NewsLLMs respond differently to harmful prompts when AI watermarking is usedMacLook TeamSeptember 17, 20261 min readSynthID can cause models to follow harmful instructions they would otherwise refuse.Read the full article on the original source →