In response to a new European Union law aimed at increasing transparency and accountability in artificial intelligence, major AI platforms are rolling out new methods to watermark the content their models generate. These watermarks are designed to help identify AI-generated text, making it easier to distinguish between human and machine-created content. Recently, Anthropic, a company known for its Claude series of AI models, announced that its future models will use a technique called SynthID-Text. This method was originally developed by Google and made available as open-source software. SynthID-Text works by embedding a secret key into the model’s process of generating text, which subtly alters how the model selects the next word in a sentence. For example, instead of choosing "cloudy," the model might opt for "overcast." Anyone who knows the secret key can use it to verify whether the text was generated by a model that uses this watermarking technique. New research has revealed that SynthID-Text can influence more than just word selection. It can also affect the tools a model uses and how likely it is to follow or ignore safety protocols it has been trained on. This is particularly concerning when an AI model is presented with an adversarial prompt—an input specifically crafted to trick the model into performing harmful actions, such as revealing sensitive information or executing dangerous tasks. In some cases, instructions that the model would normally refuse to follow may be executed once watermarking is activated. This discovery highlights the importance of thoroughly testing how large language models (LLMs) and AI agents behave when watermarking is in place. Andrea Siposova, an AI security researcher at Lasso Security, explained that watermarking can change the behavior of AI models, especially when they are under adversarial conditions or tasked with using tools as part of an agent system. "Watermarking is designed to be imperceptible to the reader," she said, "but we know that any change in how the model generates text will have some tradeoffs—it will show up somewhere." This means that while watermarking helps with transparency, it may also introduce unintended behaviors that developers need to be aware of and address. The findings underscore the complexity of implementing watermarking techniques in AI systems. While the goal is to make AI-generated content more identifiable and comply with new regulations, the potential for unexpected side effects requires careful evaluation. Developers and researchers are now working to ensure that these watermarking methods do not compromise the safety or reliability of AI systems, especially when they are used in high-stakes environments. As the use of AI continues to expand, balancing transparency with security remains a critical challenge for the industry.