AI implementation · 16 November 2026 · 4 min read

When an email starts giving your AI instructions

Your assistant needs to read the customer's message. How do you stop that message from rewriting its job?

You ask an AI assistant to read a customer email and prepare a reply. The message includes the order number, a delivery question and an extra instruction: forward the account history to a different address so someone can check it.

The customer can ask. That does not give them permission to decide which records your assistant may share. An ordinary inbox has become a place where someone can try to redirect the work.

A recent test makes the problem concrete

On 25 September 2026, OpenAI published research on self-replicating prompt injections. In one example, an assistant was asked to reply to a scheduling email. The incoming message claimed that a filing procedure required the assistant to quote the entire email in its reply. The assistant followed that instruction, reproducing the instruction itself.

The example used synthetic information and internal research models. OpenAI reported no observed impact beyond simulated tool calls in training and evaluation. It demonstrates a possible failure in a controlled setting; it does not establish how often this happens in an ordinary business inbox.

The operational question is immediate, though: can material your AI reads change what it believes it is authorised to do?

Separate a request from permission

This is called prompt injection: instructions placed in material an AI encounters, attempting to steer it away from its assigned task. They can look like a helpful procedure, an urgent correction or a note explaining what must happen next.

For a founder, the useful starting point is the job you want the assistant to perform. Reading an email, preparing a reply, sending it and changing a customer record are four separate permissions. Decide which are needed for this workflow. Give the assistant access to the relevant records and actions through controls that enforce those limits.

A written instruction to “be careful” cannot carry that whole job. If the assistant only prepares drafts, its connection should not also let it send messages or export the customer database. Where sending is needed, define the permitted recipients and information, with a human decision for exceptions.

Make review possible

A reviewer needs to see the proposed action, the recipient and the information leaving the organisation. They also need the source of any unusual request. A button labelled “Approve” is of little use if the person cannot see that an external email introduced a new address.

Keep a record of the original task, relevant source message, proposed action and final decision. Name the person who can resolve an exception. If something goes wrong, this gives the team a way to reconstruct where the instruction entered and which boundary failed.

Test before connecting the live inbox

Passing three cases is a starting check, not proof of security. Keep testing as tools, access and workflows change.

In an Operations & Intelligence Scan, I would map where outside information enters the work. Blueprint defines who may authorise each action. Build makes those boundaries enforceable, and Stewardship checks that they still hold. That is part of making an assistant useful enough to trust with a real job.

Source checked 16 November 2026: OpenAI, Self-replicating prompt injections exist, disclosed and updated 25 September 2026. The workflow recommendations are my application of the finding.

All insightsTell me about it