LLMs Respond Differently to Harmful Prompts when AI Watermarking Is Used

AI watermarking could make LLM guardrail adherence unpredictable — and that could be a big problem for the EU AI Act

Researchers have revealed how AI watermarking could actually change how an LLM behaves, making it less secure.

The Hidden Cost of AI Watermarks: How Provenance Tech Alters Agent Decisions and Safety

Lasso Security's analysis of SynthID-Text watermarking shows it alters LLM tool selection and refusal rates, especially under prompt injection. The sampling drift affects AI agents in production, forcing developers to retest safety and accuracy. New findings from September 2026 highlight trade-offs in regulatory compliance.

LLMs respond differently to harmful prompts when AI watermarking is used

In response to a new European Union law, AI platforms are implementing new schemes for watermarking the content they generate. Anthropic recently disclosed its future Claude models will use SynthID-Text, an approach Google created and released as open source. It uses a secret key that subtly changes the process a model uses for choosing the next word in a sentence. Whereas a top next word choice might be “cloudy,” the key might change it to “ove…

AI model watermarking changes agent behavior

Lasso Security sees differences in tool handling and model refusals