//Question
How Do We Prevent Internal AI Assistants from Exposing Unauthorized Documents or Customer Data?
Posted on 20th November, 2026

Harry
//Answer
Enforce permissions before data reaches the model, not after it answers. An assistant should only ever see what the person asking is allowed to see. Each control belongs at a specific layer:
Layer | What belongs here | What goes wrong without it |
|---|---|---|
User identity and permissions | SSO, roles, attributes and resource entitlements | The assistant can’t tell who is asking or what they’re entitled to |
Permission-aware retrieval | Filtering documents by the requesting user’s permissions before they enter the model’s context | Restricted content reaches the model, and any clever prompt can surface it |
AI agent and tool execution | Rechecking the user’s permissions before every sensitive action or data lookup | The agent uses its own broad access on behalf of a user who has none |
Runtime inspection and output controls | Detecting risky disclosures, enforcing policies and keeping audit evidence | Mistakes in the layers above go unnoticed, with no record for investigation |
Two rules prevent most leaks:
Retrieve as the user, not as the assistant. The most common failure is an assistant that indexes or searches with a service account that can see everything. Pass the user’s identity through to retrieval and tools, so existing document permissions still apply.
Treat output controls as the last line, not the first. If restricted data reaches the model, output filtering may catch some of it, but the model has already seen it. Permission-aware retrieval is the real control.
Then test the setup the way an attacker would:
Sign in as a low-privilege user and ask for a document you know they can’t access.
Plant an instruction in a shared document asking the assistant to reveal customer records.
Confirm both attempts fail, and that the attempts appear in your audit logs.
Akto can be added as a testing and runtime-control layer on top of these controls. It doesn’t replace your identity provider, document permissions or permission-aware retrieval. Akto Argus red teams internally built assistants for data leakage, prompt injection and privilege escalation. At runtime, Akto’s guardrails detect and can block or redact sensitive data such as personal, health and financial information. They also allow or deny tool calls by agent, tool and resource, and keep an audit trail. For Microsoft 365 Copilot, Akto Atlas adds real-time prompt controls and audit records.
Permissions decide what the assistant can see. Runtime controls confirm it stayed inside them.
Comments