Whitepaper / Security & quality
Before You Ship AI to Production
A security and quality checklist for enterprise leaders: the pre-production checks that keep an AI feature from becoming your next incident. Free, versioned, and grounded in real engagement work.
The feature is new. The exposure is too.
For a decade, application security came down to a few well-understood questions. Can a user reach data that isn’t theirs? Can input become code? Can a session be forged? Your test plans learned to ask those well. An AI feature keeps every one of them and adds a category none of them cover: the instruction. When a product routes user text into a model, that text is no longer only data to validate. It is partly a program the model may choose to run, and a test plan written for forms and endpoints walks straight past it.
The stakes scale with what you have wired the model to. The OWASP Top 10 for Large Language Model Applications (2025) ranks prompt injection first, ahead of every classic web risk, because the mitigations that contained SQL injection do not transfer. And the cost of getting it wrong has not moved in your favor: IBM’s Cost of a Data Breach Report 2025 put the average United States breach at 10.22 million dollars. This whitepaper is a small, specific set of checks to add to the gate you already run before a release ships.
What’s inside
Why AI changes your risk surface
A model in production is business logic a user can talk to. The paper explains, in plain terms, why the injection point is no longer a field to fuzz but the meaning of the words, and why your last penetration test never saw it.
A real finding, walked through end to end
The anatomy of a prompt-injection flaw in an AI text-rewrite endpoint from an engagement we ran: how a user-controlled parameter reached the model prompt, how it was demonstrated rather than asserted, and how the fix was verified clean on retest.
Ten checks you can run this week
Each check states what to verify and gives a concrete pass-or-fail test you can put to your own team, across security, quality, and governance. No platform required.
A one-page summary, built to print
The full checklist on a single page with the pass condition for each item, so it can go straight into your release gate.
The ten checks
The full paper gives each check a rationale and a concrete pass-or-fail test. Here are the headers, so you know what you’re getting.
- 01Treat every AI endpoint as its own attack surfaceSecurity
- 02Test for prompt injection before launch, by demonstrationSecurity
- 03Draw the data-in-training and data-in-context boundaryGovernance
- 04Keep a human in the loop for consequential actionsGovernance
- 05Validate and constrain what the model outputsSecurity
- 06Enforce authorization at the AI boundary, not around itSecurity
- 07Rate-limit and abuse-test the AI endpointsSecurity
- 08Map the AI supply chain: models, tools, and pluginsGovernance
- 09Log the model's decisions so you can answer for themQuality
- 10Re-scope your penetration test to include AI, then retest the fixesQuality
Who it’s for
CTOs, CISOs, and Heads of Product who are about to put a model in front of a customer, a document, or a decision. If your team is adding an AI feature to an existing product, this is the gate to run before it goes live. It reads in fifteen minutes and reasons from first principles, so it stays useful whether your model rewrites a sentence or approves a refund.
Where an outside eye helps
You can run most of this checklist with your own team, and you should. Where an independent partner earns its place is on the checks that need someone to actively try to break the feature, and on the discipline of demonstrating and retesting rather than scoring. We don’t build the software we test and we don’t resell the tools we recommend, so a finding is shaped by the evidence and nothing else. Our penetration testing treats AI endpoints as their own attack surface and includes retest as standard, our approach to AI is set out on the responsible AI page, and the full anonymized write-up of the engagement behind this paper is a published case study. On organizational certification we are direct: we hold no organizational certification today and do not claim otherwise; ISO 27001 certification for VirtuesTech is planned, stated July 2026.
Adding an AI feature and want a second set of eyes?
Bring the endpoints and the release date. We scope the test with the engineers who would run it, AI surface included.