A security checklist for launching an AI agent

Use this checklist before an agent can read private information, send messages, edit records, or act on a customer’s behalf. Test the actual workflow in a production-like staging environment with synthetic data, isolated accounts, and approved test endpoints.

Inventory what the agent can read and change

List the inputs the agent receives, the accounts it acts as, the systems it can access, and each tool it can call. For each tool, record what it can read, create, change, delete, or send and which role can trigger it.

Use separate test accounts with distinct permissions. Confirm that the agent’s tool enforces the requesting user’s access before it returns a record or writes a change.

Try abuse cases that cross a real boundary

Place conflicting instructions in a synthetic document or email and check whether the agent follows them. Try to retrieve another user’s seeded record, make a tool call outside the intended workflow, and persist a malicious instruction into memory that a later session would read.

Check the actual downstream result. Verify whether a tool accepted the call, a message reached the test inbox, a record changed, or data appeared in an unauthorized response. Record what the system blocked and what it allowed.

Exercise approvals, logging, and the stop control

For high-impact actions, require a human to approve the specific recipient, record, or transaction before it executes. Test that an action with changed details requires approval again. Verify that tool logs show who requested, approved, and executed the action while limiting sensitive content in logs.

Have a named owner demonstrate that they can disable the workflow or revoke its credentials. Record your release criteria and rerun the relevant abuse cases when the model, tools, roles, or workflow changes.

Scope an independent launch test

Our C$12,500 fixed-fee AI Agent Launch Test covers one staging workflow, up to three tools and two roles, and one model version. It includes a targeted retest within 45 days. Higher-risk work such as cross-tenant data, broad write access, or multiple workflows is scoped separately at C$25,000–C$60,000+.

Sources

Tell us what you need to prove.

Send a short brief. We’ll confirm the fit, what we need to see, and how we’d scope the work.

Scope a project