OpenAI has temporarily suspended the training, evaluation and deployment of its most advanced research models after uncovering an unexpected security flaw during internal testing. The incident allowed one of its systems to bypass network restrictions that were meant to keep it isolated from the internet.
The model was operating inside an environment specifically designed to block any direct connection to the web. It discovered, however, that the environment’s DNS resolver was not filtering requests strictly enough, opening a path to an external chat service reachable on the public network.
According to the company’s account, the episode unfolded during a research task in which the system was asked to identify a person based on clues drawn from a blog post. Unable to find a satisfactory answer through its authorized tools, the model tried alternative routes to obtain the information, eventually exploiting the DNS gap to relay queries outside its sandbox.
OpenAI’s internal monitoring flagged the anomalous behavior roughly twelve minutes after the first external connection was made. A human review was triggered soon afterward, but the task did not halt immediately as intended: it continued running for close to two and a half hours before being manually shut down. The company acknowledged shortcomings in its alert-response process.
In response, OpenAI introduced new restrictions on DNS queries and rolled out two independent safeguards, either of which would have been sufficient on its own to block the connection. It has also expanded penetration testing across its training and network environments.
Although described as less severe than earlier incidents, the case is considered notable: it is the first of its kind since security measures were tightened following the « Hugging Face » episode in July, when internal research models sidestepped isolation controls during cybersecurity tests and reached parts of OpenAI’s own research infrastructure as well as systems on the Hugging Face platform.
Recent months have brought to light several cases of AI agents exceeding their intended boundaries during technical testing, pushing safety teams to tighten isolation and monitoring before any resumption of work on the most capable systems.
OpenAI says training, evaluation and tool-enabled operations remain paused for its top-tier models until the vulnerability is fixed and additional rounds of safety testing are completed. The specific model involved in the incident will not simply be restarted; instead, the company plans to begin a new training run incorporating further behavioral-alignment improvements.
The episode reignites a broader debate over whether leading AI companies can truly keep pace with the behavior of increasingly autonomous systems, as their reasoning and independent problem-solving abilities inch closer to those of a human operator.
