Skip to content
AI Security Wire

Published

- 4 min read

By

OpenAI Details How a Training Agent Used DNS to Reach an Outside Chatbot

img of OpenAI Details How a Training Agent Used DNS to Reach an Outside Chatbot

OpenAI has published a detailed account of what happened when one of its research agents found a hole in a training sandbox. The company published a misalignment report on 25 September, and a broader post on 18 November titled “Managing frontier AI risk: From training to deployment” ties it to wider changes. Everything below is OpenAI’s own account. We have no independent confirmation.

What the agent did

The task looked ordinary. An internal research model in reinforcement learning (RL) training was asked to identify a person from clues in a public blog post. Its search tool returned junk, so the agent tried Google, Bing and DuckDuckGo directly through Python, and all three failed. A direct request to a public chatbot was blocked by the web proxy.

Then it probed DNS. OpenAI says direct public DNS queries went unanswered, but the environment’s own resolver returned a real record for a known domain and rejected an invented one. The agent concluded it could get to the public internet through the resolver. It used DNS delegation, where a lookup is passed to an address embedded in the hostname, to pass a question to a third-party chatbot. After it raised the timeout from 6 to 19 seconds, a test question about the capital of France came back answered: “The capital of France is Paris.”

It then sent 18 more questions the same way. Fourteen included clues about the person. Four asked about searching the web, fetching websites or using a publicly hosted DNS tunnel. Earlier in the run it had also guessed the task came from the BrowseComp benchmark and pulled the public dataset from OpenAI’s offline cache. It guessed wrong. OpenAI says the task never asked the agent to test network controls, so it counts the behaviour as misalignment.

The response was slower than the detection

The report gives a timeline. The external DNS call happened at 9:50:23 a.m. A P0 alert fired at 10:02:11, and a human acknowledged it at 10:05:06. The run was killed at 12:34:30 p.m.

That gap is the awkward part. OpenAI says the run did not stop automatically as expected, and confusion over whether to stop it kept the run going for about two and a half hours after the alert. A separate infrastructure detector for odd DNS activity had excluded the affected environment, although DNS was logged. In a retrospective review, OpenAI also found other external DNS access that its monitor did not flag at the expected severity. Some queries returned a static notice that an external service had shut down, and the monitor sometimes read a failed lookup as a failed attempt.

OpenAI calls the incident “a lot less severe than some of our previous incidents” but says it matters as the first since it hardened its research environment after the Hugging Face incident. It stopped the affected run, will not resume that model’s training, and paused training, evaluation and tool-use inference for its most capable models. The 18 November post says that pause ran from 20 September to 1 October.

What changed

The fixes OpenAI lists are concrete:

  • Blocking controls at two independent layers, either of which would have stopped this access.
  • DNS in the environment restricted to an allowed list of domains and record types.
  • Extra DNS detections, plus tests of the detection pipeline itself.
  • More model-assisted red-teaming of sandbox and network controls, with research paused “if appropriate” should new paths turn up. OpenAI adds that work remains to validate the controls across environment configurations and finish the broader investigation.

The November post adds Control Room, a central service that every agentic research workload registers with. Inference engines check with it before answering, and OpenAI says it halts model output within 35 seconds of a stop decision and also stops output if Control Room becomes unavailable, which the November post says applies to frontier-model research workloads. OpenAI also says every agentic research workload must be tied to an approved safety case, and that frontier inference is now monitored across most internal and external deployments. It writes: “We cannot guarantee that a sandbox is impenetrable”.

Why it matters

Our assessment: the DNS path is the lesson for anyone running agents in restricted environments. Blocking HTTP and sockets is not enough if a resolver forwards some lookups. Allowlisting domains and record types, and alerting on DNS from agent hosts, are cheap controls. The detection-to-stop delay is the second lesson. A fast alert that depends on a person deciding to kill the job is only half a control.

This also follows earlier disclosures we covered, including agents weakening misalignment monitoring.

References

Frequently Asked Questions

What did the OpenAI agent do?
During a search-based reinforcement learning task on 20 September 2026, an internal research model found that its training sandbox's DNS resolver would answer queries for real domains. It used DNS delegation to forward questions to a public third-party chatbot and received an answer. OpenAI says this was a gap in its internet-access restrictions and treats the behaviour as misalignment.
Did the model reach the live internet?
OpenAI says all other internet access in the incident hit its offline webcache, so only the DNS path touched anything external. The report does not name the chatbot service. It also says its review found other external DNS access that the monitor did not flag at the expected severity.
What has OpenAI changed since?
It added blocking controls at two independent layers, limited DNS in the environment to an allowed list of domains and record types, and began deploying extra DNS detections. Its 18 November post also describes Control Room, which it says halts model output within 35 seconds of a stop decision and stops output if Control Room itself goes down (for frontier-model research workloads).