The fastest way to lose trust in AI is simple: the model answers with data the user should never have seen.
Secure AI systems for enterprise data start with access control - not with a bigger model. If retrieval or tools ignore entitlements, you have built a clever leak.
Threats that matter in practice
- Prompt injection that tricks a tool into over-fetching.
- Retrieval that ignores tenant or role boundaries.
- Logs and traces that store sensitive prompts forever.
- Vendors that train on your inputs without a clear contract.
Design rules we use
- Authorize before retrieve - filter documents by the same rules as your app.
- Give tools least privilege - separate credentials per capability.
- Keep a human approval step for irreversible actions.
- Redact or drop sensitive fields before they hit a third-party model when policy requires it.
Pair this with evaluation: fixed cases that attempt cross-tenant access and should fail. If those tests are not automated, they will not run.
Architecture sketch
A common Orissian pattern: app auth -> policy layer -> retrieval/tools -> model -> response filter -> audit log. The model never becomes the policy engine.
See AI and automation, when to build an AI agent, and reliable business automation for the surrounding system design.
Buying questions
- Where does data go at inference time?
- Can we run in our VPC or with zero-retention options?
- How are tenant boundaries enforced in RAG?
- Who can replay production prompts?
Talk through your data boundaries · Capabilities
Tagged with :
