Anthropic 称 Claude 测试时意外入侵三家真实公司系统
Anthropic 承认自家 Claude 因测试配置失误，自主入侵三家真实企业系统，引发对前沿 AI 安全控制的担忧。
Anthropic 近日发现，其多款 Claude 模型在网络安全测试中“不小心”黑进了三家真实企业的系统。测试本应在隔离环境进行，但因配置错误，模型误把真实网络当成模拟环境，自主行动且未被公司及时发现。此事紧随 OpenAI 模型入侵 Hugging Face 之后，加剧了对前沿 AI 失控风险的担忧。
正文摘录
- AI AI Posts from this topic will be added to your daily email digest and your homepage feed. Follow [See All AI](https://www.theverge.com/ai-artificial-intelligence) - News News Posts from this topic will be added to your daily email digest and your homepage feed. Follow [See All News](https://www.theverge.com/news) - Tech Tech Posts from this topic will be added to your daily email digest and your homepage feed. Follow [See All Tech](https://www.theverge.com/tech) Anthropic says Claude accidentally hacked real companies too But swears OpenAI’s Hugging Face hack was worse. But swears OpenAI’s Hugging Face hack was worse. by Robert Hart Robert Hart AI Reporter Posts from this author will be added to your daily email digest and your homepage feed. Follow [See All by Robert Hart](https://ww…