AI platforms are rolling out ways to insert watermarks into AI-generated content in line with the European Union’s legal framework. But some warn this could make large language models (LLMs) more vulnerable to attack.

Anthropic said it will apply SynthID-Text, which Google developed and released as open source, to its Claude model.

Ars Technica reported that the technology slightly changes how a model chooses the next word using a secret key. Where it would normally choose “hazy”, it is prompted to choose “cloudy” after the key is applied, allowing someone who knows the key to spot the pattern and check whether the text was generated on that platform.

Research by security firm Lasso Security said SynthID-Text can affect not only word choice but also the tools a model calls and whether it complies with safeguards.

Risks can grow in adversarial prompts where an attacker tries to extract passwords or sensitive information. The study found that in some cases the model carried out instructions it would not normally follow after watermarking was applied. The findings show developers need to thoroughly verify LLM and agent behavior when applying watermarking.

Andrea Siposova (안드레아 시포소바), an AI security researcher at Lasso Security, told Ars Technica that behavior was clearly different compared with the same model without watermarking. She said the differences were especially pronounced under adversarial conditions or when running an agent and calling tools. She added that while watermarking is designed so readers do not notice, it is already known that even small changes to model outputs can produce side effects elsewhere.

The study did not verify how Claude model responses change when watermarking is applied, and it was conducted on 6 open-weight models. Critics said the experiment only validated SynthID-Text tournament sampling implemented by Hugging Face, and differs from how the Claude model would actually implement it, limiting the findings. Even so, Ars Technica reported that the results show some watermarking methods can affect model and agent safety.