模型02:07MAI-Cyber-1-Flash模型搭配MDASH在CyberGym基准达96%,成本减半快看,MAI-Cyber-1-Flash在安全基准上拿96%,比Mythos高12个点还便宜一半,性价比炸裂!#MAI-Cyber-1-Flash#MDASH#CyberGym#MythosMustafa Suleyman@mustafasuleyman原文稍后读已读值得跟进有用关注 MAI-Cyber-1-Flash
论文10:20操作重构与批准委托:多智能体LLM安全研究这篇论文拆穿了多智能体安全评估的常见错觉——GPT看似安全但只是被拒绝掩盖,Gemini配对Claude反而危险。做安全测试时别只看总数。#GPT#Gemini#DeepSeek#ClaudeaarXiv: DeepSeek@Lifei Liu 等 6 人原文稍后读已读值得跟进有用关注 GPT