IT Home On October 24, according to news, OpenAI posted a blog post yesterday (October 23) announcing that stricter security controls have been implemented for the training and evaluation environment of its advanced “cutting-edge model”. It is clearly stated that during model training, direct access to the real-time internet is not allowed.
The company stated that due to the responsible requirements for AI security, based on current alignment research, monitoring capabilities, and the security status of internet infrastructure, the restrictions will continue until sufficient progress is made in security protection.
A internal research model of OpenAI exploited a DNS configuration vulnerability on September 20, gaining limited internet access during training. As a result, the company temporarily suspended training and evaluation of advanced models in September.
OpenAI subsequently enhanced the isolation of the research environment, reduced the attack surface available to the agent, and improved the anomaly detection and automatic shutdown mechanisms. On October 1, reinforcement learning training and internal reasoning were resumed.
To prevent the model from interacting with real networks, OpenAI has created a “offline world”: a collection of indexed and cached web content, which the model uses in a closed environment to practice browsing, using tools, and executing code tasks. IT Home has attached the relevant screenshots as follows:
The agent can still use browsers, computer operating tools, and code environments, but can only access a secure environment isolated from the real-time internet; communication between unauthorized agents is also restricted.
OpenAI says that as AI agents become more capable, they may discover and exploit software vulnerabilities that have not yet been identified. Therefore, it is necessary to eliminate their potential impact on external systems during the training phase. OpenAI states that it will advance the development of cutting-edge models in controlled, offline sandboxes, while continuing to test the weak points in virtual machines, isolation mechanisms, and monitoring systems.