An AI Vault is a control layer between users and language models. Before a prompt leaves your environment, sensitive entities are detected and masked — names, emails, IBANs, health data, secrets — so the model never receives clear-text confidential data.
Why classic DLP is not enough
Traditional DLP was built for email and files. LLM prompts are conversational, multi-turn and often pasted from CRM, tickets or contracts. You need:
- entity detection in free text;
- fail-closed behaviour when protection cannot run;
- policy classes for agent actions (read vs send vs deploy);
- rehydration only for authorised users after the model responds.
How Vault-before-model works
- User writes or pastes a prompt.
- Vault scans for sensitive patterns and NER entities.
- Tokens replace clear text (
[PERSON_1],[IBAN_1]). - Only the masked prompt is sent to the model.
- Optional readable view is restored locally under policy.
Semantic benefits for compliance pages
Pages targeting “AI Vault”, “mask PII before ChatGPT” or “LLM data protection” should answer:
- what is masked;
- where processing happens;
- what evidence is kept;
- how this relates to GDPR minimisation and Swiss FADP duties.
Conclusion
If teams use AI without a Vault, sensitive data can leave the company in a single paste. TrustAI puts the Vault before the model so productivity and protection move together.
CTA: See the Vault demo and start a 4-day trial on TrustAI.center.
FAQ
What does an AI Vault mask?
Typically personal data, financial identifiers, health information, secrets and other confidential business content detected in prompts before model exposure.
Is an AI Vault the same as DLP?
Related but not identical. A Vault is specialised for LLM prompts, agent tools and policy tiers around AI actions.
Can the model still be useful after masking?
Yes. Structure and intent remain; only sensitive values are tokenised so the model can reason without clear-text secrets.