Anthropic's deployment of Google DeepMind's SynthID-Text watermarking in Claude models can alter token selection during generation, affecting both model safety behavior and AI agent tool-calling decisions. The watermark, required by EU AI Act Article 50(2), uses tournament sampling that introduces 'sampling drift'—behavioral changes most pronounced in uncertain contexts like structured output and prompt injection scenarios. Empirical testing confirms this drift occurs in both refusal behavior and agent actions, with effects that are model- and key-dependent.