Just Security,
No B* S**

The latest security research on building, breaking, and defending agents, systems, and humans. Everything here goes to the mechanism and the details you need, nothing past that.

Latest research

Aug 16, 2026· 16 min read

How Claude's Text Watermark Works (and How It Breaks)

Anthropic shipped a text watermark across all Claude models on August 2, 2026. The technique seeds token-level randomness with a keyed HMAC to embed a statistical signal. Two independent studies found that a single paraphrasing pass destroys the signal more than 90% of the time.

Normal generation

The weather was cold and the streets were quiet as she walked toward the old library.

Watermark signal

Watermarked generation

The weather was gray and the streets were still as she headed toward the aging library.

Watermark signal

watermarkingSynthIDEU AI Act