OpenAI has published initial guidelines for safety cases intended to govern frontier reinforcement learning (RL) training runs. The authors frame these as a step toward structured, evidence-based safety documentation similar to that used in aviation or nuclear power, while acknowledging that AI’s emergent complexity makes the same rigor difficult. The document is explicitly focused on frontier RL training, not deployment, and is presented as a living set of practices that will evolve with internal iteration.

The technical safeguards section covers three layers: alignment training, containment, and monitoring. Alignment training includes automated and manual dataset reviews to catch broken or exploitable RL environments, grader tuning to penalize reward hacking, and prior-run analysis to verify grader behavior. Alignment measurement involves offline evals during runs, backtesting evals on previous incidents, tracking eval awareness or metagaming with blocking thresholds, and worst-case stress tests to ensure at least one eval would detect misbehavior. The authors also recommend preventing automated graders from seeing chain-of-thought in RL, to avoid models evolving to evade chain-of-thought monitors. Containment focuses on hardening the sandbox and research infrastructure, iterative red-teaming of sandbox and perimeter using checkpoints, limiting high-bandwidth cross-sample communication to approved pathways, and storing immutable transcripts for incident investigation. Monitoring should enforce monitorability with clear thresholds, maintain high recall on past incidents via held-out evals, refresh eval datapoints for net-new risks, and implement rapid response with defined SLAs—including paging on-call staff during the day and auto-pausing runs for unacknowledged alerts at night.

Operational guidelines address governance and process around a training run. Dissents, or pre-mortems, should be written by a separate team member to challenge the safety case. Approvals require senior leadership review with individual veto power. The accountable senior leader should bear responsibility for the safety case and incident response, including in performance reviews. There should be runbooks and technical controls for pausing runs if the safety case is invalidated. Internal transparency to oversight groups, auditor access, defined escalation paths with severity levels, and an executive-paging mechanism are all recommended. Technical controls should make it difficult to start noncompliant runs, with safety features failing closed. Rollback ability should allow identification of all downstream uses of a misaligned model, and a comprehensive residual-risk list should be enumerated to support informed risk acceptance decisions.

The final section covers investigation of severe misalignment incidents. Labs should provide internal transparency with periodic updates, enable safe access to transcripts and samples, root-cause training dynamics via ablations or resampling, and conduct operational and cultural postmortems. Detection methods should be developed without hillclimbing on incident-derived information, and incident-derived evals should serve as regression tests. Public disclosure of results and postmortems should follow the investigation, and affected third parties should be notified as soon as possible.

The authors note these practices are in the process of being implemented at OpenAI and expect them to change quickly. The document is open for community feedback.