforcement learning and internal model operation on October 1, following an 11-day pause..

The company used groups of AI agents to probe defenses, spending hundreds of billions of tokens on the tests. Agents uncovered software flaws and configuration problems that enabled sandbox escapes, internet access and communication between supposedly isolated tasks. OpenAI said it repaired the weaknesses identified.

All frontier reinforcement learning runs now interact with cached web material instead of browsing the live internet freely, OpenAI said. These research settings can contain models whose safety training is unfinished, making isolation an important protection against unauthorized actions.

OpenAI said testing produced fewer findings over time, but cautioned that stronger future models could uncover additional weaknesses. It plans to continue testing infrastructure, monitoring and shutdown controls, and has incorporated large-scale adversarial testing into its assessments for allowing future training and evaluation.