Book a crash test

PromptBuildCrashShip

Your AI wrote it. Our dummies crash it.

Crash Dummy is the crash-test stage for apps built with Lovable, Bolt, Replit, Cursor, v0 and Base44. We send test users into your app, see what breaks, and write every fix as a plan you hand straight to your AI agent.

You don't need to read the fix. Your agent does.

Your own app only. You verify that you own it before we touch anything.

Cam B · Dummy B● Frame 2118

T+02.118s███ IMPACT ███Totaled

Dummy B is looking at Dummy A's private task.

Illustrative freeze-frame. Crash Track is our own demo app, broken on purpose. No real users, no real data.

Where Crash Dummy sits in your build

  1. PromptYou describe the app.
  2. BuildYour AI tool builds it.
  3. CrashWe send in the dummies, rate it, and write the fix.
  4. ShipYou launch with a Safety Rating you can trust.

You already have a prompt step and a build step. This is the step that was missing. And it closes the loop: the crash test writes the fix plan, your agent applies it, and the dummies check the result.

01The test

Two dummies walk into your app.

Most leaks in AI-built apps aren't clever hacks. They come from one ordinary question nobody asked: can one user see another user's stuff? So we ask it in the most literal way possible.

Test account 1

Dummy A

Signs in like a normal user. It creates a task, uploads a file and fills in its profile. Then it sits quietly.

Test account 2

Dummy B

Signs in next door. It takes everything Dummy A touched and tries to read it, change it and delete it. It also tries to give itself the paid plan and wander into the admin page.

If Dummy B gets through, so could any of your users. You'll see exactly where it got in, frame by frame, and the prompt that closes the door.

The two-dummies test is part of the Full Crash Test. It runs only on test accounts you create for us, inside an app you've verified you own.

02Crash Replay · Crash Track (our demo app)

Watch a crash. Nobody gets hurt.

This is Crash Track, a small task app we built in Lovable and broke on purpose so you can watch a crash safely. It has no real users and no real data. Every account in it belongs to us.

Every finding in your Crash Report comes with a replay like this one: what each dummy did, when, what the app answered, and the moment of impact.

Illustrative script, not a real run. It plays in your browser and doesn't scan anything live. The real recorded run of Crash Track replaces it after Crash Test Lab #1.

Crash replay · crash-track · 1000 fps · illustrative● REC · T+02.118s
Cam A · Dummy Aown tasks
Buy more dummiesprivate
Cam B · Dummy Bimpact
Buy more dummiesA's

Impact

  1. T+00.000sDummy Asigns in✓
  2. T+00.418sDummy Acreates task "Buy more dummies"✓
  3. T+01.902sDummy Bsigns in✓
  4. T+02.117sDummy Basks the database for Dummy A's task→ answered
  5. T+02.118s███ IMPACT ███Dummy B read Dummy A's taskTOTALED

Safety Rating ★☆☆☆☆ 1/5 Ship held

↳ Repair prompt ready

Frame 2118: the moment Dummy B reads a task that isn't its own.

One crash, one fix plan. Hand it to your agent (Copy for Lovable, or save it to your project) and let it rebuild. Then press Re-crash. The same dummies run the same moves, and this time Dummy B goes home empty-handed.

See the full sample Crash Report

03What you get · the Crash Report

A report you can read without a security degree.

A Safety Rating from 1 to 5 stars, calculated by a published rule. How the rating works →

  1. Scratch

    Cosmetic. Fix when convenient.

  2. Dent

    Someone could abuse this. Fix soon.

  3. Crash

    Something important is exposed. Fix before launch.

  4. Totaled

    Other people's data is open. Stop and fix now.

  • Each finding explained in plain English. What happened, who could do it, and why it matters. No acronyms without a translation.
  • Evidence you can check. We save what we sent and what came back, with every other user's data blacked out.
  • Repair prompts. Each fix is written as a prompt for the tool you built with. Copy for Lovable, Copy for Cursor, and so on.
  • A fix plan for your agent. One button, Hand this to your agent, copies the whole plan: what to fix, in what order, how to check each fix, and what never to touch. It comes as a file too (crashdummy-fix-plan.md).

In every report

The report is for you

  • A star rating and what it means
  • Plain-English explanations
  • Replays you can watch
  • "Is it safe to launch?"

The fix plan is for your agent

  • Findings in priority order, with stable IDs
  • The exact table, bucket or key to change
  • Verify steps after every fix
  • Guardrails: never switch off security to make a check pass

You stay the driver. Your agent does the repairs. You decide when to launch.

04Plans

Pick your crash.

The Stranger test

Dummy Run

$49

One dummy, no account, no password. It sees exactly what anyone on the internet can see.

  • Keys and secrets left in your app's public code
  • Database tables anyone can read without logging in (we count rows and never copy them)
  • Storage folders anyone can list or download from
  • Database functions anyone can call
  • Files that should never be public (.env, .git, source maps) and basic security settings
  • A Crash Report with a Safety Rating, repair prompts and a fix plan for your agent

Delivered by email within 24 hours

Order a Dummy Run · $49

Coming soon

Re-crash on every deploy

Airbag

$29/month

It doesn't stop the crash. It tells you the moment there is one.

AI edits are fast, and sometimes they quietly undo last week's fix. Airbag re-runs your crash test on a schedule or after each deploy, and alerts you when your Safety Rating drops.

I'd want this

Join the list. We'll email you when it's ready. You won't be charged until then.

How ordering works

  1. PayThrough PayPal. Put your app's URL in the payment note.
  2. VerifyWe email you from [email protected] with a one-line verification tag for your site and a link to our testing terms. Add the tag and reply “I agree”. For a Full Crash Test, you also sign the one-page testing authorization.
  3. CrashTesting starts only after a paid order, verified ownership and your agreement to the testing terms. Then the delivery clock starts.
  4. RepairYour Crash Report and fix plan arrive by email. Hand the plan to your agent.

If we miss the delivery time, you get a full refund. Read the service terms before you order (a draft under legal review). Questions before you buy? Email us.

05Where it runs

Runs where you build.

  • Available now

    Any browser

    Your app is on Lovable, Bolt, Replit, v0, Base44 or your own domain.

    Give us your app's URL, verify you own it, and get a Crash Report and fix plan by email.

  • Available now

    Your agent's instructions

    Lovable Instructions, AGENTS.md, CLAUDE.md, Cursor rules.

    Paste one snippet, once. Your agent then reminds you before risky deploys, and knows how to apply a fix plan safely. Get the snippet →

  • Early access · not live yet

    Inside your AI coding agent

    Lovable, Cursor, Claude Code, Bolt, Replit.

    An MCP server: tell your agent "crash-test my app." It runs the test, applies the fix plan and asks for a Re-crash. You still approve payment.

  • Early access · not live yet

    On every pull request

    GitHub.

    A GitHub Action that crash-tests your preview deploy and posts the Safety Rating on the PR.

  • Early access · not live yet

    In your terminal

    npx crashdummy scan <url> --slowmo

    Replays the crash frame by frame.

Coming later: let your agent crash its own app. Build, crash, repair, re-crash, all in a loop, without you copying and pasting.

Join early access Tell us which one you'd use first. It decides what we build next.

Building with an agent? It can read our docs too: crashdummy.dev/agents.

06Trust

We only crash your own car.

Ownership first, always.

Before any test, you prove the app is yours. You add a one-line tag to your site, a small file, or a DNS record. Ask your builder: "Add this meta tag to the head of my site." Our engine refuses to test any address that hasn't been verified, and that rule sits in the engine itself, not just in a form.

We look, we count, we don't keep.

When something is open, we record that it's open and how much, for example "the profiles table is readable by anyone: 1,240 rows." We never record what's inside. Secrets show up masked (sk_live_ab12…). Evidence is deleted 30 days after your report.

Polite by design.

We send at most 5 requests per second. Spam checks stop at 10 attempts. We never run load tests, never try to knock your app over, and never touch real payments.

Changes only to our own dummies.

In the Full Crash Test, every write happens in the test accounts you give us, and we clean it up afterwards.

Your report is private.

It goes only to you. If you want to share a public summary, you can choose to, and it never includes evidence.

Read exactly what we test and what we never do →

07Founder

Who's behind this.

Ron Rosenzwieg is a QA and test-automation engineer. He works with Playwright, API and end-to-end testing, and tests AI and MCP systems. He also has a background in security. LinkedIn

Crash Dummy is new. Founding clients get a lower price in exchange for honest feedback. We don't show logos or testimonials we don't have. Become a Founding Crash Tester →

08FAQ

Straight answers.

All questions →

What exactly is Crash Dummy?

Crash Dummy is the crash-test stage for apps built with AI tools like Lovable, Bolt, Replit, Cursor, v0 and Base44. We test your live app from the outside. First a stranger with no account tries it. Then, in the Full Crash Test, two test users try to get at each other's data. You get a Crash Report with a 1–5 Safety Rating, plain-English findings, and repair prompts you paste back into your AI builder.

Lovable (or my builder) already has a security scan. Do I still need this?

Keep using it. It's a good first line. Lovable's Quick scan, for example, runs when you publish and checks for tables without row-level security, rules that let everyone through, and vulnerable dependencies. Lovable's own docs say these tools don't replace a thorough security review.

A scan reads the blueprint. A crash test drives the car into the wall: we check what a second user can actually reach in your live app, and show you a replay of it. Run the scan, fix what it finds, then crash-test.

Will this break my app or touch my users' data?

We're deliberately gentle. We send at most 5 requests per second, and abuse checks stop at 10 requests. We never run load tests. In a Full Crash Test, all writes happen in the two test accounts you create, and we clean them up afterwards. When something is exposed, we record that it's exposed and how many rows, never the contents.

Do I need to be a developer?

No. The report is written for founders, designers and makers, not security engineers. Every finding explains what happened and why it matters. The fix comes as a prompt you can paste into the tool you built with.

Does a 5-star Safety Rating mean my app is secure?

No. No test can guarantee that. Five stars means our dummies found nothing confirmed in the areas we test, on the date of the test. Your next AI edit can change that, which is why every report is dated.

Can I crash-test someone else's app?

No. We only test apps whose owner has verified ownership, with a tag on the site, a file, or a DNS record. Unverified addresses are refused. That includes competitors, clients who haven't agreed, and "just curious."

Can my AI agent apply the fixes?

Yes, and that's the idea. Every report comes with a fix plan written for your agent: what to fix, in what order, how to check each fix, and what never to touch. Click "Hand this to your agent" (or save the file to your project), and tell your agent to apply it. Then press Re-crash to confirm. You don't need to understand the fix yourself.

Crash it here, not in production.

Pay, verify you own the app, then we test. Nothing runs before both. Not ready to buy? Try the free 15-point pre-drive inspection and check your app yourself.