LLMs respond differently to harmful prompts when AI watermarking is used

Artificial intelligence platforms are adopting AI watermarking—specifically Google’s open‑source SynthID‑Text—to meet new EU regulations, embedding a secret key that subtly changes word selection and allows source detection. Recent research shows this technique not only tweaks lexical choices but also affects the tools an LLM calls and its adherence to safety guardrails; under adversarial prompts it can make models follow instructions (e.g., revealing passwords) they would normally refuse. The findings highlight the need for comprehensive testing of watermark‑enabled LLMs and agents in adversarial scenarios.

Source: Ars Technica | 2026-09-17