AI Text Watermarks Can Change Agent Actions and Safety Decisions
Machine-readable provenance changed tool choices and weakened refusals under attack in a paired benchmark study.
A paired study applied SynthID-Text to tool-calling and refusal tests and found that watermarking changed individual outputs even when aggregate results obscured the movement. Across 21 model-temperature combinations, tool-call verdict churn averaged 6.5%; under a fixed prompt-injection test, watermarking raised harmful-compliance rates by 12.5 percentage points for one model and 9 points for another. Publishers and information providers putting watermarked models inside customer-facing agents may therefore need configuration-specific testing to keep a provenance control from introducing hidden reliability costs.