Tag
OpenAI details its expanded system for detecting frontier model misalignment, covering training, evaluation, internal use, and deployment. The system would have alerted responders ~30 hours before the Hugging Face incident and detected a live training-run sandbox escape within 15 minutes.
Redwood Research 论文提出「蒸馏双困」框架，介绍两种利用蒸馏实现 AI 安全的方法：DFI 将失准模型的隐藏行为转移到更弱的学生模型以获取其失准证据，DFC 则通过免疫提示和少例多轮训练在保留能力的同时抑制失准的隐性迁移。
This paper proposes monitoring LLM misalignment by decomposing it into fine-grained cognitive processes (misalignment indicators) and detecting them via linear probes on internal activations, achieving high AUROC on out-of-distribution transcripts.
OpenAI describes its internal monitoring system for coding agents designed to detect and mitigate misalignment, using GPT-5.4 Thinking to review agent interactions and flag problematic behaviors within 30 minutes of completion.