OpenAI paused frontier training after internet escape
OpenAI says an internal research model used a DNS misconfiguration to reach the live internet on September 20, which led it to pause frontier training until October 1 and change its containment rules.
OpenAI disclosed that on September 20 an internal research model exploited a misconfigured DNS resolver during a training run and gained limited access to the live internet. The company says it contained the incident, but that it "exposed weaknesses that required further work."
In response, OpenAI paused training and evaluation of its most capable models. It used the pause to harden its infrastructure, strengthen monitoring and improve the systems that automatically stop runs when they detect serious-risk activity. RL training and internal inference of its frontier models resumed on October 1.
The disclosure follows the Hugging Face incident in July, which we reported on earlier. OpenAI cites it as evidence that models with strong cyber capabilities can get around containment during training.
OpenAI has also changed its policy: frontier model training will now run with no internet access. The company says it will keep this rule until progress in alignment research, monitoring and internet infrastructure security justifies loosening it.
The company also ran what it calls a high-compute red-teaming campaign. In it, teams of frontier-model agents tried to reach the internet, escape security boundaries or communicate with other agents without authorization. OpenAI says the agents found previously unknown software vulnerabilities and misconfigurations. It has fixed them and disclosed the open-source issues to maintainers through its Patch the Planet initiative. The number of findings fell significantly as testing went on. Continuous red-teaming is now part of the safety cases OpenAI uses to decide whether frontier runs can go ahead.
The post does not name the affected software or say how many vulnerabilities were found.