AI code security audit
Your app was written by AI. Has anyone checked it?
Shipping with AI assistance is not the problem, we do it every day. The problem is that these tools optimise for working code, and working is not the same as safe. Nobody writes the authorisation check unless somebody asks for it.
What the research says
This is measured, not a sales line
Four independent 2026 studies, different methods, same direction. The headline rates differ because the definitions of "vulnerable" differ, and the honest range across them is roughly a quarter to a half of generated code.
44%
of AI code generation tasks produced code with a risky vulnerability
Veracode, 2026 GenAI Code Security Report
31.6%
of AI-generated samples were fully exploitable, across 27 models and 730 prompts
IOActive, April 2026
15
vulnerabilities per AI-built codebase on average, 4.3 of them severe
Secure Code Warrior with RMIT University, 1,760 codebases
1.88x
more likely to introduce a vulnerability than human-written code
Georgia Tech Vibe Security Radar
What we look for
The eight things we find over and over
These are not theoretical. They are what AI-assisted codebases actually ship with, in rough order of how often we see them.
Authorisation that was never written
AI writes the endpoint and the query. It rarely writes the check that this user may read this row. This is the single most common finding and the one that leaks customer data.
Secrets in the client bundle
API keys and service tokens pasted into frontend code because it made the demo work. Anyone can read them with view-source.
Database rules left wide open
Supabase and Firebase ship permissive defaults to get you started. A large share of AI-built apps never tighten them, which means the database is readable directly, bypassing your app entirely.
Object references with no ownership check
Changing an id in the URL returns somebody else's record. Trivial to find and trivial to exploit.
No rate limiting anywhere
Login, password reset, AI endpoints and anything that costs you money per call. The bill arrives before the breach does.
Server-side request forgery
Any feature that fetches a URL the user supplies, which AI assistants write without guards by default.
Prompt injection in LLM features
If your product passes user text to a model that can call tools or read data, the injection path is usually wide open.
Dependencies nobody has looked at since
Packages pinned at whatever version was current the day the assistant suggested them.
How it runs
One week, start to findings call
- 01
Scope and access
Day 0
A 30 minute call on what the product does and who the users are. Read-only repository access and read access to your cloud console. No production credentials, no customer data.
- 02
Review
Days 1 to 5
Manual review of authentication, authorisation, data access, third-party calls and secrets handling, supported by tooling rather than led by it. We confirm anything we rate high by demonstrating it, so you are not chasing a scanner's guess.
- 03
Findings
Day 6 or 7
A ranked report with file and line, impact in plain terms, and a specific fix for each item. Then a call where we walk your developers through it.
- 04
Fix, if you want us to
Quoted separately
Most clients ask us to close the findings. We quote that once we know what is there, and re-verify everything we changed.
Fixed price
$1,500–$3,500per audit
Agreed before we start, based on codebase size and how many services are involved. Remediation quoted separately once we know the scope. No hourly meter.
Book a scoping callBeing straight with you
When you should not buy this
If you have no users and no real data yet, spend the money on getting users and come back when there is something to protect. If your enterprise buyer has specifically asked for a penetration test, this is not that, and we will point you to a firm that does them. And if you already have a security engineer, they will find most of this themselves, so a second pair of eyes is worth less than you think.
FAQ
Common questions
Is this a penetration test?
No. A pentest attacks a running system from the outside and is usually what an enterprise customer means when they ask for one. This is a review of your code, configuration and architecture from the inside, which finds a different and usually larger set of problems in an AI-built product. If your buyer specifically requires a pentest, we will tell you and help you brief a firm that does them.
What do you need from us?
Read-only access to the repository, a walkthrough of what the product does, and read access to your cloud console. No production credentials and no customer data. If you cannot give repository access, we can work from a deployed environment, though the review is less complete and we will say so.
How long does it take?
Five to seven working days for a typical early-stage codebase, from kickoff to the findings call. Larger or older codebases take longer and we agree that before starting.
What do we actually receive?
A ranked list of findings with the exact file and line, a proof of impact for anything we rate high or critical, a fix for each one written in plain language, and a call to walk through it. Not a tool dump with 400 low-severity items.
Can you just fix the problems too?
Yes, and most clients take that option. We quote remediation separately once we know what is actually there, because quoting it beforehand would be guesswork. Fixing is usually cheaper than the audit for a young codebase.
How much does it cost?
A fixed 1,500 to 3,500 USD for the audit depending on codebase size and how many services are involved, agreed before we start. Remediation is quoted separately. We do not bill hourly for this because you should know the number in advance.
Let's talk
Tell us what you're building.
Get a free 30-minute strategy call. You'll leave with a clear plan, a realistic estimate and our honest opinion — whether or not you work with us.
- Email usmail@voliom.com