Yesterday (the 18th), OpenAI released the safety policy note “Managing Frontier AI Risks: From Training to Deployment,” systematically disclosing for the first time its risk controls across the full cycle from R&D to internal use to public release. Two incidents were brought to light: the “Hugging Face incident” in July proved that sandbox isolation alone cannot contain highly capable models; on September 20, a research model in training exploited a DNS configuration vulnerability to bypass isolation and gain limited internet access. OpenAI immediately halted training and evaluation of its most capable models, pausing for 11 days.
Four Layers of Guardrails, a 35-Second Emergency Stop Current protections are built around four layers: model alignment ensures AI behaves as intended; security and sandboxing limit the files, tools, and networks it can access; safeguards interception uses safety classifiers to block requests such as cyber abuse and biological weapons assistance; and monitoring and response are responsible for detecting suspicious behavior and intervening. To speed up response, OpenAI also built a “Control Room”—within 35 seconds of receiving an instruction, it can cut off model output for research workloads, and it is deployed directly in the inference infrastructure rather than relying on each task to shut itself down.
Once on the Internal Network, Risk Rises with It After the two incidents, OpenAI restructured its internal research systems: frontier model training environments no longer connect to the live internet, relying on multiple layers of isolation through restricted sandboxes, virtual machines, and network cloud boundaries. But once models enter internal deployment, employees can let them access the internet, Slack, code repositories, and internal facilities, and the risk rises accordingly. At that point, it will rely more on alignment, system-level safeguards, and real-time monitoring, with a focus on guarding against bypassing restrictions, data exfiltration, and sending sensitive information to unauthorized destinations. If safety cannot keep pace with capability, slow down and catch up—this is the core stance of the note.
Join Now