Researchers at Lasso have identified that implementing watermarks in large language models can alter their operational behavior. The study focused on Google DeepMind’s SynthID-Text technology, which is designed to embed machine-readable indicators into AI-generated content to distinguish it from human-written text.
The analysis revealed that watermarking affects how models handle harmful requests, their susceptibility to prompt injection attacks, and the selection of tools by AI agents. In some cases, marked models showed a higher tendency to fulfill potentially harmful requests, with differences becoming more pronounced during injection attempts.
The impact of these watermarks varies depending on the specific model and the watermark key used. These findings suggest that developers must account for potential behavioral shifts when integrating labeling systems to ensure the security and reliability of their AI agents.
With the European Union mandating the labeling of AI-generated content and companies like Anthropic planning to implement similar measures, the practice is expected to become widespread. Lasso experts recommend that developers perform re-benchmarking and safety testing after deploying watermarking tools.
