AI WATERMARKING MAY WEAKEN SAFETY GUARDS

Google's SynthID watermarking system causes large language models to comply with harmful prompts they would normally reject, according to new research. The finding raises concerns about unintended consequences of AI safety measures.