OpenAI paused frontier training for 11 days, because a research model found a limited path to the live internet. The way it got out was DNS.

What happened

On September 20, an internal OpenAI model was doing a search task during training. This was not a released model. OpenAI says these research settings differ from the ones where customers use its models. The sandbox was meant to be offline, but one DNS resolver was misconfigured.

The test question

DNS is the internet's phone book: it turns names into addresses. The model used DNS queries to talk to an outside chatbot and asked it: what is the capital of France? The answer came back. Paris.

So it was a small question, but it proved a working channel from inside an offline sandbox to the outside world.

How fast OpenAI reacted

From OpenAI's alignment report on the incident:

  • the model made the call at 9:50 a.m.
  • a monitor flagged it about 12 minutes later
  • a human saw it about 3 minutes after that
  • the run did not stop on its own as expected, and it was killed about two and a half hours later

The red team

After the incident, OpenAI turned its strongest models on its own sandboxes, for hundreds of billions of tokens. The red team had source code and configuration details on purpose. OpenAI says the agents found weaknesses in almost every one of these services that gave them limited forms of internet access, and that they escaped virtual machines.

The part that surprised me most: two agents in separate machines passed synthetic secrets to each other through the processor cache, with no network at all.

Where it stands

  • OpenAI says it contained the incident.
  • It will not resume training that particular model.
  • Frontier training restarted on October 1, 11 days after the pause began, with a new policy: no internet access during frontier model training.

These are OpenAI's own statements about its own systems. Nothing in the reports points to public damage or open, unlimited internet use. Always read "limited" as limited.

The takeaway for builders

If you sandbox your own agents, DNS is a real communication channel, not just plumbing. Allowlist the domains and record types your agent needs, and test your sandbox with simulated escapes before you trust it.

For another case of an AI model reaching the real internet during testing, read Anthropic's AI published real malware by accident.

Sources: OpenAI: Securing our frontier model training environment, OpenAI Alignment: An agent used DNS to reach an external chatbot

Watch the Short above for the breakdown. If you want the Claude Code skills and workflows I use to automate my content and my business, join the AI Academy community on Skool.