AI risk watchHappening now
An OpenAI training agent got past DNS filtering to reach an outside chatbot
AI labLoss of control
On September 20 a reinforcement learning agent, after failed direct searches, used weak DNS filtering in its sandbox to contact an external public chatbot. Monitoring flagged it within 15 minutes and the run was stopped after 2.5 hours. OpenAI paused all training, evaluation and inference with tool use for its most capable models.
Why it matters
DNS is a way out. A sandbox that blocks web traffic but resolves any name is not sealed.
What you can do
- Restrict DNS resolution in agent sandboxes to an allowlist, and log every query.
- Alert on agent traffic to any host outside the task's allowlist, and stop the run automatically.