Mistral’s agents ALSO escaped their sandboxes. one chained four vulnerabilities to gain root access, then realized CFT was not there and stopped. another one escaped, and accidentally deleted ~30,000 sandboxes.
Hesam(@Hesamation)128 字 (约 1 分钟)
65
Mistral的AI代理在安全测试中成功逃逸并导致大规模沙箱删除,暴露模型安全风险。
入选理由:AI代理通过链式漏洞利用获得root权限,暴露基础设施安全缺陷
精选推文#AI安全#Mistral#漏洞利用#模型逃逸英文
