Last year I helped build an AI security starter kit for an enterprise security vendor — good material, aimed at CISOs with security teams, AI-SPM budgets and board decks to fill. Useful if you are a bank. Less useful if you are a twelve-person firm whose entire security function is the operations manager and a strong opinion about passwords.

So I have built the version for everyone else. The AI Security Starter Kit is six documents, free, no email gate, licensed CC BY 4.0 so you can adapt the templates and call them your own.

What's in it

The 90-day roadmap. The spine of the kit. Four phases — Look, Decide, Defend, Detect & drill — that take a 5–250 person firm from "people are quietly using ChatGPT" to governed, defended and rehearsed. Roughly fifteen hours of total effort, and most of the actions are conversations and settings rather than purchases.

The staff cheat sheet. One A4 page for the wall. The three data bins, the three questions before you paste, and the one hard rule about money: any change to payment details gets verified by a call to a number you already had. AI can fake a voice on a video call. It cannot fake the number in your records.

The AI usage policy template. Two pages, fill-in-the-brackets, with an editable Word version. Two pages is a feature, not a limitation — one page of policy that gets read beats forty that don't.

Securing AI agents & MCP. The document I most wanted to exist. A chatbot advises; an agent acts — it sends the email, changes the record, moves the money. That difference is the whole security story, and the guide covers the five failure modes, ten rules for running agents in a small firm, and six hygiene rules for MCP connectors. Plain English throughout, with the NCSC's 2026 agentic AI guidance and the joint NSA/CISA/NCSC MCP guidance distilled underneath it.

The board briefing template. The quarterly two-pager: where we use AI, the top three risks honestly stated, incidents and near-misses, money, and the decisions the meeting exists to take. Also editable.

The self-assessment. Twenty-five questions, five domains, scored 0–2. The test for every answer is could you show an outsider the evidence within five minutes? Score it quarterly; watch the trend, not the number.

Why bother

The numbers behind the kit are not subtle. Survey after survey through 2025 and 2026 lands in the same place: roughly half of employees use AI tools their employer never approved, around half of those have pasted non-public company data into them, and most small firms have no AI policy at all. Meanwhile the deepfake payment fraud that made headlines as an exotic attack on a Hong Kong multinational is now commodity tooling pointed at anyone with a finance inbox.

None of the fixes require a platform. They require an inventory, one page of policy, business-tier accounts with training switched off, two fraud procedures, and a quarterly hour. The kit is that sequence written down, in the order I would do it, with the "who does this and how long it takes" stated on every step.

Take it, use it, mark your own homework with it. And if something in it does not survive contact with your firm, tell me — it is versioned, and it will be reviewed like everything else around here.

The kit lives at peterbassill.com/ai-security-starter-kit.