OpenAI proposes a cutting-edge AI training 'security argumentation' framework, suggesting to go through the security documentation before starting the run Dynamic Beating AI News: OpenAI has officially released a long article titled "Towards Safety Cases for Frontier AI Training". It is mentioned in the article that before continuing to promote cutting-edge reinforcement learning training, structured safety documents should be completed, ideally benchmarking the aviation and nuclear power industries for "safety justification". The company acknowledges that due to the complexity of emerging capabilities, it is difficult for AI to achieve the same level of rigor as traditional high-risk industries. Currently, the framework is still being implemented internally and seeking feedback from the outside world. OpenAI explicitly only covers cutting-edge RL training, and internal and external deployment require further consideration of broader alignment properties. On the technical side, OpenAI breaks down security argumentation into three layers: alignment training, isolation, and monitoring. Alignment includes automatic/manual review of RL environments, tuning scorers to prevent rewarding hackers, offline alignment of evaluations and backtesting, and prohibiting automatic scorers from reading thought chains to prevent models from learning to evade monitoring. In terms of isolation, emphasis is placed on reinforcing multi-layer sandboxes and infrastructure, using training checkpoints as red teams, limiting high bandwidth communication between samples, and using one-time write storage to solidify dialogue records. In terms of monitoring, it is required that the model maintain detectability, with high recall of historical accidents and timely evaluation of newly added risks. High priority alarms must be manually responded to or automatically suspended from training within the agreed time limit. Suggestions from the operational side include: having other teams write objection rehearsals, having executives (research leaders, safety leaders, chief scientists, etc.) veto the start of the operation with one vote, holding training leaders accountable for safety arguments and accident response, suspending safety arguments in accordance with the manual when they fail, making them public to the internal supervision committee, granting sufficient verification authority to the auditing party, setting severity levels for misalignment, and allowing on duty personnel to call the CEO. The system should default to fail shutdown and cannot run without monitoring; At the same time, it is necessary to be able to roll back downstream data and ratings contaminated by misplaced models. In terms of accident investigation, OpenAI proposes benchmarking aviation investigation practices: internal rolling notifications during the investigation period, training dynamics through ablation and resampling, conducting operational and cultural reviews, developing detection and evaluation methods that do not directly fit accident samples, and conducting regression testing using accident derived evaluations. The investigation conclusion, review, and process changes should be disclosed to the public after completion, and affected third parties should be informed as soon as possible. [Original link]