Anthropic is expanding its text watermarking system to four additional Claude models. According to reporting by The Decoder, this expansion is taking place to meet the regulatory deadline set by the European Union AI Act for December. However, the rollout includes a notable technical limitation affecting two older systems, specifically the Claude Opus 4.6 and Sonnet 4.6 models.
Watermark Bypass on Older Systems
For the Claude Opus 4.6 and Sonnet 4.6 models, developers can bypass the newly implemented text watermarking mechanism by adjusting a single sampling parameter. Anthropic has responded to this configuration loophole by issuing a compliance warning rather than deploying a strict technical fix.
Unresolved Details and Compliance
The announcement did not specify whether Anthropic plans to issue a forced technical patch for the older models or if developers will remain solely responsible for compliance via the existing warning guidelines. It also remains unclear how EU regulators will view this parameter-based workaround as the December deadline approaches.
Developers working with older Claude architectures will need to monitor how compliance requirements are enforced for these specific models.
Lead image: Lee Campbell / Pexels
