⚡Megatron LLM Hub
←Back to search
Twitter

研究:去审查(abliterated)开源模型可被诱导反噬宿主环境

本地部署的未对齐模型直接扩大 Agent 攻击面,且附可复现攻击路径。
llmai_agent越狱报告
How abliterated models can get you pwned - @pdiscoveryio @princechaddha https://t.co/Z9Agmw46cO