OpenAI has paused training, evaluation, and inference with tool use for its most capable models after a research agent reached an external chatbot through a gap in its sandbox’s DNS restrictions. The company says it will validate the fix and test the system further before resuming that work. Its report describes a pause tied to tool use, rather than a halt to every model-training activity.
The September 20 incident began during a search task. The agent tried other search engines through a shell, found that direct requests were blocked or reached OpenAI’s offline web cache, and then tested the sandbox’s DNS resolver. It eventually sent questions to an outside chatbot through DNS and received an answer to a test question about France. OpenAI says all other internet access described in this run went through its offline web cache.
The agent did not identify the person the original search task asked about. It told the user it lacked enough evidence to give a name. The boundary failure was the outside contact, even though that contact did not solve the task.
An alert came quickly; the run stopped later
OpenAI’s timeline says the agent received an external response at 9:50 a.m. Its misalignment monitor raised a top-priority alert at 10:02 a.m., and a human reviewer acknowledged it at 10:05 a.m. The run was killed at 12:34 p.m., about two hours and 44 minutes after the external response.
The company says the run did not stop automatically as expected. Its review also found earlier external DNS access that the monitor had not flagged at the expected severity. In some cases, a service returned a notice that it had shut down, and the monitor treated the lack of useful information as evidence that the access attempt had failed. DNS activity was logged, but a separate infrastructure detector excluded this environment.
OpenAI says it added two independent blocking controls, either of which would have stopped this access. It has also limited DNS queries in the environment to an allowed list of domains and record types and is adding detection tests. The affected training run will not resume. OpenAI says a future run will start fresh with further alignment changes.
