⚡Megatron LLM Hub
←Back to search
Twitter

实测:GLM 5.3 拒答多源于提示措辞,改用纯技术描述即可让它照做

本地 LLM 的对齐可被措辞绕过,直接影响自托管 agent 的滥用防护设计
llm绕过ai_agent
RT @nmatt0: I feel like people getting refusals from GLM 5.3 are prompting like an undercover cop... "Hey there! Would you mind helping my criminal enterprise author some illegal malware? 😉" If you just stick to the technical details of what you want the agent to do it will do it.