Services / AI app security review
AI app security reviews
AI coding tools are very good at producing something that works. They are much less reliable at producing something safe to put customer data into.
The problem with shipping what the AI wrote
Tools like Lovable, Bolt, Replit and v0 have made it possible for a business to build a working internal tool or customer app in a weekend, without a developer. That is genuinely useful. The gap is that these tools optimise for something that runs, and security is mostly invisible when something runs.
The same issues turn up repeatedly in AI-built applications, and they are not exotic. They are the basics, missed.
What tends to be wrong
- Secrets in the front end: API keys and tokens shipped to the browser, where anyone can read them.
- Authorisation checked in the interface only: the button is hidden, but the underlying request still works if you call it directly.
- Database rules left open: row-level security never enabled, so one customer can read another customer's records.
- No rate limiting, which makes brute force and cost-run-up attacks trivial.
- Unvalidated input reaching a database or a system command.
- File uploads with no type, size or access control.
- Verbose errors exposing stack traces, queries and configuration to users.
- Abandoned dependencies with known vulnerabilities.
What I review
- Authentication and session handling.
- Authorisation, tested at the request level rather than the button level.
- Database access rules and data isolation between tenants or customers.
- Secrets management and what the front end can see.
- Input handling and output encoding.
- File upload and storage controls.
- Dependency and platform configuration.
- What personal information the app holds, and whether that squares with your privacy obligations.
What you get
A plain-English findings list, ranked by how much damage each issue could do, with the specific fix for each. Written so that you can hand it straight back to the AI tool that built the app, or to a developer, and have the fix made.
Reviews are non-destructive and stay inside a scope agreed in writing before I start. I will need your written confirmation that you own the application or are authorised to have it reviewed.
Assessed against: OWASP Top 10 and common AI-generated code weaknesses
Common questions
Is this a penetration test?
No. A penetration test is an adversarial exercise against a production system, usually with a broader and more intrusive scope. This is a structured review of how the application is built and configured, looking for the failure patterns common in AI-generated code. For most small businesses it finds more, sooner, for less.
Do you need access to the code?
It helps considerably, and most AI platforms make export or repository access straightforward. Where code access is not possible, I can still review the running application's behaviour, within the agreed scope.
What if you find something serious?
I tell you immediately rather than waiting for the report, along with what to do in the short term to reduce exposure while a proper fix is made.
Related services
Start with a free check
I run a short external diagnostic and send you what I find. No obligation, and no pitch if there is nothing worth fixing.
I assess your systems and documents against published standards, identify where the gaps are, and give you the documentation to show you have addressed them. This is not legal advice.