LLM Watermarking Can Alter AI Agent Tool Calls and Behaviour: Study
• Lasso Security conducted a study finding that LLM watermarking, a method used to identify AI-generated text, alters how open-weight models call tools and behave. • The researchers observed that these watermarking techniques changed individual tool-call outcomes and refusal behaviors, with the most pronounced effects occurring during prompt injection attacks. • These findings suggest that security measures intended to track AI content can inadvertently degrade the reliability and functional accuracy of AI agents.
analyticsindiamag.com



