Private previewInvite-only while we complete a program review. Request access
autonomous application security

Proof first.
Then the fix.

Pwnly takes your GitHub repository through four layers of analysis, reproduces access-control and payment bugs in an isolated sandbox with fake users, and opens a fix that builds and no longer lets the exploit through.

Works today with Node.js and Next.js apps (Supabase-style auth) · read-only repository access

L1-L4
layers
5
fake roles
0
live requests
pwnly · acme/shop · scan #0142 live
FINDING #07CriticalCWE-639Hypothesis

Broken access control on GET /api/invoices/[id]

Any signed-in user can read another customer's invoice by changing the id.

proof · roles tested: user_a, user_b

$ user_b GET /api/invoices/1042

...

waiting for the probe

expected 403 · got ?

Verify fix Open pull request
sandbox logno network · fake users

Sample data from an illustrative scan. Not a customer.

Broken object-level authorization (IDOR) Missing row-level security Leaked service_role and API keys Privilege escalation Price and credit tampering Webhook signature gaps Wildcard CORS Unauthenticated admin routes Mass assignment Known-vulnerable dependencies Subscription and paywall bypass Replayed payment events

4

analysis layers

Cheap deterministic checks first, deep verification only where it matters.

5

fake roles per test

Anonymous, two ordinary users, an admin and an unpaid user.

0

requests to your live site

There is no field to enter a target. Tests run on an isolated copy.

~90s

to a first pre-scan

Measured on our own repositories. Your numbers will vary.

01Product

Findings you can act on, with the proof attached

A score, a ranked list, the exact request that broke it, and a patch you can review. Click through a sample.

62/ 100

Security score

2 critical · 2 high

1 medium · 1 low

CriticalBroken access control on GET /api/invoices/[id]
Confirmed
Criticalservice_role key committed in .env.local
Confirmed
HighTable invoices has no row-level security policy
Confirmed
HighPrice taken from the client in POST /api/checkout
Potential
MediumPATCH /api/users/[id] accepts a role field
Tested safe
LowCORS allows * on /api/public
Confirmed

Sample data from an illustrative scan. Findings are labelled potential, confirmed or tested safe.

02How it works

Four layers, one autonomous pipeline

Each layer narrows the field. What gets reproduced is labelled confirmed; the rest stays labelled potential or tested safe.

L1
L2
L3
L4

Layer 1·Included in the free pre-scan

Static checks

Leaked keys, an exposed Supabase service_role, tables without row-level security, wildcard CORS and vulnerable packages. Deterministic rules, no AI guesswork.

repo scan · deterministic rules

.env.localExposed secret sk_live_••••••••
supabase/migrations/0004.sqlTable invoices has no RLS policy
next.config.jsCORS allows *
package.jsonlodash 4.17.15: known advisory

Layer 2·Included in the free pre-scan

Map and hypotheses

We map your API routes, auth, billing and webhooks, then write concrete, testable hypotheses: who should not be able to do what.

route map · hypotheses (unverified)

authPOST /checkoutGET /invoices/[id]PATCH /users/[id]POST /webhooks/stripeprice tampering?IDOR?role escalation?

Layer 3·Launch Pass and Pro

Sandbox, with fake users

Your app boots in a throw-away container with no network and fake data. Five fake roles probe every route the hypotheses point at.

sandbox · no network · fake users · observed status codes

GET /invoices/1042POST /checkoutPATCH /users/2
anonymous401401401
user_a (owner)200200403
user_b200200403
admin200200200
unpaid_user403402403

user_b read user_a's invoice: expected 403, got 200. Reproduced.

Layer 4·Launch Pass and Pro

Proof, then a verified fix

A weakness only counts if the exploit reproduces. The fix only ships if it builds, type-checks, passes our content checks and the exploit stops working. You review and merge.

fix · invoices/[id]/route.ts

- const row = await db.invoices.find(id)+ const row = await db.invoices.find({ id, owner: user.id })
  • Type errors: not increased
  • Build and boot in a fresh sandbox
  • Content check: no new hosts, no new secrets
  • Exploit replay: blocked (403)
Open pull request (the owner decides)

03Attack surface

See which routes break, and which ones hold

Pwnly probes each route with fake roles. Flip the switch to see the same app after the verified fixes.

3 reproduced · 3 potential · 2 tested safe

GET /invoices/[id]POST /checkoutPATCH /users/[id]POST /webhooks/stripeGET /admin/*db: invoices (RLS).env.local/api/public (CORS)your appsandbox copy
reproduced potential, awaiting proof tested safe or fixedIllustrative map. Moving dots are Pwnly's probes.

04What we touch

Where we look, where we test, what we never touch

1 · Read

Your repository

Read-only access to the repositories you choose. Known secret patterns are replaced by placeholders before any excerpt reaches an AI model.

2 · Test, inside the boundary

Isolated sandbox, no network

A copy of your app, built from your code, destroyed when the scan ends.

Your app Throw-away database Payment mock 5 fake users

3 · Report

Findings and a fix

Each finding is labelled potential, confirmed or tested safe. A fix waits as a pull request for you to review.

Your live site and production data never contacted

There is no field to enter a target. Pwnly cannot be pointed at a URL, an IP address or a system you did not connect.

05Why Pwnly

A scanner flags. Pwnly proves.

The difference is what lands in your inbox: a list of maybes, or a reproduced bug with a fix you can verify.

Many scanners

Pwnly

Flag a pattern that looks risky

Reproduces the bug with fake users, or reports it as tested safe for that test

Leave triage to you

Labels every finding: potential, confirmed or tested safe

Suggest a code change

Ships a fix only after build, types, content check and exploit replay pass

Often need a live target to test

Tests an isolated copy of your code. Your live site is never contacted

Produce one report per run

Re-tests on every pull request (Pro)

You merge on a diff and a warning

You merge on proof: the same request, before and after the fix

06Trust

Built to be checked, not just believed

Security tools read your source code. These are the controls between your code and everything else.

01

Read-only repository access

You choose the repositories. Disconnect the GitHub App at any time and scans stop at once.

02

Secrets stay out of the model

Credentials that match known secret patterns become placeholders before any excerpt reaches an AI model, and are only restored locally when a patch is written.

03

No raw model output

Every reply is parsed through strict schemas. You see validated findings and fixes, never prompts or transcripts.

04

Hostile repositories assumed

Repository content is untrusted data. Paths, endpoints and patches are checked by code, not by instructions to a model.

05

Every model call is logged

Per organization and user, kept for a year, so misuse can be traced and an account suspended in minutes.

06

Supported regions only

Access follows Anthropic's Supported Regions Policy. The preview is invite-only.

07Pricing

Start free. Pay when you want proof and fixes.

Free pre-scan

$0layers 1 and 2

  • Security score
  • Count of critical and medium risks
  • Potential business-logic risks (unverified)
Request access

Launch Pass

Most popular

$79one-time, per repository

  • All four layers
  • 3 scans within 30 days
  • Proof for every confirmed finding
  • Verified fix as a pull request
Request access

Pro

$99per month

  • Everything in Launch Pass
  • Audit on every pull request
  • Security badge for your README
  • Fair-use limits apply
Request access

Payments by Stripe. Cancel Pro anytime. Purchases open when the preview does.

08FAQ

Questions, answered plainly

Do you test my live website?+

No. Every dynamic test runs against a throw-away copy of your own code in a sandbox with no network and fake data. There is no way to enter a URL or an IP address.

Does my code go to an AI model?+

Excerpts of security-relevant files are processed by Claude models from Anthropic. Secrets are replaced by placeholders first. Anthropic may retain inputs and outputs and review them for safety, as our Terms and Privacy Policy explain.

Is this a penetration test?+

No. Pwnly automatically verifies a defined class of problems: access control, ownership and payment logic, plus secrets, configuration and dependencies. It does not replace a manual penetration test.

How accurate is it?+

We have not published accuracy numbers yet, and we will not claim any we cannot measure. Every finding is labelled potential, confirmed or tested safe, and only reproduced weaknesses are called confirmed. A clean result means those tests passed, not that the app is safe.

Which stacks work?+

Node.js apps, mainly Next.js. Role-based tests assume Supabase-style JWT authentication; other setups get unauthenticated checks and limited verification.

Can I try it today?+

Pwnly is in a private, invite-only preview while we complete a program review. Request access and we will write back.

Ship fast. Know it is safe.

Join the private preview and see your first proven finding in minutes.

Request access