AI code security review for vibe-coded apps

Your AI built it.
We try to break it.

Adversarial AI code security reviews for software built with Claude Code, Cursor, Lovable, Bolt, Replit and other AI coding tools. We look beyond obvious code smells to find exploitable business logic, permission failures, payment risks, data leaks and cost-abuse paths.

Request a security review →See how it works

Read-only review. No source-code uploads required. Never send us secrets or credentials.

SEC-004 — Read-only user can create a manager
Severity: MEDIUM · Confidence: HIGH

Credible attack path
1. Viewer submits a hidden POST action
2. Endpoint checks membership, not role
3. Viewer invites a manager
4. Privilege boundary is bypassed

Existing controls?
Sibling endpoints block viewers.
This one doesn't.

Required fix:
Enforce capability on every mutation.

Required regression:
Assert DENIED — and prove the invite
side-effect never fired.
Not a linter. We reason about how the system can be abused.No padding. No finding without a credible attack path.Fix verification. We attack the patch, not just the original bug.
What can go wrong?

AI-generated code can work perfectly and still be dangerously wrong.

Vibe coding and AI coding tools are exceptionally good at making features work. Security failures often live somewhere else: in assumptions, permissions, trust boundaries, retries and the way different parts of the application interact.

01 / AUTHORITY

“Read-only” users who can still write

We inspect every mutating path for inconsistent role checks, IDOR, tenant crossover and privilege escalation.

02 / COST

APIs an attacker can turn into your bill

We trace expensive work — LLMs, third-party APIs, email, storage, queues and CPU — and ask how cheaply it can be abused.

03 / TRUST

Public identifiers mistaken for secrets

Anything visible in a browser, URL or client app is treated as attacker-visible. We check what authority it accidentally grants.

04 / PAYMENTS

Replay, webhook and entitlement failures

We examine signatures, idempotency, retry behaviour, server-authoritative pricing and fail-closed billing decisions.

05 / NETWORK

SSRF, redirects and hostile URLs

We follow the complete fetch path: redirects, DNS resolution, private ranges, response bounds and trust of forwarded headers.

06 / FIXES

Security patches that move the hole

After a fix, we look for bypasses, missed sibling endpoints and new vulnerabilities introduced by the patch itself.

The review

We don't stop at “this looks suspicious”.

A useful security finding should tell you what an attacker can do, why your current controls don't stop it, and exactly what has to be proven after the fix.

“If we can't describe a credible attack path, we don't report it as a vulnerability.”
→
Architecture & trust boundariesWhat is valuable, who can reach it, and what the application is trusting.
→
Code & business logicConcrete attack paths through roles, tenants, endpoints, payments, APIs and data.
→
AI-ready remediationEach confirmed issue includes a focused fix brief and the regression test that should lock it down.
→
Fix verificationOnce repaired, we try the original exploit again — then look for a way around the repair.
STEP 01

You send the project details

Public URL and repository link. Private repos are arranged separately with temporary read-only access.

STEP 02

We map the attack surface

Accounts, permissions, payments, tenants, external services, data and costly operations.

STEP 03

You receive ranked findings

Severity, confidence, attack path, impact, existing controls, minimal fix and regression test.

STEP 04

Your AI fixes it. We re-check it.

The patch gets an adversarial pass for bypasses, incomplete fixes and newly introduced risk.

Different by design

No 300-item bullshit vulnerability report.

Automated scanners are useful. They are not the same thing as reasoning about the application an attacker actually sees.

What we don't want to give you

  • Thirty versions of the same theoretical warning
  • “Critical” labels on everything
  • Findings where the control already exists
  • Generic OWASP commentary with no attack path
  • A fix suggestion nobody verifies afterwards

What you should get instead

  • A smaller number of credible, ranked findings
  • Concrete attacker steps and realistic impact
  • Why your current protection misses the path
  • A minimal remediation brief your coding AI can use
  • A regression test and adversarial fix review
Founding reviews

We're opening a small number of early review slots.

For AI-built or heavily AI-assisted applications that are live, approaching launch, or about to start handling real users, payments or customer data. Submit the project first; we'll confirm fit and scope before any payment.

£99founding review · no payment today
Request a review

Tell us what you've built.

Do not submit passwords, API keys, .env files, access tokens or repository credentials. For private repositories, we'll arrange temporary read-only access separately.

Adversarial Security Review

Secure intake powered by Tally
Form not loading? Open the review request form directly →
Questions

Before you hand us the keys. (You don't.)

Do you need my passwords or API keys?

No. Never send us passwords, API keys, access tokens, .env files or other credentials. A security review should not require you to give away production secrets.

What if my repository is private?

Submit the project without credentials. If we accept the review, we'll arrange temporary read-only repository access separately.

Is this just a vulnerability scanner?

No. Automated tooling can be part of a review, but the focus here is adversarial reasoning: authority, business logic, trust boundaries, tenant isolation, payments, replay, cost abuse and how controls interact.

Can I give the findings back to Claude Code or Cursor?

Yes. Confirmed findings are written to be actionable, with a specific remediation direction and a regression test your coding workflow can implement.

What happens after I fix something?

The important part: the fix is reviewed again. We look for bypasses, incomplete patches, missed sibling paths and new risk created by the repair.

Is £99 the final long-term price?

No. This is a founding-review price while we validate the service, scope the work properly and learn what makes these reviews most valuable.