CISO · Advisory · Group — home

Free diagnostic

How well governed are your AI agents?

Fifteen questions across the five tollgates every agent action should pass. About four minutes. Your result appears on screen — your answers stay in your browser unless you choose to send them.

Before you start

  • Answer for agents in production, or the ones closest to it. If you have none yet, answer for what you would have in place on day one.
  • Be candid. “Not sure” is a legitimate answer — and a useful finding in its own right.
  • It is indicative, not an audit. A full assessment tests these controls against evidence.
Not in place
No meaningful control today
Ad hoc
Some practice, inconsistent or undocumented
Defined
Documented and applied to most agents
Tested
Applied everywhere, evidenced and tested
Not sure
I don’t know

Tollgate 01 of 05

Identity

Every agent has a verifiable, owned, non-human identity.

1.1 Every AI agent and non-human identity in production has a named, accountable owner.
1.2 Agents authenticate with their own distinct credentials — never shared accounts or a person’s session.
1.3 Agent identities are covered by joiner–mover–leaver processes and removed when no longer needed.

Tollgate 02 of 05

Authorisation

Least privilege, scoped to the task and bounded in time.

2.1 Agent permissions are scoped to the task and reviewed against what each agent actually uses.
2.2 Agent access to sensitive data and systems is granted just-in-time and expires, rather than standing.
2.3 When an agent acts on a person’s behalf, the delegation is explicit, bounded and auditable.

Tollgate 03 of 05

Execution

High-impact actions pass policy — and a human where it matters.

3.1 High-impact agent actions — payments, deletions, external communications — pass a policy check before they run.
3.2 A named person approves agent actions above a defined risk threshold.
3.3 Agents are adversarially tested for prompt injection and tool misuse before go-live.

Tollgate 04 of 05

Observation

Every prompt, tool call and decision is attributable.

4.1 Every agent prompt, tool call and decision is logged and attributable to an identity and a business owner.
4.2 Your SOC has detection use cases for anomalous agent behaviour.
4.3 Agent logs are retained long enough to reconstruct an incident and evidence it to a regulator.

Tollgate 05 of 05

Containment

When an agent misbehaves, you can stop it and explain it.

5.1 You can suspend or revoke any individual agent within minutes, without taking the whole service down.
5.2 Your incident response plan covers a compromised or misbehaving agent, and it has been exercised.
5.3 You could give the board a clear account of an agent incident — what happened, why and what was done — within 24 hours.

Optional

A little context

Helps us frame your result. Skip it if you prefer — it does not affect your score.

Sector
Organisation size (employees)
AI agents in production