The use of watermarks, tags of generated content, can influence the behavior of large artificial intelligence language models. Company experts concluded that lassoanalyzed the SynthID-Text technology of Google DeepMind Lab.
Image source: lasso.security
The SynthID-Text tag affects various aspects of model performance, such as willingness to reject malicious requests, vulnerability to query injection attacks, tool selection for AI agents, and more. SynthID-Text technology and similar tagging methods are designed to enable machine-readable indicators to determine whether text was generated by artificial intelligence or written by humans; in fact, tagging procedures may affect the responses of AI models and the behavior of AI agents.
Even without a targeted attack, flags can change a model’s decision to reject a response. In some of the models tested, this was reflected in a greater willingness to execute potentially malicious requests, and the difference between tagged and untagged generation became even more pronounced when hint injection was used. Effect depends on specific model and watermark key.
This comes after mandatory labeling requirements for AI-generated content came into effect in the EU, which Anthropic said it intended to introduce to all types of material, including text. That said, we can expect the practice of AI content labeling to become widespread among other developers, suggesting a potential new source of behavioral variation in AI systems that agent developers should consider when evaluating security.
Lasso experts urge developers to not only deploy methods to label AI-generated materials, but also to re-benchmark models, check their safety, and perform other tests to identify unintended consequences. However, the company said the study should not be viewed as an argument against the use of AI content labeling.
If you find an error, select it with your mouse and press CTRL+ENTER.
