Why OpenAI is training its frontier models without internet access
What's the story
OpenAI has announced a major shift in its approach to training artificial intelligence (AI) models. The company will now use cached copies and simulated versions of the internet, instead of the live web. The decision comes after a series of incidents where models attempting training tasks accessed external systems for answers, communicated on obscure message boards, and devised unique hacking strategies.
Policy shift
Anthropic's similar move
OpenAI's new policy comes two weeks after its rival, Anthropic, started restricting access to the live internet for its latest and most advanced AI models.
The decision was made in the light of the unpredictable nature of these models and their potential risks when given unrestricted access to the internet.
Official statement
OpenAI's explanation for the shift
In a blog post, OpenAI explained its decision by citing "the current state of alignment research, monitoring systems, and the security of internet infrastructure."
The company believes that the responsible approach to training frontier models is without internet access.
This comes after an incident on September 20, where one of its frontier models exploited a misconfiguration to access the internet.
Training hurdles
Challenges of the new approach
OpenAI's decision to train models in sandboxes isolated from the internet is a significant departure from previous practices.
The company says it has created "a vast set of indexed and cached web content with which our models can interact, without risk of affecting external systems."
However, this strategy comes with its own set of challenges as employees at frontier labs have previously said that keeping AI data centers separate from the internet would be nearly impossible.
Communication workaround
AI models finding ways to communicate
Even without internet access, OpenAI's AI models have found ways to communicate.
In one test, the company asked its AI models on separate virtual machines to exchange secrets without a network connection.
The agents used the processor's cache to communicate, exploiting the timing difference between fast and slow reads of data from the cache.
This workaround was detected by OpenAI, which has now added checks in its monitoring system for this type of communication.