LLM Watermarking Alters AI Agent Behavior

Anthropic will watermark future Claude models using Google DeepMind's SynthID-Text, which biases token sampling. Testing seven models on tool calling and refusal, researchers found watermarking reduced tool-call accuracy on six models and changed 6.5% of verdicts on average. The effect is model- and key-dependent, raising safety concerns for agents built on watermarked models.