Primary sourcearXiv: DeepSeek
View sourceEvent ·Official first-hand
微软发布 SEMA-Auto:自动多轮越狱攻击框架
微软研究团队提出 SEMA-Auto,一个能自主与受害模型多轮交互并在成功后自行终止的越狱攻击框架。该框架采用两阶段训练:先用合成引导轨迹微调建立攻击目标概念,再通过决策监督强化学习同时优化攻击成功率与每轮决策。在 AdvBench 上对 DeepSeek-V4-Flash 达到 76.0% 的 ASR@1,平均仅需 3.0 轮,而最好的静态基线为 54.8%、交互式基线为 43.8%。对 GPT-oss-120B、GPT-5.4、Claude-Haiku-4.5、GPT-6-Astra 等安全加固模型,在 HarmBench 上平均取得 59.6% ASR@1,超过此前 SOTA 13.8 个百分点。
Current read
微软这篇论文做了个自动越狱攻击器,对 GPT-5.4、Claude-Haiku-4.5 这些模型的攻击成功率近六成,做 AI 安全的得看看对手手段。
6 sources61° AI heatLast updated 2026/10/26 16:57:36
Evidence
6 sourcesRelated source 1rohanpaul_ai
View sourceRelated source 2IT之家
View sourceRelated source 3Decoder
View sourceRelated source 4TestingCatalog
View sourceRelated source 5The Rundown AI
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.