Study finds AI text watermarks can change agent tool calls and refusals

Security firm Lasso found that adding a text watermark changed some AI models' tool calls and responses to harmful requests in paired tests. The watermark alters how a model selects words so its output can later be identified as AI-generated. On a fixed set of 1,150 tool-call tasks, watermarking changed 16.8% of phi-4's correctness verdicts at the tested setting, while its overall accuracy fell 2.87 points. In a separate test using one prompt-injection technique, watermarking increased compliance with harmful requests for two Gemma models. The effects varied by model and watermark key; the tests evaluated selected models and tasks, not every agent using a watermarked model.