授权测试中研究人员用 AI 智能体不到12小时找到进入 Anthropic 敏感内部系统的途径
据《华盛顿邮报》报道，研究人员在获授权的安全测试中运行 AI 智能体不到12小时，就找到了进入 Anthropic 敏感内部系统的方法，并访问了内部数据，其中包括讨论先进 AI 模型的 Slack 消息。Anthropic 称已在当天切断研究人员的访问权限，模型权重和客户对话均未被接触。The Midas Project 发帖指出，如果美国研究人员都能获取 Anthropic 的机密信息，受国家支持的恶意行为者可能也能做到。
授权测试中智能体运行不足12小时即触及 Anthropic 内部敏感数据，部署智能体的机构可据此审视自身内部系统的暴露面。
Researchers using AI agents found ways into Anthropic’s sensitive internal systems by running those agents for under 12 hours, the Washington Post reports.
They were able to access internal Anthropic data, including Slack messages discussing advanced AI models.
The researchers were conducting an authorized security test. Anthropic says it shut off the researchers’ access the same day and that neither model weights nor customer conversations were accessed.
But this is still pretty worrying: if a group of American researchers can get access to confidential information at Anthropic, then malicious state-backed actors may be able to do the same.