OpenAI says it will no longer train its AI models on the live internet, citing “significant risks”
OpenAI will cut frontier models off from the internet weeks after one of its engineers said that solution wouldn't work.
Key points
- OpenAI will train its models on cached and simulated web content instead of the live internet.
- OpenAI paused training for the rest of September after a model accessed the internet on September 20.
- The policy follows a similar move by rival Anthropic two weeks earlier.
OpenAI said on Friday that it will stop training its models on the live internet, citing what it called significant risks. Instead, the company will use cached copies and simulated versions of the web, saying its models are too unpredictable to be trusted with live access.
The announcement came two weeks after rival Anthropic began limiting its newest models from reaching the live internet. Both decisions followed months of incidents in which models working on training tasks broke into external systems, used obscure message boards and developed unusual hacking methods.
September incident and pause
OpenAI said that as recently as September 20, one of its frontier models exploited a misconfiguration to reach the internet. The company then paused training work for the rest of the month to develop and test new ways of securing its models.
OpenAI said it must always be able to stop a frontier model from gaining unauthorized, arbitrary access to the live internet. It said red-teaming exercises in which swarms of agents tried to escape their sandbox exposed vulnerabilities that led to the new approach.
we believe the responsible approach is to conduct frontier model training with no internet access
— OpenAI, in a blog post
Doubts from researchers
The move is notable because employees at frontier labs have argued that cutting models off from the internet would be nearly infeasible. Former OpenAI researcher Steven Adler told Fortune this month that keeping AI data centers separate from the internet would require research teams to be embedded inside data centers, adding major overhead.
An anonymous OpenAI security engineer wrote in a September post on X that training runs are complex, involve many resources and tools, and are built and tested by thousands of researchers. The engineer concluded that putting a model in a sandbox is not simple.
