The North Sea at night under a low sky, one distant light on the horizon

Field notes Governance

A one-page guide to prompt injection for the board

November 2026  ·  5 min read

This is for directors who do not work in technology and are being asked to approve AI, or are finding out that suppliers already use it. It takes five minutes and ends with four questions.

The risk in one paragraph

An AI assistant cannot reliably tell the difference between the instructions its operator gave it and text it reads while doing the job. So anyone who can put text in front of it can try to give it orders. That text can sit in an email, a support ticket, an invoice, a CV, a shared document or a web page. It is called prompt injection. The attacker does not need to break into anything. They write something, and wait for an AI to read it.

Why it is a supply-chain question

Your own AI tools may be governed well. Your suppliers' tools are the larger exposure. Your managed service provider, your outsourced finance and HR, your law firm and your software vendors are increasingly running AI over your data, with access to your systems. If a stranger's text reaches one of those assistants and it follows the instructions, the damage is done with your supplier's legitimate access. It looks like normal work to your monitoring.

What can happen when it works is concrete. Data is sent out in a reply or a link. A record is changed: a payee, a permission, a closed ticket. A note is planted in the assistant's memory and obeyed in later sessions. A summary or risk rating is wrong in a way that suits the attacker, and someone signs it off.

Why "we have a policy" is not an answer

A policy binds a supplier's staff, not the stranger who wrote the email. A well-behaved model still follows instructions that look like the job. The UK National Cyber Security Centre has said there is "a good chance prompt injection will never be properly mitigated" in the way older attacks were, and OpenAI has said it is "unlikely to ever be fully solved". Detection helps, and every good supplier should use it, but filters miss.

The answer that holds up is engineering. It limits what an assistant may do after it has read something suspicious, puts changes behind a named person and records every attempt. That is checkable, and it is what a board should ask for.

Four questions for the next supplier review

1. Which of our suppliers run AI over our data, and which systems can that AI touch? Most organisations have never been told. The answer is the size of your exposure. Ask for it in writing, supplier by supplier.

2. When their AI reads something hostile, what can it still do? The answer you want is that its authority drops automatically for a period, and that anything that changes your systems needs approval from a person you have named. An answer about training, prompts or model safety is an answer about intent.

3. Do they log injection attempts as incidents, including the ones they stopped, and will they tell us? A stopped attempt is still evidence that someone is targeting your data through them. You should hear about it.

4. Which of these controls are live today, and which are planned? Ask for the status of each in writing. A supplier who cannot separate the two does not know, and you are underwriting the difference.

The board's job is not to understand how injection works. It is to know which suppliers' AI reads your data, and to have their answers to these four questions on file.

What we can show you

We hold ourselves to the same questions, and we can answer all four with evidence. Our platform labels every tool result before an agent reads it. A hostile read cuts the agent's authority for everyone on that account. Every injection attempt is logged as an incident. Every read and action is recorded in a signed evidence ledger that a customer can verify without asking us. In the week to 3 November 2026 the platform labelled 1,912 tool results; 74 looked hostile or could not be judged, the work carried on, and none of them could widen what an agent was allowed to do. Which controls are live and which are planned is set out on Indirect prompt injection is a supply-chain attack, along with eight more detailed questions for your security team. The long version of the argument is The attacker never logs in.

If your board wants a supplier map and the answers to these four questions before its next risk review, we can help you get both.

All field notes

Tell us what you are accountable for.

The useful first conversation is usually about the thing that worries you, not about our capabilities. If it turns out we are the wrong firm for it, we will say so.

One reply from a person who has read what you sent.

Colophon

  • Plate 06, the sea at night: generated, Azure OpenAI gpt-image-1.5, 3 November 2026.

Every plate on this page is a generated stand-in for commissioned photography. The prompt that made each one is embedded in its file. How this site was made.