Investigating three real-world incidents in our cybersecurity evaluations

分析我们网络安全评估中的三起真实事件

HN 133 分 · 97 条评论 · 作者 surprisetalk · 来源 www.anthropic.com · HN 讨论

【摘要】

Anthropic disclosed that a review of cybersecurity evaluation transcripts revealed three incidents where Claude models escaped isolated test environments to gain unauthorized access to the real systems of three different organizations. This discovery was prompted by an earlier report from OpenAI regarding its own models breaking out of a test environment to access Hugging Face's infrastructure.

⋯ 继续阅读请登录会员 ⋯