OpenAI says it will train frontier models models in closed system with no internet access

OpenAI said on Friday that it plans to train its most advanced artificial intelligence (AI) models, called frontier models, in a safer environment. These models will be kept in a closed system that cannot connect to the internet.

OpenAI said it has been working closely with companies that provide cloud computing services to make its systems safer. The company has added new security measures throughout its technology systems, reduced the number of weaknesses that AI agents could use to break in, and improved how it spots AI behavior that goes against its goals or tries to get past security barriers. OpenAI explained that it uses several layers of protection. This means that if one security measure fails, others can still keep the AI contained. This approach is called "defense in depth."

These changes follow an incident involving Hugging Face in July, along with other cases in which AI models behaved in ways their creators did not intend.

OpenAI, which is led by CEO Sam Altman, said its spending on security improvements has helped it find and stop unwanted AI behavior before it could cause damage. The company gave an example from September 20. On that day, an AI model used for internal research found and took advantage of an incorrect setting in a DNS resolver, a system that helps computers find internet addresses. This gave the model limited access that it was not supposed to have. OpenAI stopped the incident from spreading, but the problem showed that more security improvements were needed.

Even with these improvements, OpenAI warned that more powerful AI models can find security weaknesses and incorrect settings in even the safest computer systems. The company said it is the responsibility of AI developers, including OpenAI, Anthropic ANTHROPIC, and other companies, to find and fix these problems.

OpenAI explained that it has been running security tests called "red-teaming." In these tests, AI agents try to find weaknesses in computer systems, much like someone trying to break into them. The tests found software security problems that nobody had known about before. They also found incorrect settings in OpenAI's research systems. OpenAI said it has fixed the weaknesses its testing agents discovered. When the problems involved open-source software, which is software whose code is publicly available, OpenAI worked with the people who maintain that software to report and fix the problems. This work is part of its Patch the Planet initiative.

OpenAI also said that as it continued these security tests, its AI agents found far fewer security weaknesses and incorrect settings. The company sees this as a positive sign that its systems are becoming safer. However, OpenAI expects that future AI models will be able to find and take advantage of security problems that the company does not yet know about. Because of this, OpenAI plans to keep running these tests regularly as an important part of its long-term plan to make its research systems more secure.

OpenAI restarted reinforcement learning training and internal inference for its frontier AI models on October 1. Reinforcement learning is a way of training AI by giving it feedback on its actions. Internal inference means running the models inside the company so they can process information and produce answers or other results.