Trivially bypass text watermarks by requesting a Base64 response
Anthropic plans to release text watermarking based on SynthID, which modulates the PNRG behind the token choosing mechanism deterministically in a way that lets another tool later estimate whether text was written by AI.
This is trivial to bypass that, it's current form: 1. Ask for the response/essay/email encoded in base64 2. Decode it using https://www.base64decode.org/
While the encoded text will follow the watermarked distribution, the decoded text will not.
No comments yet.