The Model Can Leak Through Its Explanation

A system may refuse the answer and still reveal the restricted fact through its reasoning, citations or confidence.

The answer

Explanations expose source names, hidden categories and inferred relationships. A polite refusal can confirm that sensitive material exists. Test the whole response surface and give restricted users only safe, non-confirming explanations.

Explanations expose source names, hidden categories and inferred relationships.

Where this breaks

A polite refusal can confirm that sensitive material exists.

The operating move

Test the whole response surface and give restricted users only safe, non-confirming explanations.

Value object — The operational card

- Forbidden conclusion

- Permitted partial help

- Citation boundary

- Logging rule

The test

Ask an unauthorised user indirect questions. Neither answer nor explanation should confirm the secret.

Sources

  1. NIST: AI Risk Management Framework — Generative AI ProfileNIST: AI Risk Management Framework

    Primary authority

  2. Swiss FDPIC: AI and data protectionSwiss FDPIC: AI and data protection

    Primary authority

  3. NIST SP 800-207: Zero Trust ArchitectureNIST SP 800-207: Zero Trust Architecture

    Primary authority

Ross BelhommePartner, Svperior / Legal

Jonathan P. De Collibus

Jonathan co-founded Svperior in 2014 and leads its cyber practice. His work sits where adversarial pressure, technical architecture, and consequential decisions meet, with experience across clinical, financial, public-sector, and private-client systems where confidentiality, continuity, and technical correctness carry material consequences.

Cyber strategy / Adversarial assessment / Security architecture / Private systems

Need to apply this to a specific situation?

Send us the initial context. If the matter fits, we will respond directly.

Send private inquiry
The Model Can Leak Through Its Explanation \