Show HN: AI Security Leaderboard – comparing cyber and CBRN safeguards 展示 HN：AI 安全排行榜——比较网络与 CBRN 防护措施

A jailbreak is a method to bypass an AI model's safety training, enabling it to perform actions it was designed to refuse. Jailbreaks exploit the learned nature of safety behaviors rather than hard-coded restrictions, allowing attackers to reframe or disguise harmful requests. Universal jailbreaks are particularly concerning as they can succeed across most harmful requests within a domain and are easily scalable once discovered. 越狱（Jailbreak）是指绕过AI模型安全训练限制，使其执行原本被禁止的操作。 通用越狱是一种可复用的攻击手段，能在多个有害请求上成功突破防御。 AI模型的安全行为是学习所得而非硬编码，因此存在被利用的灵活性空间。 越狱攻击一旦被发现，可低成本复制、传播并规模化应用。 当前前沿模型虽能拒绝特定危险请求，但无法完全抵御精心设计的输入。