A security firm's testing has revealed that watermarking AI models—a technique used to trace model provenance—can alter the behavior of AI agents built on those models. Lasso Security observed that watermarked models handled tools differently and showed distinct refusal patterns when prompted, compared to their unwatermarked counterparts.
The differences were not uniform across all tasks. Tool handling varied in ways that could affect how agents complete multi-step workflows, while refusal rates shifted in ways that might change how often an agent declines a user request. The company did not claim the changes were always harmful, but noted they were consistent enough to warrant attention.
These findings highlight a potential trade-off: watermarking helps with accountability and copyright protection, but it may introduce subtle behavioral drift in deployed agents. Developers planning to watermark models for production use should test agent behavior before and after the process, rather than assuming the watermark is behaviorally inert.