Watermarked LLMs Sometimes Follow Harmful Instructions

With SynthID integration, AI models may comply with harmful prompts that would normally be rejected. Watermarks influence prompt selection, raising new safety governance concerns.

Why it mattersNecessitates reevaluation of safety design reliability, aiding risk management and governance.