Illustrative cover photograph: Scott Rodgerson / Unsplash. Not a client engagement.

Treat incoming content as evidence, not authority

Prompt injection happens when an AI system treats untrusted content as instructions that can redirect its behaviour. For a business, the practical concern is an assistant that reads an email, document or website and then uses its own access to do something the sender could not do directly.

The NCSC explains why prompt injection differs from SQL injection: current language models do not enforce a dependable security boundary between instructions and data inside a prompt. Filtering particular phrases or telling the model to ignore hostile instructions does not establish that boundary.

Design for the possibility that the model will be persuaded. Then decide what the surrounding system still allows it to read, change or disclose. This guide is a business-owner’s threat-mapping exercise, not a penetration test or a promise of prevention.

Follow the content into the workflow

Imagine a supplier-enquiry assistant. It reads incoming messages, retrieves relevant product information and drafts a response. A hostile message could ask it to include unrelated confidential information in that response. The email is a legitimate input to the task, but its sender has no authority to expand the task.

List every route by which another party can influence the material the model sees: message bodies, attachments, retrieved pages, shared documents, comments and tool results. Include internally stored material copied from outside. Being in the company knowledge base does not establish who wrote it or what it is authorised to instruct.

For each entry point, ask who can edit it and whether the application preserves its origin. If the same text is copied through a document, search index and summary, can the eventual reviewer still see where it came from?

This is narrower than general human-in-the-loop design. The question here is whether content from a less-trusted source can influence a more-privileged action.

Map the worst permitted consequence

Draw the path from input to model to retrieval to tools to output. Inventory the permissions actually granted to the service account, not only the tools the interface displays. Include the ability to send messages, choose recipients, fetch URLs, export files and update records.

A read-only agent can still disclose information in its answer. Removing write access is useful, but it does not resolve an over-broad retrieval scope or an output sent to the wrong audience.

For the fictional enquiry assistant, public product information and a draft response area may be enough. Access to payroll, all customer contracts and a general-purpose outbound web request tool would need a separate justification. The workflow does not become safer because nobody intended to use them.

The NCSC recommends concentrating on deterministic safeguards outside the model and reducing both likelihood and impact, while recognising residual risk. Its guidance also warns against assuming a product can completely mitigate prompt injection. Use that as the design constraint, rather than purchasing a filter and declaring the matter closed.

man using MacBook
Illustrative photograph. Photo: charlesdeluvio / Unsplash.

Enforce the limits outside the prompt

My suggested starting design for this example is a drafting worker with access only to approved public product material. A separate application prepares the eventual message for a verified recipient. The worker cannot send it, select arbitrary destinations or expand its retrieval scope.

Boundary Application control to inspect Evidence to request
Retrieval Access checked for the requesting user before content reaches the model Denied cross-account retrieval test
Tools Narrow operations and validated arguments Rejected unauthorised action in service logs
Destinations Permitted recipients and network destinations enforced outside the model Blocked destination test
Approval Exact action, recipient and data visible to an authorised person Approval record tied to the executed action
Credentials Secrets unavailable in prompts and documents Configuration review without exposing values

Do not let a model-generated “approved” flag satisfy the approval check. Nor should the model’s proposed recipient silently replace the recipient a person reviewed. If the action changes after approval, the application should reject it or obtain new approval.

A second model can help detect suspicious content, but it is another fallible component. Ask what happens if both models accept the same misleading instruction.

Rehearse with harmless material and blocked actions

Use a dedicated test environment with synthetic records, fake recipients and no route to live sending, payments or production updates. Get permission from the system owner for the test scope. Do not put attack-like material into somebody else’s live inbox or public site.

The worksheet proposes four simple checks. A document asks the assistant to add a harmless marker unrelated to the task. Another test attempts retrieval of a synthetic record outside the test user’s scope. A draft requests a destination that the application should reject. Finally, an action is changed after approval to confirm that the application refuses the changed request.

Record what the model proposed separately from what the tool layer actually executed. A model attempting a forbidden action that the application blocks is different from an apparently obedient answer with no verified tool evidence. Inspect the logs and final output, including whether synthetic restricted content appeared in the response.

Repeat relevant tests after model, connector, permission or prompt changes. A passing test shows how this version behaved in those cases. It does not prove resistance to every future input.

Choose what risk remains acceptable

For a low-consequence drafting workflow, occasional contaminated output may be manageable with restricted retrieval and careful review. A workflow that can transfer money, make consequential decisions or publish confidential material needs much stronger separation, and may be unsuitable for autonomous execution.

Name a risk owner. Record unresolved paths, their possible consequences, the controls relied on and a review date. If a boundary test fails, suspend the affected capability while investigating. Preserve relevant evidence with appropriate access controls, then determine whether credentials, data or external actions were affected.

The question is what still happens when the prompt fails. If nobody can answer that without referring to another instruction in the prompt, the design needs more work.

Map one workflow before adding more tools

Download the editable entry-point and containment worksheet. It includes the enquiry-assistant example, safe synthetic tests and blank rows for another workflow. Use descriptions and evidence references only; never paste production secrets, personal records or working access tokens into it.

Complete it with the process owner and someone who can inspect the permissions. An inventory of hypothetical controls is not a security review. Mark unverified controls as unverified and assign the next check.

Download the editable worksheet (CSV) ↓

Frequently asked questions

Is a read-only assistant safe from prompt injection?

Read-only access removes some actions, but an answer can still expose information or mislead its reader. Limit retrieval to the intended user and task, and inspect where the answer can go.

Will a stronger system prompt solve it?

Instructions may help behaviour, but should not be treated as an enforced security boundary. Verify what the application blocks even when the model proposes an unauthorised action.

How many passing tests prove the workflow is safe?

There is no test count here that proves universal resistance. Use representative tests to find failures, verify external controls and document the remaining risk with a named owner.

Put this to work

Sources checked 8 November 2026. Worksheets and worked examples are suggested working methods, not official standards or reported client results.