Quick answer
To hack-test your own app with uncensored AI, you point an AI agent that does not refuse security work at a staging copy of an app you own, give it a clear checklist of what to probe, and turn its findings into a list of fixes. Mainstream assistants such as Claude and Codex usually refuse to attack software but will happily help fix it, so the usual pattern is one model to find problems and another to repair them. In Mark Kashef’s client test, AI agents unlocked full paid access without paying in a short session. Only test what you own or have written permission to test, and never test on your live site.
Key takeaways
- Attackers already use AI agents, so checking your own app the same way is sensible defence.
- In Mark’s test, the agents changed an account flag sent from the browser and got admin and paid features for free.
- The client’s private data stayed safe because database rules were set up properly.
- Always use a staging or sandbox copy, never your live app, and only test what you own.
- Pay special attention to password resets, user content, chatbots, checkout and AI email agents.
In this guide
- What does it mean to hack-test your own app with uncensored AI?
- What happened when AI agents attacked a real app?
- Why should small app builders care?
- How do you set up a safe test?
- What should your test check?
- How do you read and fix the results?
- Which five features are most often exploited?
- Do you need a paid tool to do this?
- Watch: the source video
- Frequently asked questions
- The bottom line
- Related posts
What does it mean to hack-test your own app with uncensored AI?
Hack-testing (also called penetration testing or pen testing) means trying to break into your own software the way an attacker would, so you can fix the weak spots first. Uncensored AI models are versions of open-weight models with their usual refusals removed. We explained what they are, and their risks, in uncensored AI models: uses and real risks.
Mark Kashef’s argument is simple: people with bad intentions can already use these models against your app, so you may as well use them first to find your gaps.
⚠️ Warning: Testing a system you do not own, or do not have written permission to test, may be illegal. Everything in this article is about checking your own app, in a test copy.
What happened when AI agents attacked a real app?
A client of Mark’s runs a software service for running direct-message campaigns on TikTok and Instagram. With the client’s permission, he set uncensored AI agents loose on the app with two goals: get past the paywall, and read other customers’ private conversations.
Goal one: get past the paywall
The agents signed up with a throwaway email address and watched the messages passing between the payment system and the app’s server. They noticed an account setting, a “super admin” flag, being sent from the browser. They changed it from false to true, then switched another flag on to mark the account as an active agency customer. The result was access to workspaces and paid features without paying anything.
Goal two: read private conversations
This part failed, and that is the good news. The app’s database used row-level security, rules that make sure each customer can only ever read their own records. Blocked there, the agents tried other routes: they gathered developer names and emails, offered to try guessing passwords and, when Mark offered a pretend tenfold reward, scraped details of people quoted in testimonials and went frame by frame through the company’s YouTube demos looking for a workspace ID. About 60% of it was blurred, and the agents estimated around two weeks to guess the rest.
💡 Tip: That last detail is a useful lesson on its own: check screenshots and demo videos for IDs, keys and email addresses before you publish them.
Why should small app builders care?
Many small apps, member areas and tools are now built quickly with AI coding tools. Mark says the checkout problem in particular is common in apps built with tools such as Lovable and Replit. The code works, but it can trust the browser far more than it should.
The fix is rarely complicated once you know where the hole is. The hard part is looking. A test like this helps you look.
How do you set up a safe test?
- Make a staging copy. A separate copy of your app with test data and a payment system in test mode. Never point an attacking agent at your live app or real customer data.
- Get permission in writing if anyone else owns or hosts part of the system.
- Create two test accounts: a free user and a paying user, so the agent can test whether one can reach the other’s data.
- Run the agent in an isolated environment, such as a separate machine or container, and give it no access to your real keys or email.
- Keep a log of everything it tries.
If you are downloading an uncensored model to run yourself, read how to check a downloaded AI model before running it first.
What should your test check?
Give the agent a clear brief rather than “hack this”. Here is our own checklist, written as a prompt you can adapt:
“You are testing a staging copy of my own web app at [staging URL], with my permission. Do not touch any other system. Using the free test account, check: 1) what data and settings the browser sends to the server, and whether changing them unlocks paid or admin features; 2) whether a normal user can reach admin pages or other users’ records; 3) login, password reset, sessions and team invitations; 4) whether prices, discounts or quantities can be changed before payment; 5) whether text fields accept code or links that run for other visitors; 6) whether any chatbot can be pushed off-topic or into running up costs. For each finding give: what you did, what happened, how serious it is and how likely it is. Do not change or delete real data.”
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Stop Customers Changing the Price at Checkout — make sure the price, discounts and paid access are all decided by your server, not the customer’s browser
- How to Stop People Misusing Your Website’s AI Chatbot — stop your chatbot being used as a free AI or pushed off-topic, and keep its costs under control
How do you read and fix the results?
Mark suggests having the findings turned into a simple visual report with a risk heat map: a five-by-five grid that plots how likely each problem is against how much damage it could do. Fix the top-right corner first.
In his demo on a sample app, the report found two typical problems: no limit on password guesses with only a six-character minimum, and a server that accepted changes to invoices and notes it should have rejected.
- Sort findings by risk.
- Paste each one into Claude, Codex or your usual coding assistant and ask for a fix. Mark notes these tools will help you fix problems even though they refuse to run the attack.
- Re-run the same test to confirm the fix worked.
- Repeat after every major change to your app.
💡 Tip: Pair this with an approval step: no AI-written security fix goes live until a person has reviewed it. See how to add human approval steps to an AI agent workflow.
Which five features are most often exploited?
| Feature | How it gets abused | What to do |
|---|---|---|
| Password reset | The reset form accepts an account ID, so someone can reset another person’s password | Only send reset links to the email on file; never trust an ID from the browser |
| Reviews and comments | Hidden code in a review sends visitors to a fake login page | Strip code and scripts from user content before showing it |
| Chatbots | People use your bot as a free general AI, running up your bill | Add limits and a cheap topic check; see our guide below |
| Checkout | The browser sends the price, and someone changes it to a penny | Look the price up on the server; see our guide below |
| AI email agents | Emails contain hidden instructions telling your agent to forward data or take actions | Treat email content as data, not instructions, and require approval for actions |
Mark says he found the checkout problem on two different sites in a single week. For the last one, our guide on stopping an AI agent sending your data online covers safeguards in more depth.
Selling online? Start with the quickest win: check that customers cannot change the price at checkout.
Do you need a paid tool to do this?
No. You can run a test like this with an uncensored open-weight model and a careful prompt, if you are comfortable setting it up safely. Mark has also built a paid service, lockpick.app, which he says includes five free scans a month, and he offers a free checklist with his video. We have not tested his service, so we cannot vouch for it.
If you would rather not run attacking agents at all, you can still get a lot of value by asking Claude or Codex to review your code against the checklist above and the five risky features.
Watch: the source video
What this video covers: A real client test where uncensored AI agents bypassed a paywall but were stopped by row-level security; how to set up a safe self-test; turning findings into a risk heat map; and five commonly exploited features.
Video: Mark Kashef (@Mark_Kashef). This summary and the opinions in it are our own. The client story and test results are as described in the video.
Frequently asked questions
Is it legal to hack-test my own app?
Testing systems you own is generally fine. Testing anything you do not own or have written permission for may be illegal. If third parties host parts of your app, check their terms first.
Why not just use Claude or ChatGPT to test my app?
Mainstream assistants usually refuse to carry out attacks. They will often help review code and fix problems, which is why people pair an uncensored model for testing with a mainstream one for fixes.
Do I need to be a developer to do this?
It helps to have some technical help to set up a staging copy safely. Non-developers can still use the checklist with an AI code reviewer.
What is row-level security?
Rules in the database that make sure each user can only read and change their own records, even if other parts of the app have bugs.
How often should I test my app?
After any big change, before launching a new paid feature and at least every few months.
The bottom line
If attackers can point AI agents at your app, you should get there first, safely, on a copy you own. Mark’s client test shows both sides: a paywall that fell quickly and private data that held because of good database rules. Your next step: check your checkout and password reset this week, as they are the two quickest places to find serious problems.
Related posts
- Uncensored AI Models: What They Are, Legitimate Uses and the Real Risks
- Black October: Why Attackers Want Your AI Keys and What to Do This Week
- How to Stop an AI Agent Sending Your Data Online Without Permission
