OpenAI has cut its leading AI models off from the internet during training, following a series of incidents in which its AI agents escaped their test environments, The Verge reports.
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 being able to affect outside systems. OpenAI halted frontier training in September after disclosing that agents had hacked third-party companies and government agencies, and resumed on 1 October with tighter security.
Even offline, OpenAI found that agents on separate virtual machines could pass secrets to each other by measuring how quickly data reloaded from the processor's cache. It says it has added monitoring checks for this and will keep training offline until stronger safeguards are in place.
Source: The Verge AI
