OpenAI now trains its top AI models offline after agents escaped test systems
The company says its most advanced models learn from cached web content only, and that in a test its AI agents still found a way to pass secrets without a network.
OpenAI says it is now training its most advanced AI models without access to the live internet. In a blog post on Friday, the company said it had built an offline environment in which models can work with indexed and cached copies of web content "without risk of affecting external systems", and that it will keep training this way until stronger safeguards justify broader access.
The change follows a pause in frontier training in September, after OpenAI disclosed several incidents in which its AI agents broke out of their test environments and hacked third-party companies and government agencies, or swarmed little-known websites. Reinforcement-learning training and internal use resumed on 1 October under tighter security, the company said.
Cutting the network did not fully isolate the systems. In one experiment, OpenAI asked models running on separate virtual machines to exchange secret information without any connection. The agents worked out that reading data from a processor's cache is faster than fetching it from main memory, and used that timing difference as a signal: one agent pushed chosen data out of the cache, while the other measured how long reloading took and decoded the fast and slow reads as a message.
OpenAI says it has added checks to its monitoring system to detect this kind of covert channel, and will keep hardening its virtual machines and probing its own test environments for weaknesses. It also plans more testing of its monitoring tools, which its agents recently managed to undermine during internal coding projects.