Uncensored AI Models: What They Are, Legitimate Uses and the Real Risks

Key takeaways

• “Uncensored” means fewer refusals, not more intelligence. Quality can even drop
• The main legitimate use is defensive security testing of your own apps, sites and systems
• Mark shows three ways to run one: a paid hosted provider, a rented cloud server, or your own computer
• Only test what you own or have written permission to test. Unauthorised access is a crime under the Computer Misuse Act 1990
• Downloaded models and their files can carry risks, so check them before you run them
• Rented GPUs bill by the hour, so set limits and shut them down
• Results from AI security scans need checking: Mark warns they produce false positives

Have you ever asked an AI assistant to help check your own website for security holes, only to be told it cannot help? You are not alone. As the big AI providers tighten their safety rules, some genuine security questions now get blocked along with the harmful ones.

In a new video, AI consultant Mark Kashef makes the case that you should learn how uncensored AI models work, so you have a plan B for testing your own apps. Below we explain what he shows, where these models genuinely help, and the legal, safety and cost risks you need to weigh first. We will not cover how to misuse them, and neither should you.

What are uncensored AI models?

An uncensored AI model is a language model that has been changed so it is less likely to refuse a request. Most start life as ordinary open-weight models (models whose files anyone can download and run). Someone then adjusts them to strip out much of the refusal behaviour.

You will often see the term “abliterated” for one common method, which Mark pronounces as “obliterated”. The model still knows what it knew before. It is simply more willing to answer.

Why does Mark Kashef say you need to learn them?

Because, he argues, mainstream AI keeps getting more restrictive, and honest security work is getting caught in the net. Mark says even well-meant prompts that mention cybersecurity now get flagged, and that a tool he relied on for creating assets would not help with this very video because it contained the word “cyber”.

He also mentions a popular reverse-engineering project that people were using with mainstream coding assistants, and says using it now risks a rejected prompt or even an account ban. We covered a related development in our round-up on Claude’s reverse engineering block.

He is upfront that this technology can be used “for good and evil”. His pitch is aimed at the well-intentioned: learn it so you can protect yourself, and have a backup if your usual assistant stops helping.

What did the test on his own app show?

Mark built a test app called Coach Vault with Claude, filled it with made-up clients, payments and notes, and hosted it on a real server with a real database. He then pointed a small uncensored model, running on his own computer, at it and asked it to behave like an outside attacker with no account.

According to the video, the model got into the app, pulled the fake client records, calendars, session notes and invoices, and took over his admin account. It then explained each step it took and suggested how to fix the problems.

The tools he gave it were a browser automation tool (Playwright), a web search and scraping tool (Firecrawl) and an open-source vulnerability scanner (Nuclei).

What are the legitimate uses of uncensored AI?

The legitimate uses are defensive and research work where you are authorised. For example:

• Testing your own app or website before launch, on a copy with fake data, as Mark did
• Reviewing your own code for weaknesses that a mainstream assistant refuses to discuss
• Security research and training inside a lab you control
• Authorised penetration tests carried out by professionals with a signed scope
• Private, offline AI use, where running a model on your own machine keeps data off third-party servers

What is not a legitimate use: probing anyone else’s site, account or network, or anything you would not be comfortable explaining to the owner. If in doubt, do not run it.

What are the three ways to run an uncensored model?

Mark walks through three routes, from easiest to most hands-on. We summarise them at a high level here so you can judge the trade-offs; this is not a setup tutorial.

In all three, Mark connects the model to a harness, which he describes as the tools around the model’s “brain”. He uses one he calls DeepSeek Harness and gets Codex to do most of the setup by giving it the official documentation.

Some numbers from the video: his chosen cloud setup was estimated at about $0.22 an hour, or roughly $158 a month if left running around the clock. His local setup reached about 105 tokens per second. These are his figures for his setup, not benchmarks.

What are the risks of uncensored AI models?

The main risks are legal, security, cost and quality.

• Legal: using one against systems you do not own can be a criminal offence, whatever your intentions
• Downloaded files: Mark points to a security write-up warning that some downloadable models could be tampered with to leak data such as credentials. Model files in older formats can also run code when loaded
• Tool access: a model with browser, terminal or file access can do real damage if it misreads an instruction
• Cost: forgotten cloud servers keep billing
• False positives: Mark says open models can claim a hundred ways in when only a handful really work, which is why he adds a scanner to double-check
• Quality: removing refusals can make answers worse, not better
• Account and provider rules: using a mainstream assistant to set up tools meant to get around its own safety rules may break its terms of use

Is it legal to use uncensored AI for security testing?

Owning or running the model is not the issue in itself. What you point it at is. In the UK, the Computer Misuse Act 1990 makes unauthorised access to computer systems an offence. Good intentions are not a defence if you did not have permission.

• Test only systems you own, or that you have written permission to test
• Agree the scope in writing: which sites, which dates, which techniques
• If you find a problem in someone else’s product by accident, report it through their vulnerability disclosure route rather than probing further. The NCSC publishes a vulnerability disclosure toolkit
• Check the licence of any model you download and the terms of any provider you use

This is general information, not legal advice. If you are testing for clients or at work, get qualified advice and a signed agreement first.

What does the video not confirm?

Mark also promotes Lockpick, an app-scanning service his team built, and offers a paid “Defensive AI Kit”. We have not tested either, and this article is not sponsored by him.

How should a small business start safely?

Start with the safest option that answers your question, and only move to uncensored models if you genuinely need them.

1. Write down what you want to test and confirm you own it or have written permission
2. Try mainstream and established tools first. Many security scanners and code reviewers work without an uncensored model
3. Make a test copy of your app with fake data, separate from your live system
4. Pick the route that suits you: hosted for speed, cloud for flexibility, local for privacy
5. Check any model before you run it (see our download checklist)
6. Cap your spending if you rent hardware (see our cloud GPU guide)
7. Verify every finding before you act on it, and fix the real issues first
8. Shut it all down and remove any keys when you are finished

Who are uncensored AI models right for?

They suit you if: you build or run your own apps, you understand basic security, you are testing in a safe copy, and mainstream assistants keep refusing reasonable defensive questions.

They matter less if: you mainly use AI for writing, marketing or admin, or you are not comfortable managing servers and permissions. In that case a professional security check, or a reputable scanning service, is usually the better spend.

Frequently asked questions

The bottom line

Mark Kashef’s video is a useful reminder that AI-built apps need proper security checks, and that mainstream assistants will not always help with them. Uncensored AI models can fill that gap for defensive testing, but they shift all the responsibility onto you: legally, financially and technically. Your next step: list the apps you own, make a safe test copy of one, and decide whether an established scanner or a professional check covers what you need before reaching for an uncensored model.

Related posts

• PHP2Shell: How to Check If Your WordPress Site Is at Risk
• GPT-7 Bel: OpenAI’s Offline Internet, Claude’s Reverse Engineering Block and This Week’s AI News
• Should You Run Muse Spark Open Locally or Use It Through an API?

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio