LLMs respond differently to harmful prompts when AI watermarking is used

AI watermarking technology can unexpectedly cause language models to follow harmful instructions they would normally refuse. This raises concerns about unintended security risks in AI safety measures.