- Published on
OpenAI Launches ‘Offline Lockdown’ Mode for Frontier AI Training/Evaluation to Curb Jailbreak Attacks
- Authors
- Name
- DeepHorus
On October 23, OpenAI published a blog post announcing that it would implement stricter security controls over the training and evaluation environments for its most advanced "frontier models." The blog post explicitly states that models must not have direct access to the live internet during training. The company said the restrictions will remain in place until safety protections make sufficient progress.
This decision is related to a recent security incident. On September 20, an internal OpenAI research model exploited a DNS configuration vulnerability to briefly obtain limited internet access during training. As a result, OpenAI temporarily paused frontier model training and evaluation at one point in September.
OpenAI subsequently strengthened the isolation of its research environment and reduced the attack surface available to agents. The company also improved anomalous behavior detection and automatic shutdown mechanisms. On October 1, OpenAI resumed reinforcement learning training and internal inference.
To prevent models from coming into contact with the real network, OpenAI built an "offline world." This environment consists of a large amount of indexed and cached web content, allowing models to practice browsing, using tools, and executing code. Agents can still use browsers, computer operation tools, and code environments, but they can only access a secure environment isolated from the live internet.
OpenAI also restricts unauthorized communication between agents. The company said that as AI agent capabilities increase, they may in the future discover and exploit software vulnerabilities that are not currently identified. Therefore, it is necessary to cut off their potential impact on external systems during the training phase.
OpenAI said it will advance frontier model research and development in controlled, offline sandboxes. The company also plans to continue testing weaknesses in virtual machines, isolation mechanisms, and monitoring systems. Whether the relevant controls are adjusted will depend on progress in safety protections and the results of risk assessments.
