Skip to content
PRTOTYPE.COM · Production studioAudit · Build · Maintain
16 November 2026 · Field notes · By Jason Ensor

Technical due diligence on an AI-built MVP: what fails

An investor has said yes in principle. Then a message arrives from someone you have never met, asking for access to the repository. That person is the technical reviewer, and their report shapes the rest of the round.

If you built your product with Lovable, Bolt, Cursor or Replit, this article is for you. It covers what the reviewer checks, where AI-built code usually falls down, and what to hand over before anyone asks. It is written for a UK founder, so it uses UK guidance where some exists.

One limit first. Public evidence on how investors treat AI-built startups is thin. We found no reliable figure for any discount they apply, so this article does not give one.

What a technical reviewer is trying to price

The reviewer has one job: to estimate what it will cost the investor to own this product after the money lands.

Four questions sit underneath that estimate.

  • Can the product lose customer data tomorrow?
  • How much work stands between this codebase and the next stage of growth?
  • Does the company own what it is selling?
  • Can anyone other than the founder change it safely?

Each answer ends up in the deal, as a cost or as a condition. A weakness you disclose with a plan is a cost. A weakness the reviewer finds alone is a cost plus a doubt about everything else you told them.

The six areas they open first

Reviewers differ, but the same six areas come up early because they are quick to test and expensive when wrong.

  • Data access rules. Who can read and write each table, and where that rule is enforced.
  • Secrets. Where the keys live, what each key can do, and whether one has ever been committed or shipped to the browser.
  • Sign-in. How sessions expire, how a password reset works, and whether an admin route checks the role on the server.
  • Dependencies. Which packages are in the build and which have known vulnerabilities.
  • Tests. What proves that a change did not break something else.
  • Deploy, rollback and backups. How code reaches production, how you undo a bad release, and when you last restored a backup.

None of this is special to AI. The UK government's voluntary Software Security Code of Practice, last updated on 15 January 2026, sets out 14 principles across 4 themes, and several of them read like a reviewer's checklist. One says: "Have a clear process for testing software and software updates before distribution." Another: "Control and log changes to the build environment." A reviewer may never name the code, but the questions are the same.

What usually fails in an AI-built codebase

The research on AI-generated code is consistent about one thing. The code runs far more often than it is safe.

Veracode tests AI models on a set of 80 coding tasks. In its Spring 2026 update, published on 24 March 2026, it found that "only 55% of generation tasks result in secure code". Python did best at 62%. Java did worst at 29%. Four months later, SD Times reported on 28 July 2026 that in Veracode's 2026 GenAI Code Security Report "the average security pass rate sits at 56 percent". Newer models have barely moved the number.

An older study points the same way. CodeRabbit looked at 470 open-source pull requests and reported on 17 December 2025 that "AI-authored changes produced 10.83 issues per PR, compared to 6.45 for human-only PRs." Read that with care. The study inferred AI authorship from signals in each pull request and assumed the rest were written by people.

Both companies sell tools that check code, so they have an interest in the result. The incident record below agrees with them all the same.

In an MVP, the failures tend to land in the six areas above. The access rule exists in the interface and nowhere else. A privileged key sits in the JavaScript that every visitor downloads. There are no tests, because the tool was never asked for any. And the founder cannot explain why a given file works, which makes every later change a gamble.

A reviewer's starting position is the one the Cloud Security Alliance set out in a research note on 6 April 2026: "Organizations should treat AI-generated code as unverified input, not trusted output." Your task before the review is to do that verification yourself and keep the proof.

Three 2026 cases and the check that would have caught each

A Lovable-hosted app, February 2026. The Register reported on 27 February 2026 that a researcher found 16 vulnerabilities in one app hosted on Lovable, six of which he rated critical. The article puts the exposure at 18,697 user records and names Supabase's row-level security as the kind of control that goes missing in generated code. These are the researcher's claims as reported. The check that catches this takes minutes: request every table as a logged-out visitor, then as a second user, and confirm that both are refused.

Moltbook, February 2026. Wiz published its own research on 2 February 2026. It found "a Supabase API key exposed in client-side JavaScript", and that key allowed "full read and write access to all platform data". The exposure included "1.5 million API authentication tokens, 35,000 email addresses". Wiz notes that a public key is safe to expose when row-level security is properly configured. So the check has two steps. Read the browser bundle for keys. Then test what each key can do.

The CVE count, March 2026. Infosecurity Magazine reported on 26 March 2026 on Georgia Tech's Vibe Security Radar, which tracks published vulnerabilities introduced by AI coding tools. It logged six in January 2026, 15 in February and at least 35 in March, with 74 confirmed cases in total. The method flags a commit only when it carries an AI tool's signature, such as a co-author tag, so unlabelled AI code is not counted. This third case is a trend, and the check is routine: scan your dependencies on every build, and have a person review AI-written changes to anything that touches sign-in, data access or secrets.

Who owns the code: the UK position

This section is not legal advice. Ask a solicitor before you sign a warranty about intellectual property.

A reviewer will ask whether the company owns the code it is selling. For AI-built software, UK policy is in motion. The government's Report on Copyright and Artificial Intelligence, published on 18 March 2026, draws a line between two kinds of work.

For works made with AI help, the report says "copyright should continue to protect works created with AI assistance." For works generated wholly by a computer, it says "we propose that this specific type of protection should be removed", and qualifies that with the words "in the absence of evidence of its ongoing value."

That is a proposal in a report. The report did not change the law.

The practical point for a founder is the difference between assisted and wholly generated. Keep a record that shows people directing and reviewing the work: a written specification, a commit history with human review, and notes on design decisions. Keep the contracts of everyone who contributed. Keep a note of which AI tools you used, on which plan, and what their terms say about the output. A reviewer who sees that record has little left to ask.

Where the NCSC draws the line

The National Cyber Security Centre published The 'vibe coding spectrum' approach to AI-assisted software development on 18 June 2026. It is the closest thing a UK founder has to an official position, and it is more permissive than many expect.

For a proof-of-concept or a limited-risk internal tool, it says full vibe coding "can often be perfectly fine". It then lists when to "slide left (towards manual)". The list includes "building authentication or authorisation logic", "processing sensitive personal data" and "handling secrets, tokens, or credentials".

The NCSC still allows AI for that code, and says AI can help to write authentication logic. It asks you to review what the AI produces, understand it, check it for vulnerabilities and verify that it does what you expect. In its words: "The risk is not applying the right safeguards when the stakes are high." On whether models will soon make this unnecessary, the piece ends: "But we're not there yet."

This gives you a frame a reviewer will accept. Say which parts of the product were built at speed, which parts handle sign-in, personal data or secrets, and what extra review those parts received.

The evidence pack to prepare

Send this before the reviewer asks. Ten areas, one folder.

AreaWhat the reviewer asksEvidence that answers itCommon failure in AI-built apps
Data access rulesWho can read and write each table?The policy for every table, and a test run showing a logged-out visitor and a second user being refusedRules enforced in the browser only, or row-level security missing
SecretsWhere do keys live, and has one been exposed?A list of every key and its scope, a scan of the git history and the browser bundle, and rotation datesA privileged key shipped in client JavaScript
Sign-inHow do sessions, password resets and admin routes behave?A description of the sign-in flow, with test results for expiry, reset and role checksAdmin pages that check the role in the interface and never on the server
DependenciesWhat is in the build, and what is known to be vulnerable?A lockfile, a dependency inventory and the latest scan outputPackages the tool added that nobody chose or has updated
TestsWhat proves a change did not break something?The test suite, its last run and a plain statement of what it coversNo tests at all
Deploy and rollbackHow does code reach production, and how do you undo it?A written deploy procedure, a staging environment and one recorded rollbackOne person publishing from the builder's own button
Backups and restoreWhen did you last restore from a backup?The backup schedule and a dated restore testBackups assumed from the platform and never restored
Ownership of the codeDoes the company own what it sells?Contracts for every contributor, the tool terms, and a record of human direction and reviewNo record of who or what wrote which part
Tool plan and privacy settingsWhich AI tools saw the code and the customer data, and on what terms?Tool names, plan tier and the privacy settings in forcePersonal accounts with default settings
Reliance on one personWho else can change this system?Architecture notes, and a second person who has deployed itOnly the founder can get a change through

A gap in the pack is acceptable if you name it. An empty row with a date against it reads better than a row you hoped nobody would open.

What to fix before the review, and what to disclose

Fix first whatever exposes customer data today. That means the access rules on every table and any key that has reached the browser or the git history. Rotate an exposed key; deleting the line that contains it does nothing for the copies already made.

Next, fix what a reviewer can find in an afternoon: missing backups, no rollback, no tests on the sign-in flow.

Everything else can go on a dated list. Our own reports rate each finding as critical, high, medium or low. Critical means customer data, money or the whole service is exposed now, and nothing else has to fail first. A founder who arrives with the critical and high items closed and the rest scheduled is in a strong position.

Then disclose. Say that the product was built with AI tools and name them. Say which parts a person has reviewed line by line and which parts nobody has. Reviewers find out anyway, and the record of what you checked is worth more than silence.

Avoid one thing: a rewrite in the fortnight before the review. New code has no history in production, and the reviewer will see the commit dates.

If you want a written report on the code before the investor's reviewer opens it, see our Production Audit. It costs from $4,500. You get a written report in five working days, with a fixed price against every module. If the report rates no finding critical or high, we refund the fee in full. The page on preparing for an investor's technical review lists what the reviewer will check, and there is a sample report to read first. When you are ready, book the audit.

Related reading: what it takes to get an AI-built app to production.

Production audit

Show us what’s creaking.

A senior review of your vibe-coded app or stalled pilot: security, data model, infrastructure, compliance. Report in five working days, no judgement.

Report in five working days · from $4,500 · The specification is yours to keep
Book the audit