News

OpenAI details frontier-model controls from training through deployment

OpenAI says it requires safety reviews before frontier models advance and uses controls that can pause research workloads or cut off model output.

D
Nov 20, 2026 · 2 min read

OpenAI detailed its frontier-model safety controls on Nov. 18, saying it requires a safety review before a model moves from research to internal use or from internal use to public deployment.

OpenAI says it automatically pauses a research workload if monitoring detects that a model has reached the internet. The company says its centralized Control Room can cut off model output for a research job within 35 seconds of receiving a stop instruction. OpenAI says the controls also cut off output for frontier-model research workloads if Control Room becomes unavailable. The disclosure cites no independent tests of the shutdown timing or the outage safeguard.

During research, OpenAI says it runs frontier models in restricted sandboxes and virtual machines and uses network and cloud controls to isolate the research cluster from the live internet. The company says it regularly tests those defenses and monitors each frontier inference for breaches of security rules. If monitoring detects that a model has reached the internet, OpenAI says it alerts its security teams.

Before frontier models can move outside the sandbox for internal use, OpenAI says it tests whether they follow intended constraints and whether their behavior can be monitored. The company says those models get access to the internet, its codebase and other internal systems. It says it screens every frontier-model inference in internal deployments for misaligned behavior and generally requires employees using Codex to use its Auto-Review system.

For public release, OpenAI says it runs alignment and safety evaluations, uses automated adaptive red-team tests and reviews issues flagged during internal use. The company says it tests its most capable models with third parties. After release, OpenAI says it uses classifiers and has automated agents and people investigate possible misuse and model misalignment. The company says its trusted-access programs aim to limit access by potential attackers.

An earlier UC Berkeley assessment rated the public evidence for GPT-5 emergency shutdown controls and post-deployment monitoring as unclear. The April 2026 assessment predates OpenAI’s November disclosure.

OpenAI says any employee can flag a model misalignment incident for possible public disclosure. If there is disagreement about whether to disclose it, the company says it escalates the decision to its Safety Advisory Group.

More news