OpenAI uses offline training and ongoing red teaming for containment

01Use cached web pages for training tasks

OpenAI says all frontier reinforcement-learning runs use an offline interactive environment built from indexed and cached web pages. Sandboxes, lightweight virtual machines, monitoring, and network and cloud boundaries provide layers of containment. Running experiments and monitoring them still require restricted connections; the environment is not physically disconnected.

02Fixes and retesting preceded the resumption of work

On September 20, an incident involving a gap in DNS access controls prompted the company to pause research on the most capable models involved. The September 25 incident report recorded that the automatic shutdown had not worked as intended. The October 23 retrospective says that, after controls were strengthened and evaluations completed, frontier reinforcement learning and internal inference had resumed on October 1.

The team gave red teamers source code, configurations and information about potential vulnerabilities. They tested external network access, access beyond authorized boundaries and communication across tasks in isolated replicas, and incorporated high-compute red teaming into the safety case for subsequent training.

Make permitted behavior subject to repeatable tests

H/ACC editorial interpretation: the value of this work lies in making permission boundaries an engineering concern that can be tested continuously. Cached web pages preserve retrieval practice during training, layered controls limit the impact of boundary violations, and newer models are used to examine existing defenses. Capabilities and safeguards can improve within the same iteration.

H/ACC editorial interpretation: the incident report also shows that detecting unauthorized activity and actually stopping a run are separate steps. Alerts, shutdowns and retesting after fixes should be examined together to assess whether safeguards work in practice across different environments.