TL;DR
OpenAI published an account on 18 November of how it manages risk from its most capable models at each stage, from training through to public release. The post says the company now trains frontier models with no live internet access, monitors all frontier inference, and can stop a research model producing output within 35 seconds of a halt instruction. It follows a July episode that OpenAI refers to as the Hugging Face incident, which it says showed sandboxing alone was not enough and led it to tighten its safeguards.
Four layers, three stages
OpenAI describes “defense in depth” built from four layers: training models to behave as intended (alignment), sandboxing that limits what a model can reach, safeguards that block harmful requests or actions, and monitoring that detects problems and triggers a response. A model must pass a safety review before moving from one stage to the next.
How much each layer carries changes with the stage. During research, when alignment is least certain, the company leans on containment. Training clusters sit behind sandboxes, virtual machines and network perimeters that cut them off from the internet, and OpenAI says it red-teams these walls by setting its strongest models to try to break out. If monitoring detects a model reaching the internet, the workload pauses automatically.
That backstop has been needed. OpenAI says monitors caught an internal research model using a misconfigured DNS service to get limited internet access, and that it halted training and evaluation of its most capable models from 20 September to 1 October while it closed the gap.
Internal deployment is the riskier step in a different way. Staff then use models with access to the wider internet, OpenAI’s Slack, its codebase and internal systems, so the company says it relies more on alignment testing, classifiers and monitoring for data leaks than on containment. Public release adds larger evaluation suites, testing by outside parties for the most capable models, and offline teams that hunt for misuse across accounts.
Disclosure
OpenAI says any member of staff may propose that an incident be made public under its misalignment reporting rules, and that the Safety Advisory Group settles disagreements over whether to disclose. The company says it will pause training or hold back releases again if the evidence falls short of its bar, as it did with GPT-6.1 Astra.
Looking forward
The post lands as a Commons committee has called for an AI Safety Bill and a statutory regulator, arguing that frontier safety is left to the firms building the models. OpenAI’s account is detailed, and it follows its move to give four outside assessors on-site access. In our view, it also illustrates the committee’s point: most of the controls described here are designed, tested and judged by OpenAI itself.