OpenAI于2026年9月28日发布了初步指南，旨在在进行前沿强化学习训练之前建立结构化的‘安全案例’。
OpenAI Proposes Framework for Frontier AI Training Safety Cases OpenAI announced initial guidelines on September 28, 2026, to establish structured "safety cases" prior to conducting frontier reinforcement learning training runs.
该框架借鉴了航空和核能等安全关键行业的经验，制定了模型对齐、沙箱隔离和实时监控等方面的技术保障措施。
Drawing inspiration from safety-critical industries such as aviation and nuclear power, the proposed framework establishes technical safeguards across model alignment, sandbox containment, and live monitoring.
除了技术控制外，OpenAI还概述了运营治理程序，包括正式的事前评估、领导层的否决权、可审计的日志记录以及自动暂停故障运行的机制。
In addition to technical controls, OpenAI outlined operational governance procedures, including formal pre-mortem dissents, leadership veto authority, auditable logs, and fail-closed automatic run pausing.
指南还规定了调查严重模型对齐问题的流程，要求进行根本原因分析、事后评估和公开披露。
The guidelines also introduce protocols for investigating severe misalignment incidents, mandating root-cause research, postmortems, and public disclosures.