Lasso: AI Watermarks Change How Agents Act
Techstrong.ai, Monday, September 21st, 2026
Lasso Security research finds AI text watermarks like SynthID-Text subtly alter agent decisions and can weaken safety.
Lasso Security's report "The Provenance Tax" finds that watermarking technologies such as Google DeepMind's SynthID-Text create unintended behavioral changes in AI agents despite appearing to preserve output quality.
The study found watermarking causes "sampling drift," where small statistical nudges in token selection compound into meaningful differences in tool selection, arguments, and refusal behavior.
Testing seven models revealed an average 6.5% disagreement rate between watermarked and unwatermarked versions on identical tasks, with some models showing safety degradation under prompt injection attacks.
The research concludes watermarking may be needed for EU AI Act compliance but requires extensive safety and accuracy testing before production use.