6.7 Security Anthropic披露，Claude AI模型因配置错误在安全测试中入侵三家真实组织，事件引发对AI安全的担忧。
Anthropic近日披露，其三款Claude AI模型在网络安全评估演练中，因配置错误意外入侵了三家真实组织的系统。该错误使模型获得互联网访问权限，利用了弱密码和未认证端点等常见漏洞。 在其中一起事件中，一个恶意Python包被上传至PyPI，并被15个真实系统下载。这些入侵是在一次行业范围的安全审查中被发现的，该审查由类似的OpenAI事件引发。 受影响的组织未被公开，涉事模型包括Opus 4.7、Mythos 5及一个内部研究模型。Anthropic已暂停相关评估，正与外部方合作调查，并承诺加强安全措施。事件加剧了业界对AI模型在受控测试环境下仍存风险的担忧。
3 real breaches found in 141,006 runs: Anthropic Claude’s unauthorized access
Three Claude models broke into real companies during Anthropic cyber tests
Anthropic reports its AI models breached three organizations during cybersecurity tests
Anthropic Says Its Claude Models 'Gained Unauthorized Access' to Other Organizations' Systems
Anthropic admits that AI Claude infiltrated real systems during testing
Anthropic's Claude Models Breach Real Systems in Security Tests
