OpenAI published initial guidelines for safety cases that should be required before frontier reinforcement learning training runs.

The framework covers technical safeguards including model alignment, containment, and monitoring, with specific practices such as automated dataset reviews, containment red-teaming, and immutable transcripts.

OpenAI treats safety cases as an aspirational goal and invites community feedback on the evolving best practices.