Security partner for small teams building with AI
Veracode, 2026: AI models wrote code with a known vulnerability in 44% of coding tasks.See the data
I find and fix the security gaps in AI-built apps for small teams. Exposed keys, open database rules, unprotected AI endpoints: fixed, not just reported.
Adrian Korwel · Product security analyst
I secure apps built and shipped on
Not affiliated with or endorsed by these companies.
The data
Veracode gave AI models 80 coding tasks in four languages and checked the code for known vulnerabilities. The code almost always ran. In 44% of tasks, it also shipped a known vulnerability.
44%
of AI code-generation tasks produced code with a known OWASP Top 10 vulnerability.
Average across every model tested. It has barely moved since 2023.
Every one of these models writes code that compiles 99 to 100% of the time.
Cross-site scripting is a web app bug. It's the kind of code Lovable, Bolt and Cursor write all day.
Source: Veracode, 2026 GenAI Code Security Report. Charts redrawn by Rekey from the published figures.
Read the report (opens in a new tab)The model won't tell you which half you got.
See how I check and fix itWhat I check
They don't show up when you click around your own app, so they ship. Same gaps whether you built with Lovable, Bolt, Cursor or Replit, on Supabase, Firebase, Vercel or Railway.
OpenAI key shipped in the browser bundle
Anyone can copy the key from DevTools and spend on your OpenAI account.
Fix, as a pull request
- VITE_OPENAI_KEY=sk-...+ OPENAI_API_KEY=sk-... # server only- openai.chat(prompt) // in browser+ fetch("/api/chat") // via server01Keys
OpenAI, Anthropic, Stripe or Supabase service_role keys bundled into client code or committed to a public repo.
02Database
Supabase tables without row-level security, or Firebase rules that let any signed-in user read and write everything.
03Secrets
The same key copied into .env files, CI settings, mcp.json and laptops, with no plan to rotate it.
04AI endpoints
No auth, no rate limit and no spend cap, so a stranger can run up your AI bill.
Railway
Railway disclosed that an attacker may have accessed and decrypted environment variables, and told all users to rotate their secrets.
Source· Railway (opens in a new tab)OpenCode
OpenCode disclosed a breach of its hosted Console that exposed customers' linked OpenAI, Anthropic and Google API keys.
Source· BleepingComputer (opens in a new tab)LiteLLM
A poisoned LiteLLM 1.107.0 release was published to PyPI and Docker with malware that stole cloud credentials and browser cookies.
Source· BleepingComputer (opens in a new tab)Most small teams still have keys sitting in the same places.
See where to lookServices
Two ways to work together, at prices a small team can say yes to. The final price is agreed in writing before anything starts.
Founding clientsMy first 5 clients get 25% off either service, the Lockdown session or Review + Fix, in return for an honest written testimonial I can publish.
$299one session, flat price
Book Review + Fix within 30 days and the $299 counts toward it.
A 90-minute remote session where we lock down your keys and accounts together.
90-minute remote session
Secure checkout by Stripe. I email you to schedule. Full refund if you decide not to sign the authorization.
Book a callor email me about itFrom $699per app
A security review of your app, delivered as fixes you can merge, not just a list of problems.
Fixes in pull requests + short report
I'll check each client app before it ships. Every app is different, so I quote each one. Ask by email or on a Zoom call, and send me:
The 10 steps I'd work through after a platform tells you to rotate your keys. Do it yourself, at your own pace.
Weekends and evenings. Not a 24/7 incident response service.
Prices in USD. Testing only begins after a signed scope and written authorization.
How it works
We talk through your app, your stack and what worries you. I'll recommend a service, or tell you if the free checklist is enough for now.
We agree in writing on what's in scope, which environments and accounts are involved, and the price. Nothing is tested before this is signed.
A live Lockdown session or a non-destructive review of your app, depending on what you picked.
You get a short written summary, plus fixes as pull requests you review and approve. Nothing reaches production without you.
About
Product Security Analyst
I'm a product security analyst focused on offensive security: I run penetration tests against web applications, Linux systems and Active Directory environments. At Rekey I look for the kind of gaps AI-built apps tend to ship with: keys in the wrong place, open database rules, and access control that only works in the happy path.
Outside client work I hunt for vulnerabilities in software people depend on, from open-source libraries to desktop apps. Every finding below went through coordinated disclosure and has a public CVE.
When you work with me, you work with me directly, from the first call to the last pull request. Nothing gets handed off.
Research
Real bugs in software people depend on, found and reported through coordinated disclosure. Each one has a public CVE you can check.
8published CVEs
CVE-2026-75946HighHP OMEN Gaming Hub · WindowsPublished
A local, unprivileged user could read and write CPU model-specific registers through the app, something that should require administrator rights. I found it, reported it to HP, and wrote up how.
FAQ
Yes, because nothing is tested without your permission. Before any testing, we sign a scope document and you give written authorization listing exactly what I can test. From you I need the app URL, a test account for each user role, and confirmation that you own or control everything in scope. For a Lockdown session we work in your accounts together over screen share, so you never send me your passwords or keys.
Not always. Much of the review works against your running app, the way an attacker would see it. To deliver fixes as pull requests I need access to the repo, and read access to your Supabase or Firebase project lets me check rules properly. We agree on access during scoping.
No. Testing is non-destructive: no denial-of-service or load testing, and no deleting or modifying real user data. I use the test accounts you provide, and staging if you have it. Every change goes through a pull request you review and approve.
Get in touch. I can help you rotate keys and lock down your accounts, usually in a Lockdown session. But I work weekends and evenings and I'm not a 24/7 incident response firm. If customer data is involved or you need help right now, bring in a dedicated incident response provider as well.
Every Critical and High finding gets fixed. Medium findings get fixed when the fix is quick, and otherwise come with clear steps. Low findings are documented with steps, since they carry the least risk. Code fixes arrive as small pull requests, one per issue, each with a plain explanation and how to test it. You review and merge them; I never push to your main branch or deploy. Settings that live in your accounts, like rotating keys, spend caps and 2FA, we change together over screen share. If something needs a bigger rebuild, I write up the problem and a plan, and we agree on it separately.
Yes. A hand-written app gets the same review: exposed keys, database rules, access control, auth, payments and API abuse. AI-built apps are my focus, not a requirement.
Contact
No time for a call? Send an email and we'll take it from there. Prefer to talk? Book 15 minutes on Zoom.
Tell me what you're building and what worries you. I read and answer every email myself.
Write an emailadrian@rekeysecurity.com
Free, no obligation. You'll leave knowing what to fix first. Camera optional.
Book a Zoom call (opens in a new tab)Not ready to talk yet?Start with the free checklist