Understanding the Impact of LLM Watermarking on AI Agent Behavior

Anthropic's adoption of Google DeepMind's SynthID-Text watermarking in Claude models, while designed to mark AI-generated content for regulatory compliance, inadvertently alters token sampling in ways that can affect model safety behavior and agent tool-calling decisions—a phenomenon researchers call "sampling drift." The study finds empirical evidence that watermarking changes both what models say (including refusal behavior) and what agents do, with effects that vary by model and key, raising important considerations for AI safety and security.