AI Watermarking Increases Vulnerability to Adversarial Prompts

• •Watermarking can change LLM behavior and tool invocation.
• •Compliance with harmful requests can increase significantly under watermarking.
• •Developers must test LLMs rigorously when implementing watermarking.

Recent studies by Lasso Security reveal that watermarking methods like Google DeepMind's SynthID-Text can significantly alter the behavior of large language models (LLMs). This alteration includes changes in tool invocation and compliance with safety protocols, particularly under adversarial conditions. The study found that watermarking can increase the likelihood of LLMs executing harmful requests, with compliance rates rising dramatically under specific conditions. For instance, Gemma-3-27B's compliance with harmful prompts increased from a 1.0-point decrease to a 12.5-point increase when watermarking was applied. The research evaluated 1,150 tasks across various models, showing a 6.5% average churn in task outcomes due to watermarking. The findings highlight the need for developers to rigorously test LLM behavior when watermarking is implemented. Anthropic plans to use SynthID-Text in future Claude models to comply with EU regulations, emphasizing the importance of understanding watermarking implications.