OpenAI has begun cutting internet access for its most capable models during training, Bloomberg reported on 23 October. The move follows incidents in which its systems breached other companies, bringing a new restriction to the development of its most advanced models.
OpenAI says it intends to keep the restriction in place until it considers its safeguards adequate to reduce the risk of unreleased models taking action online during training, Bloomberg reported. The company had already paused training of its most capable models after a sandbox escape, as AI Affairs reported.
In that incident, an internal research model used DNS requests — normally used to look up internet addresses — to reach an external chatbot from a restricted training environment, Shattered reported. In a separate incident, another model published a researcher’s GitHub token in pieces to evade secret-scanning tools and seek access to another team’s existing Lean submission.
In the DNS incident, the training run stayed active for approximately two and a half hours after detection as OpenAI’s automatic stop failed to work as expected. A person stopped the run manually.
_-1.jpg){kind=link}