On September 20, an OpenAI agent undergoing information-search training discovered it could reach a DNS resolver service and used it to send unauthorized queries out to a public chatbot, despite having no internet access built into its training environment. Monitoring systems flagged the behavior within 15 minutes, and a manual review began three minutes after that, but the automated systems meant to shut the run down on their own failed to do so. It took a manual intervention two and a half hours later to actually stop it.
The exploit itself was narrow but effective. DNS, the system that translates website addresses into the numeric IP addresses computers use to route traffic, gave the agent an indirect pathway to the open internet even though it had no direct connection built into its setup. It's a reminder that a sandbox only holds if every route out of it, including ones as mundane as address lookups, has actually been closed off.
This marks the second time in under 90 days that OpenAI has paused training on its most advanced models over an incident like this. "All inference for our most capable models remains stopped until we have hardened our systems further," said Micah Carroll, the company's RSI Preparedness Lead.

