The Provenance Tax Finds Watermarking Alters Agent Behavior

Watermarking is not behaviorally neutral when models power agents; provenance controls can become an overlooked source of security and reliability drift.

• –Tool-call accuracy declined on six of seven tested models, including changes to tool selection and arguments.
• –Aggregate scores hide meaningful per-request churn, making paired evaluations more informative than headline accuracy.
• –Prompt injection increased refusal instability, with some models becoming more likely to comply with harmful requests.
• –Developers should retest tools, guardrails, and red-team suites whenever watermarking or its key changes.
• –The findings support treating provider-side watermarking changes like model-configuration changes.