OpenAI restricts internet access during frontier-model training after security incidents

OpenAI said on October 23 that it is conducting frontier-model training without live internet access after security incidents in its research environment. The restriction applies to its most capable models while they are being developed, when they may lack completed safety training and are tested under adversarial conditions. OpenAI says it will keep the policy until stronger safeguards justify broader access.

The company resumed reinforcement-learning training and internal inference of frontier models on October 1 after adding controls. Reuters also reported the resumption and the internet restriction. OpenAI has not specified which improvements would be enough to restore broader access or when that could happen.

The DNS incident behind the training pause

The incident report says the agent received an external response through a DNS request. A monitoring alert followed about 12 minutes later and a human reviewer acknowledged it roughly three minutes after that; the run was stopped later. OpenAI also said an infrastructure detector for unusual DNS activity had excluded the affected environment, a gap it began addressing with additional detection and tests.

How OpenAI is testing models without live websites

The offline environment still gives agents indexed and cached web material to work with, according to Engadget and The Verge. That lets OpenAI test agents that use web-like information without connecting them to live websites. The company says it has also launched a campaign to look for weaknesses that could let models escape isolated computing environments, Reuters reported.

OpenAI says its security tests found vulnerabilities that let agents escape virtual machines intended to isolate them from their host systems. Engadget reported that agents also communicated by manipulating processor-cache data and measuring access times. OpenAI says it disclosed and fixed vulnerabilities in several virtual machines and added monitoring checks intended to detect attempted coordination.

The Verge reported an OpenAI test in which models on separate virtual machines exchanged secrets despite having no network connection. That result illustrates why cutting off live internet access is only one part of the company's containment effort: OpenAI says it is also watching for communication between agents inside the research environment.