What makes this a 'control blueprint' rather than an AI policy?

Policies describe intent. A control blueprint describes decisions: where AI can run, what it can reach, which assets it can change, who can stop it, and how you recover if the system fails. The playbook is structured around those decisions and the evidence needed to prove them.

Who should read A Control Blueprint for AI?

Boards, audit committees, CISOs, CIOs, chief AI officers, AI programme leads and infrastructure teams. It is written to be read by both the leadership and the operators who will implement the controls.

What is the AI kill switch doctrine?

It is a design pattern for controlled stop states, from throttling an agent, through model withdrawal, to a full hard stop, where the stop path is independent of the system being stopped. Part four of the playbook covers it in detail.

Which research does the playbook draw on?

AISI's Frontier AI Trends Report, NCSC frontier AI guidance, NIST AI RMF, OWASP GenAI and agentic guidance, MITRE ATLAS, RAND's Securing AI Model Weights work and the International AI Safety Report 2026.

Does the playbook recommend specific tooling?

No. It is deliberately vendor-neutral. It sets out the control decisions and evidence a mature AI operating model needs, and leaves tooling choices to your architecture team.

How long is it and how is it delivered?

40 pages, roughly 28 minutes to read. Request it with your work email and a secure link arrives within a minute; access is tied to that address.

A Control Blueprint for AI

A Control Blueprint for AI
2026 Playbook

Control the path. Govern the agent. Protect the asset.

A 40-page 2026 playbook on AI infrastructure, open weights, agents, kill-switch design and operational recovery for boards, CISOs and AI leaders.

40 pages

28 min read

Boards and audit committees

A Control Blueprint for AI: 2026 Playbook cover

Get the playbook

Emailed to you within a minute

01Foreword

Written by the people who build the controls.

AI governance is moving quickly, and the danger is treating it as a paperwork exercise while agents, models and infrastructure become part of the operating fabric of the business. A Control Blueprint for AI is a deliberately different document.

Mark Fermor · Co-Founder, Firevault

The doctrine

Control the path. Govern the agent. Protect the asset.

Every Firevault playbook is grounded in the same doctrine behind our Offline Secure Storage® platform, so a decision taken in one chapter still holds in the next.

It is not a model-security checklist; it is a control blueprint for the people who now have to decide where AI runs, what it can reach, which assets it can change, who can stop it, and how the business recovers if the system fails or is abused.

Grounded in AISI, NCSC, NIST AI RMF, OWASP GenAI, MITRE ATLAS, RAND and the International AI Safety Report 2026, and written by Mark Fermor, it gives boards and CISOs a common vocabulary, Control the Path, Govern the Agent, Protect the Asset, and the operating cadence to run AI control as a discipline, not a policy.

02What's inside

6 parts. Written for the people who own the decision.

Each chapter opens with the decision it exists to help you make, and closes with the evidence you should expect back.

The AI infrastructure reality

Why AI is now critical infrastructure, and how control planes, data planes and recovery planes must be separated across compute, data, identity, facilities and recovery.

AI infrastructure is now a national security issue
One shared path means one shared failure

Attack surface and the AI data centre

Chained attack surface analysis, prompt, identity, tool, API, data, recovery, and the concentration risk sitting inside modern AI clusters.

Attackers rarely need to own the model if they own the route
Model theft is now a data-centre security problem

Open weights and model assets

Protecting weights like source code plus crown jewels: custody, safeguards, release control, hashing, offline archives and destruction records.

Open-weight releases change the risk boundary permanently
Model custody is a lifecycle discipline, not a one-off decision

Agents and action authority

Identity, scope, purpose, expiry, approval gates and revocation for AI agents. Why recommendation is not delegated authority.

Agents need identity and expiry, like people do
Control the tool, not only the prompt

The AI kill switch doctrine

Controlled stop states, agent stop, model withdrawal and the hard stop. Designing the safe landing zone and proving it works.

A kill switch must have levels, not just an off button
The deeper the consequence, the more independent the stop path must be

Policy, adoption and operating cadence

From principles to operations. Five board questions, control outputs, a pilot pattern with a control boundary, and a working operating cadence for AI control.

A pilot is judged on reach, authority, evidence and recovery
AI control is an operating rhythm, not a policy binder

03Who it's for

Read it if you're accountable for the decision.

Boards and audit committees
CISOs, CIOs and chief AI officers
AI programme and platform leads
Data-centre and infrastructure teams

Sectors we hear from

Financial servicesGovernment & defenceCritical national infrastructureTechnology & AIHealthcare & life sciences

05Frequently asked

Questions leaders ask before requesting.

Something else on your mind? Reply to any Firevault email or write to founders@fire-vault.com.

06More Firevault playbooks

Each playbook is written for the people who own the decision, practical, UK-grounded and free to request.

Ready to read it?

Request A Control Blueprint for AI. Access opens on this page straight away, so you can read it here and download the PDF. We email a secure link to the same address as a backup.

Get the playbook Talk to Mark