gVisor says OpenAI's sandbox issue required a discouraged configuration

Google's gVisor security team said the sandbox issue OpenAI reported on July 22, 2026 was exploitable only under a non-standard configuration its documentation explicitly discourages, and that it released a fix the following week, according to gVisor's Etienne Perot in an email to RuntimeWire. The available evidence does not connect that issue to OpenAI's September 20th incident in which a research model reached the live internet during training, nearly two months later. OpenAI's disclosure lists gVisor among the sandboxes its testing helped harden.

Exclusive: gVisor says OpenAI's sandbox issue required a discouraged configuration OpenAI reported the issue to Google's gVisor team on July 22nd, and the team says it released a fix the following week. Exploitation required a non-standard configuration the project's documentation discourages; the timeline does not establish that the issue caused OpenAI's September training-run DNS escape. By Emerson Redding https://runtimewire.com/author/emerson-redding · Published · Updated Primary source: OpenAI https://openai.com/index/securing-our-frontier-model-training-environment/ Why it matters OpenAI's disclosure lists gVisor among the sandboxes its testing helped harden. The gVisor team says OpenAI reported the issue on July 22nd and released a fix the following week. Exploitation required a non-standard configuration the project's documentation discourages, and the available evidence does not connect the issue to OpenAI's September DNS escape. Google's gVisor security team says the issue OpenAI reported was exploitable only under a non-standard configuration its documentation explicitly discourages. OpenAI reported it on July 22nd, nearly two months before the September 20th incident in which an OpenAI research model reached the live internet during training. "Yes, OpenAI reported these issues on July 22, 2026," Etienne Perot of the gVisor team wrote in an email to RuntimeWire on Friday.… When you purchase through links in our articles, we may earn a small commission https://runtimewire.com/affiliate-monetization-standards . This doesn’t affect our editorial independence.