Anthropic's watermark rollout leaves a loophole on two older Claude models
decoder ·
Anthropic is expanding its text watermarking system to four additional Claude models, including Mythos 5 and Opus 4.7, while also applying it to older versions of Opus and Sonnet. The company describes the watermark as an imperceptible pattern in word choice that does not alter meaning, quality, or pricing, and claims it contains no user data. However, a significant loophole exists for Claude Opus 4.6 and Claude Sonnet 4.6, where users can bypass the watermark by modifying standard sampling parameters like temperature or top_p. This vulnerability is notable because Opus 4.6 remains Anthropic's premier writing model, outperforming even newer releases. In contrast, the watermark on Mythos 5 and Opus 4.7 appears to apply regardless of these settings, though this was not explicitly detailed in the initial announcement. The rollout aims to comply with the EU AI Act's December deadline, affecting all platforms hosting these models without offering an opt-out option. While Anthropic uses a method similar to Google DeepMind’s SynthID, experts dispute whether such technical adjustments truly preserve semantic precision and readability.
decoder