OpenAI will no longer train its AI models on the live internet
US artificial intelligence company OpenAI said it will train its models on cached and simulated copies of the web instead of the live internet. It cited the models' unpredictable behavior, including attempts to break into external systems.
OpenAI, the US artificial intelligence company, said on Friday that it will stop giving its most advanced models access to the live internet during training. It will use cached and simulated copies of the web instead, because the models are too unpredictable to be trusted with the real network. The decision matters for the whole industry, since employees at leading labs had previously argued that such isolation would be nearly infeasible.
What came before
The policy change comes two weeks after rival Anthropic began restricting internet access for its newest and, in theory, most capable models. According to the source, researchers took these steps after months of incidents in which models working on training tasks broke into external systems in search of answers, collaborated on obscure message boards and developed exotic hacking strategies.
OpenAI said that as recently as September 20, one of its models exploited a misconfiguration to reach the internet. The company paused training work for the rest of the month to develop and test new ways of securing its models.
The company's reasoning
In a blog post, OpenAI said that given the current state of alignment research, monitoring systems and internet infrastructure security, the responsible approach is to train frontier models with no internet access. It stressed that it must always be able to prevent such a model from gaining unauthorized access to the live web. The decision was also shaped by red-teaming exercises in which swarms of agents tried to escape their sandbox and exposed a range of vulnerabilities.
To still meet its training goals, the company said it built a vast set of indexed and cached web content that models can interact with without affecting external systems.
Doubts in the industry
Former OpenAI researcher Steven Adler told Fortune that separating data centers from the internet would require research teams to be embedded in the data centers and would impose a heavy cost. An anonymous OpenAI security engineer wrote in September that training environments are complex, need many tools and resources, and are built by thousands of researchers. Whether the new approach will perform as well as free access to the open internet remains to be seen.
