LLM Watermarking Alters AI Agent Tool Choices

Lasso Security tested Google DeepMind's SynthID-Text watermark across open models and found it changed which tools agents selected, the arguments they passed, and whether they refused harmful requests. Average churn hit 6.5% across 21 model and temperature combinations, with phi-4 at 16.8%. Anthropic plans SynthID-Text for future Claude models, and the EU AI Act mandates detectable AI outputs.