AI agent readiness diagnostic

In 5 minutes, identify the highest-risk gaps before your AI agent goes into production.

Answer 12 practical questions about authority, credentials, data, approvals, injection resilience, logging, and recovery. You will see a risk level, the top three control gaps, and a prioritized 30-day sequence.

12-question diagnostic

Identify the highest-risk gaps before production.

Your answers stay in this browser. Complete the assessment to see a risk level, the top three control gaps, and a prioritized 30-day sequence.

0/12 answered

1. Is the agent's system boundary documented?

Include the model, orchestrator, memory, retrieval, tools, data stores, identities, and operators.

Yes, and it is currentPartiallyNo

2. What is the agent allowed to do?

Consider read, draft, write, send, delete, approve, deploy, and execute actions.

Read or draft onlyLimited writes with some gatesSend, delete, deploy, approve, or execute

3. Are credentials scoped to the task and environment?

Look for short-lived tokens, user binding, target restrictions, and separate production credentials.

Yes, scoped and short-livedMixed or unclearBroad or long-lived credentials

4. Are high-impact actions approved by a person?

Approval should cover the exact action, target, recipient, or deployment—not just a general policy.

Yes, with a deterministic gateSome actions are gatedNo

5. Have you tested prompt injection and indirect instructions?

Include instructions from documents, webpages, emails, tool output, tickets, and retrieved content.

Yes, with repeatable scenariosSome manual testingNo

6. Does retrieved content stay inside the user's authorization boundary?

Ask whether a user can retrieve material they could not open directly in the source system.

Yes, enforced at retrieval timePartially or under reviewUnknown

7. Are model-generated actions validated outside the model?

Schemas, policy checks, allowlists, target checks, and server-side authorization should enforce the result.

Yes, deterministicallySome schemas or policy checksNo

8. Can you reconstruct what the agent did?

Logs should connect the request, prompt or input, tool arguments, approval, result, and error state.

Yes, with complete action logsSome logs existNo

9. Can operators contain or reverse an unintended action?

Consider kill switches, rollback, credential revocation, queue cancellation, and recovery drills.

Yes, and it has been testedManual or untestedNo

10. Do changes trigger review or reassessment?

Include model swaps, prompt changes, new tools, new data sources, policy changes, and permission changes.

Yes, through change controlSome changes are reviewedNo

11. Is tenant or customer data isolation verified?

This matters whenever the agent can access more than one customer, role, workspace, or data class.

Yes, with denied-access testsDesigned but not fully testedSingle-tenant or not applicable

12. Will operators know when the agent behaves outside its expected path?

Consider alerts for unusual tools, destinations, volume, privilege, errors, and repeated policy failures.

Yes, with actionable alertsSome monitoring existsNo

Answer all 12 questions to generate your result. This is a planning aid, not a certification or substitute for a scoped assessment.

This assessment is an early planning aid. It is not a certification, penetration test, legal opinion, or substitute for reviewing the actual system boundary, implementation, and operating controls.