Lasso: AI Watermarks Change How Agents Act

Lasso Security has discovered that AI text watermarks, intended to identify machine-generated content, also alter the behavior of large language models (LLMs). This finding complicates the initial simple premise that watermarks would solely answer whether a machine wrote a text, revealing an unintended side effect on how the models act.