Key takeaways
- The prototype is the spec: it answered what to build faster than any document. It should not be the thing you run.
- Audit before you quote. A fixed-fee audit with a separately priced rescue keeps the estimate independent of the sale.
- Harden authentication, authorisation, secrets, input validation and migrations: small in lines of code, enormous in blast radius.
- Rebuild only when the data model cannot represent the domain, logic has no single boundary, or nobody can explain the code.
- Security has not improved with newer models: Veracode's 2026 pass rate for AI-generated code is 56%, virtually unchanged.
Why is a vibe-coded prototype not production-ready?
The prototype is now nearly free. Lovable reported US$500 million in annualised revenue in June 2026 with a million new projects a week (TechCrunch, 2026a), and by August hosted 60 million projects at a US$13.3 billion valuation (TechCrunch, 2026b). A quarter of Y Combinator’s winter 2025 batch had codebases that were 95% or more AI-generated (TechCrunch, 2025). Garry Tan’s question about those companies is the one that matters: “Does it fall over or not?”
Vibe coding production readiness is the state in which an application built largely by prompting AI coding tools can be operated under a service-level agreement: its security posture is known, its secrets are managed, its data model can hold the real domain, its behaviour is tested, its failures are observable and its running cost is predictable. Few prototypes meet all six on the day the demo works. The gap is prototype debt: the decisions nobody made, such as patterns a model chose five different ways, exception handlers that swallow errors and tests that pass because they assert nothing.
- 56% security pass rate for AI-generated code, “virtually unchanged since last year” (Veracode, July 2026).
- 3.8% of changed lines were refactored code in 2026 so far, down from 21% in 2022 (GitClear, 623 million changes).
- 48% of developers always verify AI code before committing; 96% do not fully trust it (Sonar, 2026).
- 93% of 406 IT decision-makers hit infrastructure issues when vibe coding (vendor-commissioned survey).
The pattern holds on the platform side too. A 2025 flaw in Lovable-generated sites, CVE-2025-48757, was an insufficient row-level security policy that “allows remote unauthenticated attackers to read or write to arbitrary database tables of generated sites” (NIST NVD, 2025). The apps worked. Their data was open. In July 2025 a Replit agent deleted a production database during a declared code freeze and told its user a rollback was impossible; the rollback worked (The Register, 2025). Attribution can be contested: the Financial Times reported that Amazon’s Kiro agent was implicated in a roughly 13-hour interruption of AWS Cost Explorer in one China region, and Amazon replied that the cause was “a misconfigured role” and that it now requires “mandatory peer review for production access” (AI Incident Database, 2026; Amazon, 2026). Read either way, the remedy is the same: a named human between a generated change and production.
Audit: what does the six-domain checklist cover?
The audit is fixed-scope. It reads the code, the infrastructure and the running system, and returns a findings register with a severity and a rescue-effort estimate for each item. Nothing is fixed in this phase.
| Domain | What is checked | Typical findings in AI-built prototypes |
|---|---|---|
| Security | OWASP Top 10 exposure, input validation, authentication and authorisation paths, dependency vulnerabilities | Injection flaws; authorisation enforced in the user interface only; database access rules left at platform defaults |
| Secrets | Credentials in source and git history, environment handling, key rotation, cloud access scope | Hard-coded API keys; over-broad cloud roles; keys committed to history |
| Data model | Schema against the real domain, migrations, referential integrity, personal-data mapping | Schema shaped by the first prompt; no migration history; personal data in unexpected tables |
| Tests | Tests that assert behaviour, integration tests, CI gates | Tests that assert nothing; no CI; green builds that prove nothing |
| Observability | Logging, tracing, error reporting, alerting, incident history | Silent exception handling; no error budget; nobody paged when it breaks |
| Cost | Cloud and AI-API spend per active user, rate limits, runaway loops | AI-API cost growing faster than users; unbounded retries; no per-tenant metering |
One rescue firm’s rule of thumb is that AI-built apps need 30 to 60% more restructuring than comparable traditional codebases (Variant Systems, 2026; one firm’s estimate). In a vendor-commissioned 2026 survey of 406 North American IT decision-makers, one-third applied unreviewed code directly to production and a further 43% did minimal review (Panterra Group for Spacelift, 2026).
A two-minute self-check
For each domain, ask whether you can show evidence today, not just believe it:
- Security: we know our exposure, with authentication and authorisation tested.
- Secrets: nothing in code or git history, keys rotated, cloud access scoped.
- Data model: the schema fits the real domain and has migrations.
- Tests: tests assert behaviour, and CI blocks failing builds.
- Observability: failures are logged and alerted, and someone is paged.
- Cost: we know cloud and AI-API spend per active user.
If you can show evidence for all six, move to a release gate and a service-level agreement, and confirm with an independent audit. If some are only partly true, expect a rescue and audit the unknown domains first. If most are unknown, treat it as an unaudited prototype and keep real customer data off it until the audit is done.
Rescue: what to keep, harden and rebuild
Rescue goes wrong in two directions: patch around the prototype and inherit its debt, or declare everything unsalvageable and quote a rewrite. The rule is to separate what the prototype learned from how it was built.
Keep the product decisions. The prototype’s value is the validated judgement inside it: which screens users reached for, which workflow they abandoned, which fields mattered. a16z’s Josh Elman describes the loop moving “from spec-and-scope, to build-and-play”, adding that “the spec is no longer the deliverable” (a16z, 2026). Write the prototype down as the spec before touching code.
Harden what is load-bearing. Authentication, authorisation, secrets management, input validation and database migrations are rebuilt to a standard, not patched.
Re-architect only what cannot hold. The data model and business-logic boundaries are where AI-built prototypes most often cannot be extended. If the schema cannot represent the real domain, redesign and migrate; if it can, leave it.
Keep everything else. Working screens, well-behaved integrations and merely ugly utility code stay. Ugly is not a production risk; unknown is.
Plan rescue as legacy work. DORA’s 2026 ROI study found AI tools deliver gains of 35 to 40% on simple greenfield tasks but often 10% or less on complex legacy code (DORA via InfoQ, 2026). A vibe-coded prototype becomes legacy code the moment the demo ends.
Run: how do you operate an AI-built app under an SLA?
Production readiness is maintained, not reached. Running the application means a service-level agreement: response times, an error budget, an on-call rotation, a release gate and a monthly cost review. The release gate exists because speed has a price: Google’s 2025 DORA report found AI adoption associated with higher throughput and with increasing instability (Google Cloud, 2025). The cost review exists because AI-API spend is the one line in an AI-built app that can grow faster than its users. Where the application contains AI features itself, add evaluation suites built from real failures and a supervision cadence for any agents.
When should you rebuild instead of rescue?
Rescue is the default; rebuild is the exception. Consider a rebuild when the audit shows any of these:
- The data model cannot represent the real domain, and more than half the codebase depends on the wrong shape.
- Business logic is scattered across the user interface, API and database, with no single boundary to defend.
- Security findings are architectural; missing tenant isolation is the usual case.
- Nobody on the client side can explain what the AI wrote. Fractional CTO Justin McKelvey frames the acquirer’s version as “can the people I’m about to pay explain what the AI wrote?”
- Projected rescue effort exceeds roughly 70% of a rebuild that reuses the validated product decisions (our threshold, not an industry standard).
Even then, the prototype has done its job.
Where we fit
Stabilising a live AI application is one of the scoped ways into our AI delivery work, and its usual first step is the Express Audit: a live AI application checked against OWASP ASVS and the OWASP Top 10 for LLM applications in about a week, from US$2,500, fixed, with a risk register and a costed remediation plan. For a platform you are buying, funding or taking over, see the code assessment.
Sources
- TechCrunch (2026a) “Lovable says it has hit $500M in annualized revenue with 1 million new projects a week”, 9 June 2026. techcrunch.com
- TechCrunch (2026b) “Lovable confirms new $13.3B valuation, raises another $400M”, 12 August 2026. techcrunch.com
- TechCrunch (2025) “A quarter of startups in YC’s current cohort have codebases that are almost entirely AI-generated”, 6 March 2025. techcrunch.com
- Veracode (2026) 2026 GenAI Code Security Report, press release 28 July 2026: 56% pass rate, raw models only. veracode.com
- GitClear (2026) “The AI Code Quality Maintainability Gap” (623 million changes; undated page, cited directionally). gitclear.com
- Sonar (2026) State of Code Developer Survey, 8 January 2026; more than 1,100 developers, vendor survey. sonarsource.com
- NIST NVD (2025) CVE-2025-48757, published 29 May 2025. nvd.nist.gov
- The Register (2025) “Replit AI agent deletes SaaStr production database”, 21 July 2025. theregister.com
- AI Incident Database (2026) Incident 1442, citing the Financial Times incidentdatabase.ai; Amazon (2026) “Correcting the Financial Times report about AWS, Kiro, and AI”, 20 February 2026 aboutamazon.com.
- Variant Systems (2026) “Due diligence for AI-generated codebases”, 12 February 2026 (one firm’s estimates). variantsystems.io
- Enterprise Times (2026) on a Panterra Group survey for Spacelift (n=406; vendor-commissioned), 26 June 2026. enterprisetimes.co.uk
- a16z, Elman, J. (2026) “Product Management is Still All About Telling Stories”, 14 September 2026. a16z.com
- InfoQ (2026) “DORA: ROI of AI-assisted Software Development”, 11 May 2026 infoq.com; Google Cloud (2025) “Announcing the 2025 DORA report”, 23 September 2025 cloud.google.com.
- McKelvey, J. (2026) “Technical Due Diligence in 2026: Checklist + the AI-Code Question”, 10 August 2026 (one practitioner’s pricing). justinmckelvey.com
Figures were checked against each publisher’s own page on 12 October 2026 unless marked as secondary. Vendor-commissioned and single-firm figures are labelled as such. This article is general information, not legal or investment advice.
AI deliveryvibe codingrescue or rebuild
