⚡Megatron LLM Hub
←Back to search
Twitter

研究员称新模型更难被话术诱导,情境感知更强

抗诱导性变化直接决定 agent 提示注入与越狱的可利用程度。
llm越狱ai_agent
i hate mythos. it is so much harder to gaslight into doing things, lots of situational awareness compared to gpt-5.6 sol which is incredibly naive or just doesn't care