Agent Permission Containment Review (QA)

Agent Permission Containment Review (QA)

1. 01

   Inventory every AI agent and automation with write access to client accounts

   List each agent by framework, the credentials it holds, and the systems it can mutate. Include marketing automations, scrapers, and reporting bots, not just security-labeled tooling.

2. 02

   Score each agent against the minimum-access test before approving continued operation

   An agent that reads analytics should not hold publish or delete rights. Downgrade scopes to the narrowest set that still completes the defined task.

3. 03

   Verify a kill switch and rollback path exists for every production agent

   Runtime governance layers such as Vaultak can pause or reverse agent actions without code changes; confirm the mechanism is tested, not just configured.

4. 04

   Run a boundary test in a staging clone of the client environment

   Attempt an out-of-scope action and confirm it is blocked and logged. Document the result with a timestamp for the client file.

5. 05

   Check secret handling across agent configs and repositories

   A scanner like Sentrint flags hardcoded credentials and dependency risks, then emits fix prompts your developers can act on directly.

6. 06

   Log findings in a client-facing containment register with owners and review dates

   Each entry names the agent, the residual risk, the responsible person, and the next review date. This register becomes the evidence trail if a breach is ever questioned.

7. 07

   Re-run the review on a fixed cadence tied to retainer renewal

   Quarterly is the floor for client-facing agents; monthly for anything touching payment or personal data.