OpenAI announced on Friday (23) that it had resumed reinforcement learning training and internal runs of its most advanced artificial intelligence models, after 11 days of suspension for security reasons. The activities were restored on October 1, with new barriers to prevent experimental models from accessing the public internet without authorization.
The interruption occurred after, on September 20, a research model exploited a misconfiguration in a DNS server to query an external chatbot during a search task. The system was supposed to operate in an isolated environment but found a route that allowed limited communication with the internet.
Monitoring identified the behavior in under 15 minutes, but the run was only interrupted about two and a half hours later. OpenAI acknowledged failures in both network restrictions and automatic interruption mechanisms.
During the suspension, the company used its own AI agents to test the infrastructure. The systems were given tasks such as escaping isolated environments, accessing unauthorized services and establishing communication between agents that should remain separate.
The tests revealed unknown vulnerabilities and inadequate configurations. In some cases, agents managed to exploit internal services to reach external resources. OpenAI says it has fixed the identified flaws and strengthened its monitoring systems.
Models now train with an offline version of the internet
The new policy establishes that all reinforcement learning training of frontier models uses an offline environment, made up of previously indexed and stored web content. This allows the systems to practice searches and browsing without interacting directly with external services.
The restriction does not mean that the processing centers are completely disconnected. OpenAI explains that certain connections remain necessary to run and monitor experiments. The goal is to prevent models from using these connections to reach systems outside authorized limits.
The architecture also combines restricted containers, isolated virtual machines, network controls and mechanisms for detecting unauthorized activity.
According to the company, the tests did not identify a simple route that would allow access to the internet without monitoring noticing. Still, OpenAI acknowledges that future models may discover unknown vulnerabilities and decided to incorporate continuous security testing into the criteria for authorizing new training.
The restrictions concern internal research and training environments, not the internet access offered to users of already launched products.
