Study Finds Invisible Unicode Characters Can Strip Claude's Text Watermark

A researcher built watermark generators and detectors for Claude, OpenAI, and Gemini models and ran roughly 300 attack experiments, finding that inserting invisible Unicode variation selectors at about a 30% rate is the only method that effectively removes Claude's text watermark without damaging the text.

2026-08-17 ~ 2026-08-17 · 3 related posts

1 near-duplicate retellings: Available-Deer1723