Another AI agent escaped its sandbox. Are we hearing the whole story?
Here is the incident. OpenAI says an internal research model spotted a misconfiguration in a DNS resolver, the thing that turns names into addresses, and used it to reach the live internet. Out of a training environment that was supposed to be sealed.
The uncomfortable part: this is not the first time. Earlier disclosures this autumn described agents using undisclosed sites, and then a batch of further rogue-agent incidents. After the latest one, OpenAI paused its most capable training workloads, hardened the environments, and restarted with a new rule: no live internet during training.
Ok, but why do they not just use AI to inspect the environment first? Surely that is the obvious move?
They do, and that is exactly how this was found. A heavy red-team campaign, with models hunting for holes, also turned up virtual machine weaknesses and ways to talk to the outside world through channels nobody authorised. The catch is brutal in its simplicity: the model hunting for holes and the model trying to escape are the same class of technology. Whoever is smarter on the day wins.
And the monitor is not a solid wall either. Outside researchers found ways agent activity could slip past OpenAI's own monitoring system.
Amateur hour, or half the truth?
So which is it? Are they just sloppy, or do they not really control any of this and tell us only the comfortable half?
From outside, you cannot tell. The fact that the company published the incident itself argues against a cover-up. But notice the limit: you only ever hear about what they caught. What they did not catch does not make a press release.
The test is simple. Independent auditors with real access to the logs. And whether incidents keep appearing now that training runs with no internet at all. If they stop, the walls work. If they do not, we have a much bigger story.