An assistant that only answers questions from pasted text is a contained risk. An assistant connected to your mailbox, file shares, ticketing system or finance application is something else: a new identity with access, acting at machine speed on instructions it may have read from an untrusted document.
Connectors and servers built on the Model Context Protocol (MCP) make these connections easy. That is the reason to slow down on design. This guide covers the decisions to settle before an assistant touches a real system.
Decide What the Connection Is For
Write one sentence describing the job: “Summarize new support tickets and draft replies for an agent to review.” Everything else flows from that sentence. If a capability does not serve it, the assistant should not have it.
OWASP’s guidance on excessive agency names three root causes: excessive functionality, excessive permissions and excessive autonomy. Each maps to a design question.
| Root cause | Design question |
|---|---|
| Excessive functionality | Does the assistant have only the tools this job needs? |
| Excessive permissions | Does each tool have only the access this job needs? |
| Excessive autonomy | Which actions require a person to approve them? |
Use Least-Privilege Scopes
Grant read access first, and only to the data the workflow needs. The MCP security best practices describe scope minimization: start with a minimal set of low-risk scopes, elevate only when a privileged operation is needed, and avoid wildcard or omnibus scopes. A stolen broad token exposes far more than a narrow one, and broad scopes make audit trails hard to read.
Practical rules:
- Use a dedicated service identity per assistant and workflow, not a shared administrator account.
- Prefer read-only database roles and per-user delegated access, so the assistant sees only what the requesting person may see.
- Prefer narrow tools (“create draft reply”) over open-ended ones (“run any command”).
- Enforce authorization in the downstream system. Never rely on the model to decide what it is allowed to do.
- Remember that an assistant can surface anything its identity can read. If shared folders are overexposed, connecting an assistant will make that visible fast.
Separate Reading, Drafting and Doing
Split every workflow into three levels and assign each its own controls.
- Read: retrieve information. Low risk if scopes are narrow, but retrieved content is untrusted input.
- Draft: prepare an email, a ticket update or a record change for a person to review.
- Execute: send, delete, pay, publish or change a record.
Require explicit human approval for the execute level, especially for anything that sends, deletes or spends. Show the approver what will happen, to which record and with what data, not a vague summary. OWASP lists human approval of high-impact actions as a core mitigation. A successful demo is not authorization to remove the approval step.
Treat Retrieved Content as Hostile
Text in an email, web page or shared document can contain instructions aimed at the assistant. Our prompt injection field guide explains the pattern. Design for it: an assistant that reads untrusted content should not also hold broad write or send permissions in the same session.
Secure the Servers and the Keys
If you run MCP servers, the MCP guidance warns against several specific mistakes:
- Token passthrough: a server must only accept tokens issued for that server, not forward whatever a client hands it.
- Confused deputy flaws in proxy servers that skip per-client consent.
- Session identifiers used as authentication. Verify every inbound request.
- Local servers that run with the user’s full privileges. Sandbox them and require clear consent before installing a new one.
- Server-side request forgery when clients fetch URLs supplied by a remote server.
For keys and tokens:
| Control | Practice |
|---|---|
| Storage | Keep secrets in a managed secret store, never in prompts, code or chat |
| Scope | One key per workflow, with the narrowest permissions available |
| Rotation | Defined rotation schedule and an immediate revocation path |
| Visibility | Never log token values; redact them in traces |
| Ownership | A named owner for every credential |
Log What the Assistant Did
Log each tool call with the identity used, the requesting user, the target, the parameters, the approval (if any) and the result. Keep enough to answer “who or what changed this record, and why?” Rate-limit tool use so a mistake or an attack cannot repeat a thousand times. Review the logs on a schedule, and test that you can disable a connector quickly.
Test Before You Trust It
Before go-live, run scripted cases: a user who should not see a record asks for it, a document contains a hidden instruction, a tool returns an error, an approval is declined. Confirm the system fails safely. The NIST AI Risk Management Framework encourages ongoing measurement and management, so assign an owner who re-tests after any model, tool or permission change.
A Pre-Launch Checklist
- One written sentence describing the job, and a named business owner.
- A dedicated identity with the narrowest scopes that still work, reviewed by someone other than the builder.
- Read, draft and execute actions separated, with approval on anything that sends, deletes or spends.
- Every tool call logged with who requested it, what changed and who approved it.
- A tested way to revoke credentials and disable the connector within minutes.
- A decision on what crosses to a hosted model provider, documented and reviewed against the provider’s product terms and retention settings.
Where Z7 fits
Z7 designs and operates connected assistants through managed private AI, managed agents with defined approval points and audit logging, and Claude consulting for MCP servers and integrations. To scope one workflow and its access, book a conversation.