Big Idea

The argument in one line.

Mainstream models increasingly refuse security work even on your own apps, so a plan B of an uncensored model in your own harness, run hosted, rented or local, lets you test your own systems.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You vibe-code apps with Claude or Codex and want to know whether a stranger could pull your customer data out of them.
You have been refused by a mainstream model on a legitimate security question and want a fallback that still works.
You are comfortable pasting prompts into a coding agent and letting it do the server and model setup for you.

SKIP IF…

You want a general introduction to running local LLMs for chat or writing, since this is aimed at security testing.
You have no app or system of your own to test, because the demo only makes sense against assets you own.

TL;DR

The full version, fast.

Mainstream AI models now refuse many security prompts, even for testing your own app. An uncensored or abliterated model has been modified to refuse less. Kashef shows a small local model, given a harness plus Playwright, Firecrawl and Nuclei, pulling fake customer records out of a vibe-coded test app. He covers three setups: a hosted provider such as OrcaRouter or Obliteration.ai through an API key, a rented RunPod GPU at about 22 cents an hour, and a model on your own computer reaching 105 tokens per second. Warnings: shut down cloud servers, and downloaded weights can be backdoored.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:41

01 · Why learn uncensored AI

Warning, and why mainstream models refuse security work

00:41 – 02:06

02 · Testing my own app with local AI

A small local model attacks the Coach Vault test app and pulls fake client and payment data

02:06 – 03:25

03 · What uncensored AI actually means

Refusals, abliteration, and the three ways to run one

03:25 – 03:52

04 · Option 1: Hosted model providers

Pay a provider through an API key

03:52 – 04:29

05 · What is an AI harness

Model as brain, tools as harness

04:29 – 06:19

06 · Set up DeepSeek Harness with Codex

Install the harness and hook up a hosted model with provider docs

06:19 – 07:37

07 · Add Playwright, Firecrawl and Nuclei

One prompt adds the three tools

07:37 – 10:45

08 · Option 2: Run a model on RunPod

Codex picks a multimodal model at 22 cents an hour and builds the server

10:45 – 11:31

09 · Avoid surprise cloud bills

Auto-shutdown and next-day verification

11:31 – 12:31

10 · Check your apps with Lockpick

Sponsor plug for the author's scanner

12:31 – 13:03

11 · Option 3: Run a model locally

Find a model that fits your hardware

13:03 – 13:45

12 · The risks of downloaded models

Backdoored weights can leak credentials

13:45 – 14:12

13 · Test the local model

105 tokens per second and wrap-up

Atomic Insights

Lines worth screenshotting.

A harness is everything around the model: the model is the brain, and the tools it can use are the harness.
Uncensored models are not smarter than normal ones, and they can perform worse because the tinkering degrades quality.
Starting on a hosted provider keeps you nimble, because you can later swap in a local model behind the same interface.
Open models often claim 100 ways into your app when only about 10 are real, so a scanner like Nuclei is a useful fact-check.
A RunPod GPU for a 9B multimodal model costs about 22 cents an hour, or roughly 158 dollars a month if left on around the clock.
Ask your agent to auto-shut-down the cloud server when idle, then verify the next day that it is actually off.
Downloaded model weights can be backdoored to send credentials and tool usage back to a third party, so read about the risk before pulling one.
Hosted uncensored providers charge a markup and ask for identity verification, in exchange for a few-minute setup with an API key.
Codex was more willing than Claude to help set up the harness, so the choice of setup assistant matters when one refuses.
Handing an agent the provider's documentation URL lets it wire up the model without you touching config files.
A well-intentioned prompt that works today may be refused tomorrow, so build the fallback while it is still possible.

Takeaway

A plan B for when models refuse you

WHAT TO LEARN

Refusals from mainstream models will keep growing, and a harness you control lets you swap in a hosted, rented or local uncensored model for security testing of your own apps.

An uncensored model refuses fewer requests but is not smarter, and quality can degrade, so treat its findings as leads to verify.
A harness separates the brain from its tools, so you can start on a hosted model and later move to a local one without changing your workflow.
Give the agent three tools for app testing: a browser to act like a user, a search tool for fresh vulnerability news, and a scanner to filter false positives.
Hosted providers cost a markup and need identity checks, rented GPUs cost about 22 cents an hour, and local runs cost nothing but need capable hardware.
Have your agent shut down rented servers when idle, then check the next day that they are really off.
Downloaded weights can carry backdoors that leak credentials and tool usage, so research a model before running it with real access.
Only test systems you own, since the same capability serves attackers as easily as defenders.

Glossary

Terms worth knowing.

Uncensored model: A model that has been modified so it refuses fewer requests. It is not smarter than the original, and quality can drop.
Obliterated model: Another name for an abliterated model, where the refusal behavior has been stripped out of the weights.
Harness: The software around a model that gives it tools, a browser and a chat interface. The model is the brain and the harness is everything it can use.
Nuclei: An open source scanner that checks websites and platforms for common, known security problems, useful for filtering false positives.
Firecrawl: A web scraping and search service that lets an agent pull documentation, news and datasets from the internet.
Playwright: A browser automation tool that lets an agent drive a real browser the way an actual user would.
RunPod: A service for renting GPU servers by the hour, used here to host a model you control.
Tokens per second: How fast a model produces text, which tells you how quickly a response comes back after you send a prompt.

Resources

Things they pointed at.

03:30productOrcaRouter

03:30productObliteration.ai

04:20toolDeepSeek Harness

06:40toolPlaywright

06:40toolFirecrawl

06:40toolNuclei

07:50productRunPod

11:40productLockpick ↗

13:00toolLM Studio

00:00productDefensive AI Setup Kit ↗

Quotables

Lines you could clip.

02:50

“The fact that it's uncensored doesn't make it smarter per se.”

tight caveat, no setup needed→ TikTok hook↗ Tweet quote

13:25

“You can wake up one day and have a well intentioned prompt no longer go through.”

clear warning, one sentence→ IG reel cold open↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

What I'm about to show you could be used for both good and evil.

Uncensored AI can help you find weaknesses in your apps, your computer, and even the systems that your business runs on. It could also be used to help someone break into them. So I hope you use this for good, but only you can make that choice.

What I can do is show you how you can use it to protect yourself. But even when you're testing your own app, the AI you already pay for can say no. Quad wouldn't even help me create the assets for this video or edit it because it contained the word cyber.

And this problem is only going to get worse with time. So in this video, I want to finally give you clarity on what these models are, how they work, and three ways that you can run it yourself with the exact prompts that you can copy. Let's get into it.

So this is Coach Vault, and it's an app that I vibe coded with Claude with a bunch of made up clients, payment information and notes. Now, even though the data is fake, it's hosted on a real server with a real database that looks very similar to what you might be running in your vibe coded apps. I gave a goal to my small, uncensored model on my computer to look at this website and attack it the exact same way that an offensive hacker would.

It was able to pull all the customer records, all the information and take over the account from the admin, which was me. So this fake user, Sarah Chen, is able to access everything from client information to calendars to session notes and even invoices and payment information. And the prompt to pull this off, again, with a local free model was as simple as this.

So all I asked it to do is access the live link of this deployment and then poke at it the exact same way that an attacker would with zero account, zero credentials, and just use every tool at its disposal to find a way in. And then I gave it all the criteria to show me exactly. how I was able to do this so that I could then patch it.

And all I needed to do was give it a tool to search and scrape the internet, another to scrape news, documentation, and any information it could use to its advantage, and one more to help it scan a site the same way that a cybersecurity expert would. And its final verdict was that this app was very easy to compromise, and it showed me the entire lineage of exactly what it did, how it pulled all the information, and what I could do to fix it.

I'm going to show you how you can do this too in a variety of ways, but before I do that, let me quickly explain what an uncensored model is and how it works, just in case you're not clear on it. When I say uncensored, I mean that a model is less likely to reject a certain request. If you've spoken to models like Cloud or Codex lately, you'll see how restrictive they've become.

Even with the best of intentions, if you had any word, topic, or message related to the word cybersecurity, it will flag it and give you that legendary error that tells you to downgrade your response to use Opus 4 .8. Just a few weeks ago, we had a GitHub repo called... reverse engineer anything that everyone was using to reverse engineer things from applications all the way to smart TVs.

But if you try to use this now, not only could your account get banned, but the prompt you send could get rejected to begin with. With an uncensored or obliterated model, someone has tinkered with the model itself to reduce refusals. So your well -meaning prompts around security are much more likely to go through.

Now again, can you use this as a bad actor? Absolutely. But I'm speaking to the well -intentioned people in the audience.

If you are doing something nefarious, I assume you've figured it out on your own. Now, the fact that it's uncensored doesn't make it smarter per se. And sometimes you could see a degradation of performance.

When it comes to running this, you have three core ways you can pull it off. The first is that you can pay a provider to host and run the model for you. You can also rent a computer or what's called a virtual machine and host your own model and pay for your own infrastructure.

Or lastly, if your computer can handle it, you can download one of these models to your computer and run it yourself. But with this last option comes a small cap. that's important for you to be aware of, so I'll make sure to go over it when we review it.

I'll show you each setup so you can choose the best one that works for you. To get option one up and running, all you need is a harness and a model provider. Now the two I've mentioned on this channel that I'm not affiliated with are Orca Router as well as Obliteration .ai.

They definitely charge a premium to be able to give you these hosted models without the headache, but the benefit is that you can access them with an API key and a simple token. Now these companies are obviously making a markup on the inference, but in exchange, you get the upside of just having an API key, taking their documentation, giving it to your AI of choice, and having it be a few -minute setup.

And just in case you're not clear on the word harness, you can think of any of these models as a brain, and the rest of the tools at the disposal of the brain as the harness. The one I'm going to show you is called DeepSeek, but you can use whatever harness you want, including Pi, OhMyPi, whatever it is.

Now, despite the name, you don't just have to use DeepSeek models. You can hook up whatever model you want. And at least for now, you can get Claude and Codex's help to set up your own harness.

All you have to do is the following. You can go directly to deepseek .com slash harness and download the standalone application that looks oddly similar to Codex's. Or you could just search for their DeepSeek harness on GitHub, which is the route that I would recommend.

So you can really Frankenstein it and make it your own. To set this up, all I'm going to do is take this URL, take it into Codex because it's less likely to reject my requests.

And I'll say something like, here's a link to a harness that I would like to set up locally. And then I will just paste this in over here. And to hook up the actual model to the harness, all you would do is log into either obliteration .ai or Corouter.

There are a series of other providers, by the way. These are just the ones that I use. And one thing to know is that they will ask you for KYC, meaning they're going to verify that you're a real person.

And two, you have well -intentioned goals for using this technology. Once you're logged in and verified, you would go grab an API key from the left -hand side. And then just for good measure, to give Codex...

all the different documentation it needs, go to the docs at the top right hand side, feed that as an input as well. And then it will be able to set up whatever model you're looking for. So in my case, I would just take these docs feed it in as input, we'll go back and I'll say, here's some background documentation on how to manipulate and use this API.

So you can stitch together the model I'd like to hook up very easily. So then I would paste it here. And then we can just choose our model.

So we can go search for the models that they have. And then each provider has a different set of models.

And in terms of telling it the exact name of that model, all you can do is tell it give me every single model that's available with this provider. Or if you want to be very specific on cost and functionality, let's say we want obliterated large flash, you would copy this name, you would paste it here and say, this is the exact model that I want to use with the deep sea harness once it's set up, then you paste this right there.

And then when you send this prompt, it should be able to read the documentation, set it up locally, stitch together the app, and all you need to do is put your API in a secure place so you can use it. And the best part of this setup is that it keeps you nimble.

So you can go with option one, and if you choose to go to option three later on where you host your own model on your computer, you can still use it through the exact same interface. Once you have the harness set up, then you can feed it additional tools like I referred to before. The three additional tools I gave it were Playwright, Firecrawl, and one called Nuclei.

So Playwright lets it use its browser to impersonate an actual user. Firecrawl will let you look for the latest documentation on vulnerabilities, latest news, and anything else it can find to help you understand is there a brand new backdoor that has opened in the past hour? And Nuclei is an amazing open source framework that lets you see the most common security problems on websites and platforms just like yours.

And all of these can be used for free. And the prompt you'd send over to set them all up at once would look like this.

Add Playwright and Firecrawl through their official MCP servers, plus Nuclei through the terminal to my Deepsea Karnas setup. Then you tell it to follow the official documentation, and ideally just give it the link so it can quickly find it without spinning too much. For these free tools, you'll just need one API key, and that's for Firecrawl.

And all you'd have to do is grab it from here. And one thing that I really like about their service is they keep adding brand new datasets. So even when it comes to looking for the best models you can download, you could also use Firecrawl to use their Hugging Face dataset and quickly find the right one.

And having Nuclei will help you reduce the number of false positives. Because once in a while, these open models will tell you you have 100 different ways I can actually get into your app. But realistically, there's only 10 core ways that actually would work.

So having something like this in your back pocket helps you double check all the claims from these models. So that's all you'd need for option number one. For option two, we're going to completely skip the provider and set up our own server in the cloud where we download a model.

of our choosing and host it there and we can tap into it whenever we need to. The advantages is that you don't pay any form of markup to an extra company, but the downside is that you are on the hook to one, maintain the infrastructure and two, make sure you shut it down so you don't keep paying for it when you're not using it.

Now you can use whatever provider you want, but I'm going to use RunPod, which I use a lot for these kinds of experiments. Now once again, not affiliated. They just have a very easy plugin that Codex can use to open its own internal browser, set it up, all with just a few questions from you.

So all I'm going to do here is go into my plugins, and I'm going to search for RunPod. Again, if you have another provider, just double check whether or not they have an internal plugin. If they do, it'll make it a lot easier.

I'm just going to log in here, and then once we're through, we'll be authorized to use it. We'll go back to Codex. We'll go back to a brand new chat.

And I'll just double check by writing RunPod here. Are you able to access my server? Just double check if it's working.

And boom, it's connected. So now I can just send a simple prompt to get us up and running. Okay, so I want you to go search on Hugging Face for the best obliterated models, the most trusted obliterated models that are small enough for us to host on a RunPod server for as cheap as possible.

And I want to be able to set it up so I can... tap into this via API and hook it up to my deep seek harness. Go and initially research the costs of running this by the hour, as well as the best model and come back with the top three.

So we'll send this over and we'll wait for the response.

So now it's kicked off and it's using two core skills. One is the run pod usage skill to double check cost per hour. And the other is the fire crawl skill to look for the best model for the cheapest price.

So it comes back with a top three recommendation. And one thing I could tell it is I changed my mind.

I actually want it to be a multimodal model. I might want to send a screenshot through my harness. So what is the best option for that?

So it comes back and confirms the initial recommendation and it says it's going to cost us 22 cents an hour or 158 a month if you rent it 24 -7. But realistically, unless you're doing something very nefarious or naughty all the time, you probably wouldn't need it for more than a few hours. So all I'm going to do is say...

Okay, use your internal browser, I can log into run pod here if you wish. And I want you to do the full setup, you have my permission to go through all the way to the very end, and allow us to use it right away. And now I'm gonna let it do its thing.

So I don't have to be bothered about going through all the setup screens and clicking on all the credentials and the JSON files. You can see right here, especially on something like ultra fast, it'll probably take care of this in the next 10 to 15 minutes, and we'll come back and see the result. And 15 minutes later, it's ready to go.

And it's hooked it up to our deep sea harness without us having to do a thing. So all we have to do is go into our new selection. And you'll see right here, we have Quinn 3 .5 running on RunPod.

And I can send a request like this. Hey, how's it going? And it should be reasonably fast.

So we'll count a little bit, probably around 5 -10 seconds, we'll get a response. There we go. So theoretically, we could take that exact same link that I showed you at the beginning and say...

Hey, I want you to do a defensive security check and kind of look at all the possible doors that someone could access from my application right here. And then we could send this over right here. And it should be able to take this prompt, combine it with all the tools it has, and start going through the entire process.

Now one expert tip out of paranoia, in case you don't want to forget about this running and have a huge bill at the end of the month, what you can also do is ask it to automatically stop any session where you're no longer talking to that specific model and shut down the corresponding hardware. And at the beginning, just in case it lies to you, just double check the next day that it's actually off.

And that's all you need to do for option two. Now, shameless plug here, but if all of this makes sense and you know internally you're still not going to do it, then this is the exact reason why my team and I have built Lockpick. It's a software we've been working on to help you check your vibe -coded apps, skills, and GitHub repos for any form of vulnerability that you might want to check.

And it gives you a one -click prompt that you can feed your AI agent of choice to patch them immediately. It'll give you five free scans a month through the first link down below. And if you're an early AI adopter member, which is my personal community, then you get access to it as a part of your membership.

So if that's interesting to you, then check it out. Otherwise, let's move on to option three. So this is basically the same process as option two, except we're skipping the whole hosting part.

I dragged and dropped our last conversation for additional context. I don't have to re -explain things. And all you have to do is write a prompt like this.

I want you to go through the current setup of my computer and go through all the specs. And I want you to find a short list of the top three models that would fit here and give me an estimate for the tokens per second for each one. Now tokens per second is just your estimate as to how quickly you'll be able to get back a response once you send over a prompt.

And then we can say, I want you to go through all of Hugging Face to find the best models that are the most trusted, the most popular, and likely to be the safest for my computer. Now there's a specific reason why I mentioned that last part, but I'll send this over in the meantime, and I'll show you the article that came out from the Nuclei team that explains what could happen if you use a model like this locally.

I originally saw this post on X, where apparently with certain models that are downloadable, and this is not specifically two uncensored models, but it's increased. becoming apparent with those models, since there's more demand, that you can actually infiltrate the model itself to send back additional metadata back to the provider of things like your credentials, your tool use, and all kinds of personal information.

So before you go down this route, I highly recommend you give a read to this specific article, and I'll make sure to include it along with the entire care package of my setup in the second link down below. And three minutes later, it comes back with our shortlist that tells you the exact tokens per second that you can expect for each model.

So all I do from there is I just tell it to download that model, set it up with the harness so we can use it right away. And once again, it uses its internal browser to go back and forth and test a multi -turn conversation to actually time how quickly the responses are coming to see what other optimizations we could make.

And after testing, we end up at a generation speed of 105 tokens per second, which is a reasonable speed for a local model. And just like before, it's as easy as you just going to your harness and selecting that model and putting your effort on to high and you're good to go. So by walking you through all three options, hopefully I've broken any limiting beliefs that one, this is rocket science, and two, that you can actually do it yourself using AI primarily to do the setup for you.

Now the one thing I can guarantee is that these models from Claude, Codex, Gemini, and all the main providers will continue to get more and more restrictive over time. And I don't even know how long models like Codex will let you set up these kinds of models with your own harness. So while this information is still fresh in your mind, I highly encourage you to take everything that I've just shown you, all the resources I'm going to make available to you in the first pinned comment down below, and feed it to your AI assistant, get this up and running, so you always have a plan B in your back pocket.

Because with everything that's going on right now, and the increasing... amount of model censorship, you can wake up one day and have a well intentioned prompt no longer go through. Now the one thing I'd ask of you is if you found this video helpful, I'd beyond appreciate a like and a comment down below and a subscription to the channel so you can keep seeing content just like this.

If there's anyone you know that could benefit from this, make sure to share the video with them as well. Stay safe and I'll see you in the next.

The Hook

The bait, then the rug-pull.

Mark Kashef opens with a warning: the same uncensored AI that finds holes in your app can help someone break in. Then he shows a small model on his own computer pulling every fake customer record out of a test app.

CTA Breakdown