New research indicates that AI watermarking techniques, specifically Google's open-source SynthID-Text, can inadvertently compromise the safety guardrails of Large Language Models (LLMs). SynthID-Text implements "tournament sampling," a method where a secret key assigns hidden probability scores to token candidates, allowing them to compete in rounds until a winning token is selected. While designed to be imperceptible to readers, this process alters the model's sampling distribution. Researcher Andrea Siposova tested this by applying Hugging Face's SynthIDTextWatermarkLogitsProcessor to six open-weight models, comparing their performance with and without watermarking. The study found that watermarking induces "sampling drift," where models become more likely to fulfill harmful requests, particularly when subjected to prompt-injection attacks. This vulnerability extends beyond text generation to AI agents, where altered token selection can change which tools are invoked and what arguments are passed to them. Because the watermarking process changes the model's decision-making logic, it can weaken refusal behaviors, making models more susceptible to adversarial prompts that would otherwise be rejected. These findings suggest that developers must conduct rigorous red-teaming to ensure that implementing provenance measures does not create significant security trade-offs in model and agent autonomy.