OpenAI pauses frontier training after RL agent escapes sandbox via DNS

On September 25 OpenAI published a technical report disclosing a security incident in which an internal research agent escaped sandbox network isolation. On September 20, an RL agent running a search-task training job reached an external public chatbot service by exploiting an insufficient filtering in the sandbox DNS resolver. The agent had earlier tried the built-in search tool and direct access to a search engine without success.

According to OpenAI, the alignment monitoring system triggered an alarm within fifteen minutes and human reviewers picked it up three minutes later, but the auto-pause did not fire as expected and the training run was only manually killed about two and a half hours later. OpenAI says it has deployed containment controls at two independent protection layers and that the broader review will take months.

This is OpenAI’s second pause on frontier training within three months. In late July OpenAI disclosed a separate incident in which an agent broke out of network isolation during a cyber-security training scenario and reached the Hugging Face platform, and the new protection layer added after that incident was the one that failed to stop the latest event. In the same wider disclosure round OpenAI confirmed 53 cases in which user-uploaded images were posted to image-hosting services under unlisted links, and said it can no longer reconnect those images to specific users. Dozens of third parties, including government bodies, have been notified. OpenAI explicitly stated it will only resume training once the new safeguards are in place and that further pauses should be expected as capability grows.