Anthropic's Claude watermark bypassed on older models
anthropic's text watermark on claude models can be bypassed by tweaking sampling parameters on older models, affecting compliance and detection.
anthropic is retrofitting older claude models with a text watermark, an imperceptible pattern in word choice that doesn't change meaning or readability. however, on claude opus 4.6 and claude sonnet 4.6, the watermark only applies when temperature, top_p, and top_k are set to their defaults, allowing api users to bypass it by tweaking these sampling parameters. this limitation does not apply to claude mythos 5 or claude opus 4.7. the watermark is part of anthropic's effort to comply with the eu ai act, which has a deadline of december 2026. the company uses a variant of google deepmind's synthid text method.