AI brief
Lasso Security observed differences in tool handling and model refusals when AI model watermarking was applied.
Why it matters: Watermarking intended to label AI outputs can also change how agents handle tools and refuse requests.
Written by AI from The Register AI + ML's published text. Read the original for full details.
More coverage of this story (1)
Research finds AI watermarking changes model behavior
LLMs respond differently to harmful prompts when AI watermarking is used
Ars Technica AI · 17 Sept 2026, 18:33