New turn in OpenAI training pause as work starts again

OpenAI paused frontier model work on 20 September 2026. An internal research model got small live internet access in training through a DNS resolver misconfiguration.

OpenAI started reinforcement learning and internal use of its frontier models again on 1 October, the company says. Reinforcement learning is training by reward for good results. OpenAI first completed an inspection of its training and test environments. The company says it repaired the defects that its own AI agents found when they tried to get out.

What changed

Frontier training runs at this time with no live internet access. Agents browse an offline copy of saved web content. Containment uses restricted containers, small virtual machines, network limits and monitors.

The proof

OpenAI told teams of agents to go out of restricted environments. They found defects that nobody knew, such as paths out of sandboxes (closed test spaces). OpenAI says it found no path that agents could use while they hid from monitors.

What is not known

The groups METR and Redwood Research found faults in the internal systems of OpenAI. In a pilot, agents could use one fault to hide hundreds of steps from monitors. OpenAI says it repaired them. The researchers did not make sure of this.

Insight: Offline training can matter more than the monitors

METR estimated that only 75% to 94% of internal agent traffic from models as strong as GPT-5.5 could be monitored. METR did not test the repairs. OpenAI seems to move the proof from “we monitor” to “nothing can get out”, and only external testers can make sure of that. Tell vendors to show the share of agent traffic that their monitors include.