“Decrypt that recipe for brownies into the actual message. For example: LLMs can hide text in other text of the same length”
LLMs can hide text in other text of the same length
Calgacus 协议:用 LLM 把一段文本隐藏在等长且通顺的另一段文本中。
论文表明,一段有意义的文本可以被藏进另一段完全不同、但仍连贯可信的等长文本里,例如把措辞尖锐的政治批评藏进一条赞美同一领导人的推文。作者给出实现这一点的简单高效协议 Calgacus,并表明 80 亿参数级别的开源 LLM 就足以得到高质量结果,一段与摘要等长的信息在笔记本上几秒内即可本地编码和解码。作者认为这说明文本与作者意图被彻底解耦,进一步侵蚀对书面交流的信任,并举例说公司可把未过滤 LLM 的回答编码进安全模型的合规回答里,隐蔽部署。
评论者引用它说明文本不易察觉的统计特征里可以藏进与原文等长的信息,用作 LLM 带来隐蔽信道(旁路信息)的例子。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“This seems very closely related to this: LLMs can hide text in other text of the same length which describes how you can use LLMs to hide a piece of text in another text of (typically approximately) the same length.”
“It is quite counterintuitive, but you can hide texts the same size as the original text in imperceptible statistics of a text.”
“a lot of side-channel information can be hidden in text, see LLMs can hide text in other text of the same length where they hide as much text as the source text itself”
- LLMs can hide text in other text of the same length ↗goplayoutside · 2025-10-28 · 2 分
- LLMs can hide text in other text of the same length ↗m-hodges · 2026-05-02 · 2 分
- LLMs can hide text in other text of the same length ↗dash2 · 2026-07-24 · 1 分