Post: Claude secretly bakes in C2PA manifests into text and images
ExplanationOk2014 · reddit · 2026-09-02
A user discovered that Claude inserts C2PA (Content Credentials) manifests into generated content, acting as obscure watermarks. This affects both text and images, with the images containing a hidden 15-20kb C2PA.org manifest. The user analyzed the embedded data using ChatGPT to understand what information is being stored and raised concerns about this non-obvious watermarking practice.
More from Safety
- One diagram explains the difference between CSRF and XSS attacks — _jaydeepkarale · 2026-10-18
- Grady Booch slams Anthropic for tolerating up to 50 deaths and $1B damages per model — Grady_Booch · 2026-10-18
- Single Neuron Suffices to Bypass Safety Alignment in LLMs, Paper Finds — burny_tech · 2026-10-18
- Polymarket gives OpenAI just 8% odds of a full training pause by end of 2026 — Polymarket · 2026-10-18
- OpenAI engineer over alleged hidden recording device seeks bail, denies Saudi escape plan — Polymarket · 2026-10-18
- Free comic "The Collective" offers a strange retelling of the Hugging Face hack — Flugelwagen · 2026-10-18