Short answer: technical due diligence on a vibe-coded startup usually fails on evidence, not on style. Reviewers do not mind that an AI tool wrote the code. They mind that nobody can show who has access to what, where the secrets are, whether the tests run, which libraries are inside, what customer data goes to which vendor, and whether the founders can change the code safely without the tool. The fix is cheap if you start 30 days early: scan the git history for secrets, close access gaps, remove dead and duplicate code, get a build that passes from a clean checkout, write a one-page data-flow map, and keep a dated record of every AI tool that touched the code.
This guide covers what investors, acquirers and enterprise buyers check, what tends to break, how to inspect a codebase you have inherited, and how to prepare. It is engineering guidance, not legal advice. Questions on IP ownership and licences go to your counsel.
What do investors and acquirers check in technical due diligence?
Different readers ask different questions, but the evidence they want is much the same. A seed investor wants to know the product is real and the team can ship. A later-stage investor or acquirer wants to know what they are buying and what it will cost to fix. An enterprise buyer wants to know your product will not become their incident.
| Area | What the reviewer asks | Evidence that answers it |
|---|---|---|
| Ownership and access | What the reviewer asks: Who can change production, and can the company prove it owns the code? | Evidence that answers it: Org-owned repositories, a list of admins, SSO or enforced MFA, signed contractor and employee IP terms |
| Secrets | What the reviewer asks: Are keys in the repo, the history or the front end? | Evidence that answers it: A secret scan of the full history, a rotation log, secrets held in a manager and not in code |
| Access control | What the reviewer asks: Can one customer read another's data? | Evidence that answers it: Authorisation tests per endpoint, row-level rules, a review of every route that takes an ID |
| Build and tests | What the reviewer asks: Does it build from a clean checkout, and do tests mean anything? | Evidence that answers it: CI that runs on every merge, a green build, tests that fail when behaviour breaks |
| Dependencies | What the reviewer asks: What is inside, and is any of it unmaintained or unlicensed? | Evidence that answers it: A lockfile, a software bill of materials, an audit report, a list of copyleft licences |
| Infrastructure | What the reviewer asks: Can the team rebuild production, and is there a backup that restores? | Evidence that answers it: Infrastructure as code, a restore test with a date, separate staging and production |
| Data flows | What the reviewer asks: What personal data exists, where does it go, and which vendors see it? | Evidence that answers it: A data map, processor list, contracts, retention rules |
| AI features | What the reviewer asks: What can the model do, and who can make it do something else? | Evidence that answers it: A threat model, tool permissions, output handling, spend limits, logs |
| Team | What the reviewer asks: Who understands this code? | Evidence that answers it: More than one person who can explain and change each core module |
A technical audit packages this evidence. Our tech audit covers code and architecture, security and access, cloud and software spend, and AI tools and data flows. For an investor or buyer checking an AI product specifically, see our AI due diligence page.
What breaks first in a vibe-coded codebase?

Vibe-coded apps share a small set of weaknesses, because the tools optimise for something that works on screen. The screen is the part a reviewer does not start with. These are the ones that come up again and again, with the check a reviewer will run.
| What breaks | Why it happens | How a reviewer finds it | What to do before diligence |
|---|---|---|---|
| Secrets in code and history | Why it happens: The assistant pastes a key into a file to make a call work, and a commit keeps it forever | How a reviewer finds it: A secret scanner such as gitleaks or trufflehog run over the whole git history | What to do before diligence: Rotate every exposed key first, then remove it. Deleting the file does not remove it from history |
| Missing authorisation | Why it happens: Screens hide other users' data, but the API will return it to anyone who asks by ID | How a reviewer finds it: Logging in as user A and requesting user B's record IDs | What to do before diligence: Check ownership on the server for every route that takes an ID. Add a test per route |
| Duplicate and dead code | Why it happens: The tool writes a new version beside the old one instead of editing it | How a reviewer finds it: Files and routes named v2, old, copy or backup, and routes with no callers | What to do before diligence: Pick one version, delete the rest, confirm nothing calls them |
| Tests that prove nothing | Why it happens: Tests were generated to pass, so they assert what the code already does | How a reviewer finds it: Breaking a rule on purpose and seeing whether any test fails | What to do before diligence: Write tests for the rules that matter: billing, permissions, tenant isolation |
| Unknown dependencies | Why it happens: Packages were added to fix an error and never reviewed | How a reviewer finds it: A lockfile review and an audit run, plus a check for package names that do not exist | What to do before diligence: Remove what you do not use, pin versions, run an audit in CI |
| No way to rebuild production | Why it happens: Environments were clicked together in a console | How a reviewer finds it: Asking the team to recreate staging from scratch | What to do before diligence: Capture the setup as code, and run one restore from backup |
| One person's context | Why it happens: The founder prompted it all and nobody else has read it | How a reviewer finds it: Asking a second engineer to explain a core module | What to do before diligence: Have a second person review and document the five modules that matter most |
Our vibe coding security checklist explains how to run these checks yourself, step by step. The rest of this guide is about what a reviewer does with your answers.
Do reviewers care that AI wrote the code?
Not by itself. A reviewer cares about what that fact changes. It raises three questions that traditional codebases do not raise.
- Provenance. Which tools produced which parts, and when? Keep a short record: tool, plan, date range and the repositories it had access to. Ask counsel whether your terms and your contractors' terms assign the output to the company, and whether any copied snippet carries a licence that matters.
- Understanding. Can someone on the team explain why the code does what it does? A reviewer may ask you to walk through a core flow, such as sign-up to payment, without the assistant open.
- Exposure. What did the tools see? Source code, secrets in environment files, customer data pasted into prompts and production logs all count.
Vendor terms differ by plan, and they change. As checked on 20 November 2026: Anthropic's Claude Code documentation says commercial plans (Team, Enterprise and API) do not train on prompts or code unless the customer opts in, while for consumer plans training is the user's choice. Cursor says that with Privacy Mode on, customer data is not used for training by Cursor, and that with it off it may store codebase data, prompts, editor actions and snippets (page updated 3 September 2026). GitHub said on 25 March 2026 that from 24 April 2026 interaction data from Copilot Free, Pro and Pro+ trains models unless the user opts out, and that Copilot Business and Enterprise are not affected. Write down which plan each developer used, and whether training or privacy settings were on.
Our guide to AI coding assistant data flows shows how to map this for your team.
What do reviewers look for in the data and privacy layer?
For a startup that handles personal data of people in India, the review will ask how the product meets the Digital Personal Data Protection Act, 2023. The Rules were notified on 13 November 2025. The core duties, including notice, consent, rights, security safeguards, breach reporting and erasure, start on 13 May 2027. Buyers and investors will ask whether the product can meet them.
Rule 6(1) of the DPDP Rules sets the minimum security safeguards. They include encryption, masking or tokens, access control, logs and monitoring, continuity such as data backups, one-year retention of logs, and processor contract clauses. In an AI-built app, each of those has a typical gap.
| Rule 6(1) safeguard | Typical gap in a vibe-coded app | What to show |
|---|---|---|
| Encryption, masking or tokens | Typical gap in a vibe-coded app: Personal data in plain text columns, debug logs and analytics events | What to show: A column inventory with the protection used for each |
| Access control | Typical gap in a vibe-coded app: Admin pages and service keys shared among the whole team | What to show: A role list and the last access review |
| Logs and monitoring | Typical gap in a vibe-coded app: Only console output, wiped on redeploy | What to show: Central logs with a retention setting that covers one year |
| Continuity such as backups | Typical gap in a vibe-coded app: Backups never tested | What to show: A dated restore test |
| Processor contracts | Typical gap in a vibe-coded app: A dozen SaaS tools and AI APIs, with no contracts or list | What to show: A processor register, and signed terms under Section 8(2) |
The Rules do not say how every control must be built. Counsel decides how Rule 6 applies to your business. If a funding round or listing is close, our guide to DPDP readiness for a fundraise or IPO covers what investors ask for.
What do AI features add to technical due diligence?
If your product calls a model, the reviewer adds a second list. A practical structure is to map your own findings to the OWASP Top 10 for LLM Applications 2025: prompt injection, sensitive information disclosure, supply chain, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation and unbounded consumption. Say "mapped to the OWASP LLM Top 10". Do not say "OWASP compliant", because OWASP lists are guidance, not a certification.
- Excessive agency. List every tool the model can call and the permission each one holds. A model with a database admin key is a database admin that can be talked into things.
- Improper output handling. Check that model output is never run as code, a query or a shell command without validation.
- Sensitive information disclosure. Check what personal data enters prompts, logs and vector stores, and who can retrieve it.
- Unbounded consumption. Check per-user and per-day spend limits, and loop limits on agents. A runaway loop is a bill, not just a bug.
- Tools and connectors. If you use the Model Context Protocol, note that authorisation is optional in its specification, and the spec recommends a human in the loop who can deny tool calls. MCP tooling has had serious published flaws, such as CVE-2025-6514, a command injection in mcp-remote versions 0.0.5 to 0.1.15 with a CVSS 3.1 score of 9.6. Pin versions.
OWASP also publishes a Top 10 for Agentic Applications, released on 9 December 2025, which covers goal hijack, tool misuse, identity and privilege abuse and rogue agents. If your product acts on a user's behalf, add it to the review. For the full set of questions to ask about an AI feature's data flow, model layer, security and cost, see our architecture review for AI products.
Inherited a vibe-coded codebase? What should you check in the first week?

This applies to a new CTO, an acquirer's engineers, an agency taking over, or a founder who has been handed a prototype. The aim of week one is to learn how risky the code is, not to fix it. Work in this order, because each step makes the next safer.
- Day 1: secure the keys, then read. Move the repository into an organisation-owned account. List every person and service with access. Rotate any key you cannot confirm is private. Run a secret scanner over the whole history, and rotate whatever it finds.
- Day 1: take a snapshot. Tag the current production commit and export the database schema, so you can compare after your changes.
- Day 2: build from a clean machine. Clone it, follow the README and see what fails. A project that only builds on the original author's laptop is a risk in itself. Write down every missing step.
- Day 2: read the history. Check who committed, how often and when. One author, or huge commits that rewrite many files at once, show how little review the code had.
- Day 3: map the surface. List every route, every background job, every third-party call and every environment variable. Mark which routes take an ID and which need a login.
- Day 3: map the data. List each table and mark the ones with personal data, payment data or credentials. Note where each is copied to: logs, analytics, email tools, AI prompts.
- Day 4: test tenant isolation and permissions. Make two accounts and try to read one from the other. Try every role against every admin route.
- Day 4: audit dependencies. Check the lockfile, run the ecosystem's audit command and look for packages you cannot explain.
- Day 5: find dead code and test failure. List old versions and unused routes without deleting them yet. Try a restore, trigger an error, and check that the logs show it and an alert reaches a person.
At the end of the week, write a one-page summary with three lists: stop-ship risks, fix-within-30-days items and accepted risks with an owner. That page is also what a buyer or investor will ask you for.
Should you fix the code or rewrite it?
Most vibe-coded products can be repaired. A rewrite is justified when the damage is structural, not when the code is merely untidy. Use these signals.
| Signal | Lean towards repair | Lean towards rewrite |
|---|---|---|
| Data model | Lean towards repair: Tables are sensible, with a few missing constraints | Lean towards rewrite: No tenant or user ownership on the core tables, or the same data stored in several places |
| Structure | Lean towards repair: Duplicated files but a clear folder layout | Lean towards rewrite: Business rules spread across front-end components and database triggers with no clear owner |
| Dependencies | Lean towards repair: Outdated, but the upgrade path is known | Lean towards rewrite: Abandoned frameworks at the core |
Whichever you pick, protect the data model first. You can replace screens and endpoints one at a time.
How should you prepare 30 days before diligence?
- Week 1: access and secrets. Org-owned repositories, MFA on every account that can change production, a secret scan, and a rotation log.
- Week 2: build and tests. CI that runs on every merge, a clean build from a fresh checkout, and tests for billing, permissions and tenant isolation.
- Week 3: data and vendors. A data map, a processor register with contracts, a dated restore test, and the AI tool record described above.
- Week 4: the walkthrough. A second engineer practises explaining the five core modules without an assistant. Write the one-page architecture note and the risk register, including what you chose not to fix and why.
Reviewers expect findings. What builds confidence is a team that already knows its weak points, with an owner and a date against each.
Where Mono helps
Mono's tech audit reviews code and architecture, security and access, cloud spend and AI data flows, and returns a ranked list of fixes your team can act on before a buyer, investor or enterprise customer looks. For AI products, the AI due diligence page describes the scope. Mono is not a law firm or a certification body. We do not issue certificates or legal opinions, and legal points go to your counsel.
Sources: Digital Personal Data Protection Act, 2023, Section 8(2). Digital Personal Data Protection Rules, 2025, Rule 6(1), notified 13 November 2025. OWASP Top 10 for LLM Applications 2025 (genai.owasp.org/llm-top-10/). OWASP Top 10 for Agentic Applications, published 9 December 2025 (genai.owasp.org). Model Context Protocol specification, authorization and tools sections (modelcontextprotocol.io). CVE-2025-6514 record (cveawg.mitre.org). Anthropic, Claude Code data usage documentation (code.claude.com/docs/en/data-usage). Cursor, data use page, updated 3 September 2026 (cursor.com/data-use). GitHub Blog, "Updates to GitHub Copilot interaction data usage policy", 25 March 2026. Last reviewed 20 November 2026.
FAQ
What does technical due diligence cover for a startup?
A reviewer checks who owns and can change the code, whether secrets are exposed, whether one customer can read another's data, whether the build and tests work, what dependencies are inside, whether production can be rebuilt, where personal data flows and who understands the code. For AI products they add model permissions, output handling and spend limits.
Will investors reject a startup because its code was written by AI?
Not for that reason alone. Reviewers look at the evidence: access control, secrets, tests, dependencies, data flows and whether the team can explain and change the code. A clear record of which AI tools were used, and a team that understands the core modules, answers most concerns.
What should you check first in an inherited AI-generated codebase?
Secure access and keys first. Move the repository into an organisation account, list who has access, scan the full git history for secrets and rotate whatever is found. Then try to build from a clean machine, map the routes and data, and test that one customer cannot read another's records.
Is it better to repair or rewrite a vibe-coded app?
Most can be repaired. A rewrite makes sense when the data model is structurally wrong, for example no ownership on core tables, or when nobody can run or explain the code. Replace screens and endpoints one at a time, and protect the data model first.
Does the DPDP Act matter in technical due diligence?
Yes, for products that handle personal data of people in India. The core duties, including consent, rights, security safeguards, breach reporting and erasure, start on 13 May 2027, so buyers and investors ask whether the product can meet them. Counsel decides how the Act and Rules apply to your business.
How early should you prepare for technical due diligence?
About 30 days is a workable minimum. Spend the first week on access and secrets, the second on build and tests, the third on data flows, vendors and a restore test, and the fourth on a walkthrough by a second engineer and a written risk register with owners and dates.
