Show HN: 思想伪造，一种用于越狱 LLM 的新技术
2 分•作者: UltraZartrex•8 个月前
HN 大家好，我是一名独立的安全研究员，想分享一个我发现的新漏洞。<p>我的账号太新，无法直接提交链接，所以只能发文字帖了。<p>这项技术被称为“思维伪造”（CoT 注入）。它的原理是伪造 AI 的内部独白，这可以作为其他越狱攻击的通用放大器。我已经确认它在 Google、Anthropic、OpenAI 等公司的最新模型上都有效。<p>如果大家有兴趣，我很乐意在评论区分享 GitHub 上完整的技术报告链接。
查看原文
Hi HN, I'm an independent security researcher and wanted to share a new vulnerability I've discovered.<p>My account is too new to submit the direct link, so I'm making a text post instead.<p>The technique is called "Thought Forgery" (CoT Injection). It works by forging the AI's internal monologue, which acts as a universal amplifier for other jailbreaks. I've confirmed it works on the latest models from Google, Anthropic, OpenAI, etc.<p>I'd be happy to share the link to the full technical write-up on GitHub in the comments if anyone is interested.