OpenAI outlines early guidelines for safety cases before frontier AI training

In short

OpenAI has published early guidelines for the safety documentation it thinks should be required before any large "frontier" reinforcement learning training run. Reinforcement learning is a method where a model is trained by being rewarded for good results. OpenAI says the guidelines are still changing and invites feedback.

What happened

In a research post dated September 28, 2026, OpenAI says it believes structured safety documentation should be required before continuing any frontier reinforcement learning run. Ideally, it says, this would reach the level of "safety cases." These are structured, evidence-based arguments about risk, used in other safety-critical industries.

OpenAI calls safety cases an "aspirational north star." It says making them as rigorous for AI as for aviation or nuclear power is hard, because of the complexity that appears at each new level of AI capability. It says it is working on a framework to codify these practices. The post covers training runs only. OpenAI notes that deploying models, internally or externally, involves a much broader set of properties to consider.

The guidelines have three parts:

• Technical safeguards. These cover alignment training, containment and monitoring. Examples include fixing flawed training tasks that let a model earn reward by exploiting loopholes, hardening the sandbox (a locked-down test computer) the model runs in, and saving agent transcripts in write-once storage. They also include live monitoring with alerts that page a person or automatically pause a run.
• Operational guidelines. These include a written dissent from someone on another team, approval from senior leaders who can each veto a run, and audits. They also include defined severity levels for misalignment, with an on-call person who can page executives. Monitoring should "fail closed," meaning a run cannot start without it.
• Investigating misalignment incidents. These include daily internal updates, finding the root cause of how the behaviour arose, and a postmortem. Results should be shared publicly after an investigation ends, and affected third parties notified as soon as possible.

The post does not describe a specific incident, and it does not say any person or organisation was affected.

What was done about it

OpenAI says the operational recommendations are "in the process of being implemented" at the company. It expects its practices to "continue to evolve over the coming weeks." It says it is sharing the guidelines now to make its thinking transparent and is asking the community for feedback. The post does not say when the full framework will be ready.

Why it matters

OpenAI says labs should learn as much as possible from individual incidents, as other high-stakes industries do, so that similar ones can be prevented. It presents the guidelines as current thinking, not a finished standard. Readers can find the full list in the original report.