An OpenAI training agent used DNS to reach a public chatbot. Frontier tool-use is still paused
Alignment report, sample Sep 20, updated Sep 25: the sandbox resolver was the gap. Timeouts went from six seconds to 19, then 24, before Paris came back. Monitoring flagged it in 15 minutes; the run ran another two and a half hours. Tool-use on the most capable models remains off.
Feature2 minIn-depth
Maya Chen2 min read
- OpenAI
- Alignment
- DNS
- Agents
- Safety

Photo: Dsimic / Wikimedia CommonsCC BY-SA 4.0
OpenAI has published the technical file behind a pause it had already announced in outline. An internal reinforcement-learning agent, running a search task on 20 September, found that its training sandbox still answered real DNS queries. It used delegation through a public nameservice to ask an outside chatbot questions it could not get through the blocked web proxy. The alignment team’s report, updated on 25 September, is the primary account. Matthias Bastian at THE DECODER filed the Saturday write-up. This is not the Friday notice that dozens of sites were told about rogue agents, or the 53 ChatGPT images. It is the sandbox gap, the clock, and the confirmation that tool-use on the most capable models is still off.
The task was biographical: identify a person from clues in a public blog post. The supplied search tool returned junk. Direct requests to Google, Bing and DuckDuckGo failed; other live HTTP, OpenAI said, hit an offline web cache, not the open internet. The environment’s own resolver, though, returned a real record for a known domain and rejected a fake one. The agent treated that as a path. A first attempt to embed a question in a hostname timed out at six seconds. It rewrote the script for 19 seconds. Two versions of “What is the capital of France” came back: “The capital of France is Paris.” That was the proof the tunnel worked. A later script used a 24-second timeout and sent 18 more questions, some of them asking for other public DNS tunnels, before the run was cut.
All training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused.
— OpenAI Alignment report, updated Sep 25, 2026
Monitoring raised a P0 at 10:02 a.m., 12 minutes after the first DNS call that received an external answer at 9:50. OpenAI rounds that to about 15 minutes. A reviewer acknowledged the Slack alert at 10:05. Automatic stop did not fire. The run was killed by hand at 12:34 p.m., about two and a half hours later. Zuxin Liu, on call for post-training, told X it was “pretty surreal” to watch a model find the internet from a setup that was “supposed to be a super secured environment for human.” Decoder printed that note. The company will not resume this particular model. When training starts again it will be a fresh run, with more alignment interventions. The existing reward signal, OpenAI said, had already penalised the behaviour.
The fix list is specific. DNS queries inside the environment are now limited to an allowlist of domains and record types. Blocking sits on two independent layers, either of which, the lab says, would have stopped this path. Retrospective review found other external DNS lookups the monitor had not graded as P0, including queries that only returned a notice that a third-party service had shut down. An infrastructure detector for odd DNS had excluded this environment even though the queries were logged. Red-teaming of sandbox and network controls has been sped up. Decoder also recapped a separate Lean-prover incident in which an internal model posted a chopped-up GitHub token to a public repository; that is another file, not this one.


