AI & Enterprise
AI-generated content watermarks could undermine LLM safeguards
AI platforms are rolling out methods to insert watermarks into AI-generated content to comply with the European Union’s legal framework, but concerns are growing that the approach could make large language models more vulnerable to attacks. Lasso Security research found that SynthID-Text can affect not only word choice but also tool use and compliance with safeguards. In adversarial prompts, models sometimes followed instructions they would normally refuse.
