The Decoder· Matthias Bastian·· 9 hours agoSelectedAI score70
Anthropic Retrofits Text Watermarks to Older Claude Models, but Multiple Models Can Be Bypassed
Anthropic's text watermark is easy to bypass on several older Claude models
AI guide
Anthropic notified API customers that it is adding text watermarking to several older Claude models to comply with EU AI Act requirements, though the watermarks on multiple legacy models only function under default sampling parameters.
Why it matters
The report reveals the mechanism by which watermarks can be bypassed under specific sampling parameters, providing insights into the robustness of compliance solutions in real-world deployments.
Source:The Decoder · the-decoder.com