MS Paint and Photos inivisibly watermark even locally generated output with GUID
- ID
- 17380
- Status
- summarized
- Published
- 24 Aug 2026, 11:28 PM
- Fetched
- 25 Aug 2026, 4:07 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://xusheng.dev/posts/reversing/mspaint_invisible_watermark/main/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 25 Aug 2026, 4:08 AM
- Tags
- Audience
- developersai_ml_learners
What happened
Reverse engineering by Xusheng Li reveals that Microsoft Paint and Photos ship local ONNX models for AI image generation, but still send prompts to a remote server for moderation. The server returns a GUID that is invisibly watermarked into the locally generated image pixels, and this behavior is not controlled by the visible-watermark setting.
Why it matters
If you ship or audit AI image tooling, assume 'local generation' does not mean 'local-only' — prompt text is still sent to a remote moderation endpoint and a server-issued GUID is baked into output pixels. This affects privacy claims, provenance/C2PA handling, and any pipeline that assumes locally generated images carry no server-side identifiers.
Discussion angle
What 'local AI' actually means for privacy and provenance: Paint runs models on-device but still phones home for moderation and embeds a server GUID — how should builders communicate this to users?