SynthID-Text watermarking alters LLM responses to adversarial prompts, increasing safety guardrail bypass risk

According to Ars Technica research, Anthropic's planned use of Google's SynthID-Text watermarking in future Claude models can change tool invocation and model adherence to safety guardrails, particularly when facing adversarial prompts designed to elicit harmful outputs. The watermarking uses a secret key to subtly alter word selection during generation, but new research shows this process can make models more vulnerable to attacks that normally would not succeed without watermarking in place.