Security questionnaires for AI-built apps: a UK guide
Your first large customer likes the product. The pilot went well and the champion inside the company wants to sign. Then procurement sends a long spreadsheet, and the deal stops until you fill it in.
If you built the product with Lovable, Bolt, Cursor or Replit, some of those rows will be hard to answer. This guide explains what the document is, what changed in 2026, which questions an AI-built app tends to fail, and what evidence gets each one through. If the spreadsheet is already in your inbox, start here.
Why the questionnaire arrives before the contract
A large company cannot put its data into your system until its security team agrees. Legal will not finish the contract while that approval is open. So the questionnaire comes first, and it comes from people you have never met.
Their subject is supplier risk: who can reach their data, where it is stored, how it is backed up and what you do after a breach. A questionnaire is the cheapest way for them to find out.
The difficulty for an AI-built app is that every question assumes a process. Someone decided how access works. Someone wrote it down. Someone can show a record that it happened. An AI tool gave you working code and none of the records.
The questionnaires and frameworks in use
You will meet five names. They are different kinds of thing, and founders often mix them up.
SIG. The Standard Information Gathering questionnaire from Shared Assessments. Mitratech, writing on 5 March 2026, describes it as "a unified standard for assessing vendor risk across multiple domains".
CAIQ. The Cloud Security Alliance's questionnaire for cloud services. It is a set of yes or no questions built on the Cloud Controls Matrix. Version 4.1 of that matrix, released on 27 January 2026, has "207 controls across 17 security domains".
SOC 2. A report from an accountant. The AICPA's own page, as read on 16 November 2026, describes SOC as "a suite of service offerings CPAs may provide".
ISO 27001. BSI, as read on 16 November 2026, calls it "the global information security management system (ISMS) standard". It covers how the company manages security, well beyond the code.
Cyber Essentials. The UK scheme. The NCSC page, as read on 16 November 2026, calls it "the minimum standard of cyber security recommended by the Government for organisations of all sizes". It has five technical controls: firewalls, secure configuration, security update management, user access control and malware protection.
Many buyers skip all five and send their own spreadsheet. The questions are much the same.
Questionnaire, attestation, certificate: three different things
A questionnaire is your own statement. You fill it in and nobody checks it at the time. That makes it easy to complete and dangerous to exaggerate, because a buyer can hold you to the answers later.
An attestation is someone else's opinion. SOC 2 is the common one. It comes in two types. The AICPA published a panel discussion on SOC engagements on 6 December 2023. In it Neha Patel, a partner at Weaver and Tidwell, said: "For a type 1, the service auditor is evaluating controls as of a specific date." The same page, describing SOC 1, says a type 2 report also covers the operating effectiveness of the controls.
A certificate is a pass against a standard. For ISO 27001, the BSI page describes "a two-stage audit for your systems and documents" before certification. Cyber Essentials is far lighter. The NCSC page, read on the same day, gives a price "starting at £320 +VAT" and says of the Plus level: "The same protections, but with more rigorous, independent technical testing."
A security audit of your code is none of these three. It tells you what is wrong and what to fix. That is the work that lets you answer a questionnaire truthfully, but it does not put a badge on your website.
What changed in 2026
Cyber Essentials now says in plain words that cloud accounts are in scope. The IASME notice of 12 February 2026 announced the Danzell question set and version 3.3 of the requirements. They apply to "all assessment accounts created after April 26, 2026". Two lines matter to a startup. The first: "Cloud services cannot be excluded from scope." The second makes multi-factor authentication "a mandatory requirement for all cloud services" wherever the service offers it. If your product runs on a hosted database and a hosting platform, those accounts are in scope.
The SIG wrote AI into its standard questions. The same Mitratech article of 5 March 2026 says: "Although no new risk domains were added, the 2026 version explicitly codifies AI governance and operational resilience". It also reports a new mapping to ISO 42001.
The Cloud Security Alliance published an AI questionnaire. Its AI Controls Matrix reached version 1.1 on 14 July 2026, with "247 control objectives across 18 domains". The companion questionnaire "has been updated to 320 questions". These cover AI inside a product. If yours calls a language model, expect some of them.
Buyers are being told to ask how the code was written. A Cloud Security Alliance research note of 6 April 2026 observes that questionnaires "do not yet routinely ask whether applications were AI-generated". It calls updating them "a necessary governance step". The question is on its way.
One platform changed its defaults. The Supabase changelog of 28 April 2026 states: "New tables in the public schema will no longer be exposed to the Data API automatically." New projects could choose that setting from that date, and the changelog scheduled it for all existing projects on 30 October 2026. A Supabase post of 7 January 2026 adds: "For tables created via the dashboard, RLS is enabled by default." RLS is row-level security, the rule that decides which rows each user may read. The same post treats tables created by migrations or outside tools as a separate case. Check what your own project does today. A safer default does not repair tables that already exist.
The questions an AI-built app cannot answer yet
The hard questions fall into two groups.
The first group asks about process. How is code reviewed before it ships? How do you know what third-party packages you depend on? The UK government's Software Security Code of Practice, updated on 15 January 2026, is voluntary and has "14 principles split across 4 themes". Several map straight onto questionnaire rows. Principle 1.1 reads "Follow an established secure development framework." Principle 2.2 reads "Control and log changes to the build environment." Principle 3.2 reads "Implement and publish an effective vulnerability disclosure process." An app that one person prompted into existence has none of these on paper.
The second group asks about the system itself: access rules, secrets, backups, logs. Here the answer is often worse than "we have no document". The control may be missing. We described the usual failures in our guide to vibe coding rescue.
Evidence that satisfies each one
A reviewer accepts two things together: a short written rule, and a record showing the rule was followed. The table lists twelve common rows.
| Typical question | Why an AI-built app fails it | Evidence that passes | Framework reference |
|---|---|---|---|
| Do you follow a secure development lifecycle? | Nothing is written down | A two-page policy and a named owner | Code of Practice 1.1 |
| Is code reviewed before merge? | One person prompts and deploys | Branch protection settings and recent reviewed pull requests | Code of Practice 1.3 covers testing before release |
| How do you restrict access to customer data? | Access rules are missing or copied in until errors stopped | The access rules per table, with a test that one user cannot read another's rows | Cyber Essentials: user access control |
| How are secrets stored and rotated? | Keys sit in client code or the repository | A secrets manager, a scan of the repository history, rotation dates | Buyer's own wording |
| Do you keep an inventory of third-party components? | The tool chose the packages | A generated dependency list and an update schedule | Code of Practice 1.2 |
| When was your last penetration test? | Never | A dated report from an independent tester, with fixes tracked | Buyer's own wording |
| What do you log, and for how long? | Only the platform's default logs exist | A list of logged events, the retention period, a sample entry | Code of Practice 2.2 covers the build environment |
| Are backups taken and restores tested? | Backups were never switched on or never restored | The backup schedule and a dated restore test | Outside the five Cyber Essentials controls |
| Do you have an incident response plan? | No plan and no contact | A one-page plan with names, steps and customer notice terms | Buyer's own wording |
| Which sub-processors handle our data? | Nobody has listed the services in use | A published list with each service's purpose and region | Data protection terms in the contract |
| Is MFA enforced on cloud services? | A single login with a password | Screenshots of enforced MFA on each cloud account | Cyber Essentials requirements v3.3 |
| How can a researcher report a vulnerability? | No route exists | A published disclosure page and a monitored address | Code of Practice 3.2 |
Where you cannot give a row yet, say so and give a date. "Planned, with an owner and a deadline" is an answer a reviewer can accept. A false "yes" is a statement the buyer relied on, and it will surface later.
Answering "was this built with AI?"
Say yes. The commit history shows it to anyone who reads the repository, and a denial that is found out costs you the deal.
Then answer the question behind it, which is what you did about the risk. The same CSA note of 6 April 2026 gives the standard to meet: "Organizations should treat AI-generated code as unverified input, not trusted output." It goes on to name static analysis and dependency scanning as mandatory gates in the build pipeline.
So a good answer has three parts. Name the tools and the parts of the product they wrote. Describe the checks every change passes before release. State who reviewed the sensitive areas by hand: sign-in, access rules, payments and anything that touches personal data.
A sensible order for a UK startup
Fix the system first. Paperwork that describes controls you do not have is the worst position to be in. Close the access and secrets gaps before you write a word.
Then write the short documents. A development policy, an incident plan, a sub-processor list and a disclosure page. Each fits on a page or two. Collect the records as you go.
Get Cyber Essentials. It is the smallest piece of work of the three, and it is the government's recommended minimum. The 2026 rules also force the MFA and cloud-scope work you need anyway.
Leave SOC 2 and ISO 27001 until a contract asks for one. Both are larger undertakings: an accountant's report in one case, a two-stage audit in the other. Ask the buyer which they will accept and whether a completed questionnaire will do for the first year.
Our Production Audit tells you what to fix in the first step. It costs from $4,500 and the report arrives in five working days. Every finding is rated Critical, High, Medium or Low, and the report gives a fixed price against every module. If the report rates no finding critical or high, we refund the fee in full. It is a written report. It does not produce a SOC 2 report, an ISO 27001 certificate or a Cyber Essentials certificate.
[See what the audit covers](/services/production-audit)