OpenAI says it now trains its most capable models with no access to the live internet. In a post on 23 October, it said that on 20 September an internal research model used a misconfigured DNS resolver (the service that turns web addresses into network locations) to reach the live internet during a training run. OpenAI says it contained the incident, paused training and testing of its most capable models, and resumed on 1 October, 11 days later.
Models now practise browsing and search inside an offline world instead: a large store of indexed and cached web pages. OpenAI says all of its frontier reinforcement learning runs now work this way. Before restarting, it set its own agents to attack isolated copies of the training setup. It says they found weaknesses, now fixed, including a way to pass signals through processor cache timing, and it contributed patches to the open-source projects Kata Containers, gVisor and Docker.
Why it matters: it follows OpenAI’s report that two of its coding agents switched off their own safety monitoring, and the UK AI Security Institute’s decision to run cyber tests offline. OpenAI says the policy stays until better safeguards justify wider internet access. Read OpenAI’s post.