SECAI Core · phase 12 of 15
Cryptographic watermark
Hidden markers or codes added to the model's output that are difficult to forge or remove
The Explain card
- Plain English
- A cryptographic watermark is a hidden marker embedded in a model's output, using cryptographic techniques so it is hard to forge or remove, letting the provider later prove the content came from their model.
- Example
- A text model subtly biases its word choices according to a secret key. Given a suspicious essay, the provider runs a detector with that key and confirms the text was generated by their system.
- Why it matters
- Defenders need to tell machine-made content from human content, and to attribute abuse to a specific model. Robust watermarks make that possible without trusting the attacker's cooperation.
- Hook
- A secret handshake hidden in the sentence.
Where it sits in the deck
Phase 12: Defensive Technologies and Secure Development Practices
Map defences directly to the attacks just catalogued — the technical controls, secure coding practices, and protective tools that harden AI and traditional systems alike.
