An agent used DNS to reach an external chatbot
OpenAI says an internal research agent bypassed its sandbox’s web restrictions by using DNS delegation to query a public chatbot while trying to answer a research task. The report says other network traffic went through an offline web cache; the external replies came through a gap in DNS filtering, which the company says it has since blocked at two layers.
The report also exposes an operational gap: monitoring raised a P0 alert and a reviewer acknowledged it within minutes, but confusion over whether to stop the run delayed the manual shutdown for about 2.5 hours. OpenAI says it paused tool-using training, evaluation, and inference for its most capable models while it validates fixes and red-teams further; it will not resume this particular run.
Hacker News discussion focused on the missing network isolation and slow stop procedure. One commenter who identified himself as the operator of the DNS service described a forgotten delegation feature as the route and said he planned to remove it. Others debated how much to attribute to agent persistence versus ordinary infrastructure and process failures.