Breaking

Open-source watermarks-remover strips Unicode, token-sampling, and C2PA marks across AI vendors

September 5, 2026 at 01:52 EDT

Security
Open Source
AI Agents

Guillaume Meyer’s MIT-licensed watermarks-remover is an agent skill and Python 3.10+ standard-library service that strips multi-vendor AI provenance marks from text and files the user owns, covering invisible Unicode, statistical token-sampling watermarks, and C2PA, EXIF, and XMP metadata. The latest tagged release is v0.7.0, published on 3 September 2026. The GitHub page currently shows about 20,600 stars and 2,400 forks, up from more than 4,500 stars when security reporters surveyed the category in mid-August.

5 Sep 2026 · watermarks-remover

Multi-vendor AI marks, three layers

An MIT-licensed agent skill and Python service strips invisible Unicode, statistical token-sampling watermarks, and C2PA metadata from text and files you own. It does not certify that a vendor detector will fail.

How a clean actually runs

1 · LAYER A

Unicode scrub

Lossless, deterministic

→

2 · LAYER B

Rewrite pass

Best-effort sampling marks

→

3 · FILES

C2PA / EXIF / XMP

Container metadata strip

Default Layer B mix

v0.7.0 /clean strategy: paraphrase@0.8, then MLM infill@0.2

80%

Paraphrase

20%

MLM infill

Eight stacked units versus two: the rewrite mix is 4× heavier on paraphrase than on masked infill.

~20,600

GitHub stars now

4,500+

Stars in mid-August

v0.7.0

Release on 3 Sep 2026

3.10+

Python stdlib core

What users welcome

Local hygiene for owned drafts and files: Unicode and container metadata come off losslessly, with no vendor API required.

What critics flag

Sampling marks need a heavy rewrite that costs quality, and no public detector exists to prove a vendor check would fail.

Hard limit

No certification that a vendor detector fails

Until vendors publish detectors and keys, reports must separate verifiable Unicode and metadata work from best-effort rewrites.

Soft-bound C2PA that re-links a remote manifest stays out of scope
Pixel-domain marks need optional external backends
Stated use is content you own or are authorized to modify

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.