OpenAI says senior leaders should each be able to veto a frontier training run
OpenAI has published a list of practices it says should govern a frontier training run, and it opens on a requirement rather than a principle: "We believe we are entering a new era in which structured safety documentation should be required before continuing any frontier reinforcement learning training run," the document says. Among the operational recommendations, a safety case "should be reviewed by members of senior leadership, who should each have the ability to veto the run", with a research organisation lead, the Head of Safety and the Chief Scientist given as examples.

OpenAI published the guidelines on September 28 and says they are still being implemented.
None of it is in force. OpenAI calls these "some initial guidelines" and closes the operational half by saying that "These represent our current recommendations and are in the process of being implemented at OpenAI. We expect our practices to continue to evolve over the coming weeks." The page is dated September 28 and limits its own scope: "this document is focused on frontier reinforcement learning training; internal and external deployment require considering a much broader set of alignment properties".
A safety case would cover three things
"Safety cases should cover three aspects of the technical stack: alignment training, containment, and monitoring," OpenAI writes, so that a model "does not try to take misaligned actions", would find it hard to break containment if it did, and would be caught by monitoring "before harm could occur".
Several of the recommendations are concrete enough to check. Transcripts should be saved "immutably for a reasonable period of time (for example, in write-once storage accounts) for all RL training, evaluation, and internal deployment", which OpenAI justifies as preventing inadvertent edits and enabling a reliable investigation afterwards. Automated graders should not see a model's chain of thought, "to prevent models from evolving to evade chain-of-thought monitors", and monitors for a model noticing it is being tested should run on alignment evaluations "with specific blocking thresholds".
Unacknowledged alerts would pause the run
The monitoring recommendation has an operating rota attached. Priority alerts should "result in action within a defined SLA", and the example OpenAI gives is to "page security on-call and the run owner on incidents during the day, and auto-pause runs for unacknowledged alerts at night". Safety features should "fail closed", so that it "should not be possible to start runs without appropriate monitoring enabled, or to disable the monitor from within RL training, evaluation, or an internal deployment".
Around that sit a set of internal checks: a member of another team should write a dissent against a drafted safety case "to find potential holes" in it, and the senior leader responsible for a run should be accountable for both the safety case and any incident response "including as part of performance reviews". The company has paused tool-use training after an agent reached non-public files.
Investigations would report daily and publish at the end
For a severe misalignment incident, OpenAI recommends that "incident investigations should have periodic updates presented internally (e.g., daily updates for ongoing investigations)", and that employees have "defined pathways to get more access, including raw transcripts and sampling from misaligned models, if it is safe and relevant for their work". Results should not stay inside: "Investigation results, postmortems, and operational changes should be shared with the public following the conclusion of the investigation. Affected third parties should be notified as soon as possible."
Safety cases themselves "should be made available to internal oversight groups (e.g., the Safety and Security Committee)", and auditors should get "sufficient access to verify that the claims of the safety case are valid and sound, and to raise gaps if found". The incidents OpenAI has disclosed so far, including a model that used a leaked API key during training, came out through its own misalignment reports, published while these recommendations were still being implemented.
No model, run or incident is named anywhere in the document, there is no date by which any of it applies, and OpenAI concedes the limit of its own analogy. It treats safety cases as "an aspirational north star we are building towards", harder to make rigorous for AI than for aviation or nuclear power "due to the emergent complexity at each new level of AI capability".