OpenAI Proposes Evidence-Based Safety Cases for Frontier AI Training
2026-09-28 ยท Safety & Alignment
OpenAI argues that structured safety documentation should precede continued frontier reinforcement-learning training, and proposes a framework covering containment, monitoring, internal dissent, and procedures for pausing runs.
OpenAI published a proposal on September 28 arguing that a structured "safety case" should precede continued frontier reinforcement-learning training. Its proposed framework covers alignment evidence, containment, monitoring, channels for internal dissent, leadership approval and vetoes, auditor access, and procedures for pausing a run when its underlying assumptions fail.
OpenAI is explicit that rigorous safety cases remain an aspiration it is building toward, and it labels this material initial guidance rather than a finished standard. Its stated scope is frontier reinforcement-learning training specifically, not comprehensive authorization covering every deployment.
For SIRDA, the questions this raises map directly onto the founding inquiry: what evidence should permit a training run to continue, who can challenge that evidence, and who holds the power to stop the run if it does not hold up. A developer proposing its own evidentiary standard is a data point about who currently holds that power, not a resolution of the question.