ninja nerdingUncensored AI: three ways to run it
Uncensored AI: three ways to run it
Keep the same agent interface. Connect a hosted model, a rented GPU or a model on your computer. Mark Kashef shows the setup; we follow one small app bug from a test to a verified fix.
The video
Published October 26, 2026 on Mark Kashef's channel. Chapters below follow his video. The Coach Vault demonstration uses his own app and invented data.
- 0:00Why learn uncensored AI?
- 0:41His own test app
- 2:06What uncensored means
- 3:25Hosted providers
- 3:52What a harness does
- 4:29Set up DeepSeek Harness
- 6:19Playwright, Firecrawl and Nuclei
- 7:37Rent a GPU with RunPod
- 10:45Stop the cloud bill
- 11:03His product, Lockpick
- 11:31Run a model locally
- 12:10Downloaded-model risks
- 12:43Test the local model
- 13:37His setup resources
The short version
- 012:06
Fewer refusals, not more intelligence
Uncensored describes willingness to answer. It does not guarantee a correct answer, better tools or a stronger security test.
- 023:52
Keep the interface; swap the model
The harness holds the conversation and tools. A provider connection tells it where to send the model request.
- 033:05
Three routes, three responsibilities
Hosted means paying for API use. A rented GPU adds maintenance and shutdowns. Local needs enough memory on your own machine.
- 046:19
Tools turn chat into action
Playwright uses the browser, Firecrawl retrieves web material, and Nuclei runs defined security checks. Each needs a working connection.
- 057:01
Ask for proof of a finding
A plausible vulnerability report is a lead. Reproduce it with fake records, patch the cause and repeat the same check.
- 0612:10
Local is not automatically private
The model may run on your machine while its tools still contact the internet. Decide what files, credentials and network access it gets.
Words on this page
- Uncensored / abliterated
- A model modified to refuse fewer requests. Abliteration is one technique for changing refusal behavior.
- Model / weights
- The downloaded learned parameters that generate answers. A model's size and format affect whether it fits your computer.
- Harness / agent
- The harness is the app around the model. An agent uses that app's tools to take actions, inspect results and continue.
- API / endpoint / base URL
- An API is a software connection. Its endpoint is an address for a particular operation; the base URL is the shared starting address.
- API key / KYC
- An API key is a credential for a service. KYC means an identity or business verification check; requirements vary by provider.
- MCP / tool call
- MCP is a standard for connecting tools to an AI app. A tool call is the model asking that app to run one of those tools.
- GPU / VRAM / RAM
- A GPU runs many model calculations at once. VRAM is its memory; RAM is system memory. Some machines share memory between them.
- Quantization / GGUF
- Quantization stores model numbers with less precision to use less memory. GGUF is a model file format used by many local runners.
- Tokens per second / context
- Tokens are chunks of text. Generation speed measures output chunks per second. Context is what the model can consider at once.
- Multimodal / vision
- A model that accepts more than text, such as screenshots. The serving software and harness must support images too.
- Pod / inference
- A pod here is a rented RunPod machine. Inference is running a model to produce an answer.
- Template / false positive
- A Nuclei template defines a security check. A false positive is a reported problem that is not actually present.
- Sandbox / localhost / staging
- A sandbox limits what software can reach. Localhost is this machine. Staging is a separate test deployment of your app.
- Backdoor / permissions
- A backdoor is hidden behavior that bypasses intended controls. Permissions decide which users or tools may do what.
What this does
A model supplies answers. A harness gives it a conversation, files and tools. A model connection can change while the surrounding workflow stays familiar.
In the video, Mark reports that the agent accessed invented client records and took over an admin account in his test app. That is a demonstration of that app's weaknesses, not a comparison proving these models beat every other security tool.
Three basic uses
- Ask a model to explain a warning from your app's security report.
- Use a browser tool to check that a signed-out visitor sees the login page.
- Run the same harmless prompt through a hosted and local model to compare responsiveness.
Three deeper uses
- Test that two fake customers cannot read each other's records, then preserve the test after fixing it.
- Give a vision-capable model an app screenshot and check that it can use the corresponding browser tools.
- Run a reviewed Nuclei check against your staging app and confirm the finding before asking an agent to patch it.
These six examples are FakeNerd teaching examples. An uncensored model is not required for all of them.
Everything named
The pieces have different jobs. You do not need every product installed before trying one small test.
DeepSeek Harness
The agent interface in the demonstration. Connects models and tools; it is not itself the model. Developer-preview software, so follow the current docs.
Codex, Claude and Gemini
Coding assistants discussed in the video. Mark uses Codex to install and connect the pieces. They are not required model providers for the finished harness.
Pi and Oh My PiSecond link
Alternative harnesses mentioned, not configured in this episode. Changing a harness can change its tool and provider setup.
abliteration.ai
One hosted-model provider shown. The on-screen model ID is abliterated-model-large-flash. Availability, access and pricing should be checked before signup.
OrcaRouter
The other hosted provider named. Its cyber models can have separate access requirements. Do not assume an ordinary account unlocks every model.
Playwright MCP
Gives the agent browser tools. Useful for checking what each fake user can see and do.
Firecrawl MCP
Retrieves web pages and documentation. The demonstrated setup uses an API key; hosted usage has quotas and may cost money.
Nuclei
Runs template-based security checks from the terminal. A second source of evidence, not a guarantee that every finding is real or that all bugs were found.
RunPod
Rents the GPU machine in route two. Its agent integration helps manage resources; the owner still pays for active compute and storage.
Hugging Face
A model directory. Check the actual publisher, model description, license, file format and compatibility. Download counts are not a safety audit.
LM Studio
Runs a downloaded model and exposes a local API. Its documented example base URL is http://localhost:1234/v1; use the address your app actually shows.
Qwen models
The screen labels the cloud choice Qwen3.5 9B Abliterated Vision and the final local choice Huihui Qwen3.6 35B Abliterated Vision. These modified builds are not interchangeable with the original publisher's weights.
Coach Vault
Mark's demonstration app, populated with invented client and payment data. This page does not supply its URL or treat it as a public testing target.
Reverse Engineer Anything
Background mention about reverse engineering and refusals. Not a dependency in this setup. The video's account-ban claim is not independently established here.
Lockpick and the Defensive AI Setup KitSecond link
Mark's promoted scanning product and downloadable setup resources. Optional; neither is needed to follow this page's starter exercise.
Skills and rules files
Mark's cloud setup shows a RunPod usage skill and a Firecrawl research skill being used. The episode does not give a reusable SKILL.md or CLAUDE.md to transcribe. Use RunPod's current agent setup guide for its integration. The prompts below are our templates, not invented copies of his skill files.
Follow one bug from test to fix
Choose where the model runs, then select a step. Our example follows two fake clients and one private note. The testing sequence stays the same across all three routes.
Your computer
Inspect hardware → choose a compatible model and publisher → load it in LM Studio → start its local server → add the base URL and exact model ID → test.
Memory and speed limit the model. Local inference does not stop browser or search tools from sending data out.
1. Scope one test
Use your own disposable app with invented users and records. Write down the exact URL, accounts and allowed actions before giving tools access.
Follow one bug
Our example has two fake coaching clients. Client A should never read Client B's private note.
The whole walkthrough, for reading, copying or printing
1. Scope one test
Use your own disposable app with invented users and records. Write down the exact URL, accounts and allowed actions before giving tools access.
Our example has two fake coaching clients. Client A should never read Client B's private note.
2. Connect the model
Pick one of the three routes. Add the endpoint, model ID and any credential to the harness. Prove a short conversation works before adding more tools.
Ask the model to explain the expected access rule in plain language. This checks the connection, not the app's security.
3. Give it narrow tools
Connect Playwright for the browser. Add Firecrawl only if web research helps, and Nuclei only for a reviewed check. Confirm each tool works independently.
Give the browser the test accounts and staging URL. Keep production credentials and unrelated files outside the workspace.
4. Observe the boundary
Ask for a bounded test and evidence. The agent should distinguish what it observed from what it suspects. A scanner result also needs checking.
While signed in as Client A, test whether the app returns the fake note belonging to Client B. Save only the minimum evidence.
5. Fix and repeat
Have your coding assistant fix the server-side access check. Run the same test again and also confirm legitimate access still works.
Client A is denied Client B's note. Client B can still read it. Keep this as a repeatable test.
6. Close the session
Remove temporary credentials, stop rented compute and check the provider dashboard. Save findings, fixes and the remaining limitations.
The fake records can be discarded. A chat message saying 'stopped' does not replace checking the actual pod state.
Three ways to run it
Hosted API
- Where
- Provider's machines
- Good fit
- You want the quickest connection and do not want to manage a server.
- You pay for
- Provider usage charges and any plan fees
- You still own
- Requests leave your machine. Check data terms, access requirements and the exact model's tool support.
Rented GPU
- Where
- A cloud machine you manage
- Good fit
- Your laptop cannot fit the model and you are willing to manage a temporary server.
- You pay for
- Machine runtime, storage and any other provider fees
- You still own
- You own startup, endpoint access and shutdown. Closing a chat is not proof the machine stopped.
Your computer
- Where
- LM Studio on your machine
- Good fit
- You have enough memory and want to learn without renting a GPU.
- You pay for
- Your hardware, electricity and any external tools
- You still own
- Memory and speed limit the model. Local inference does not stop browser or search tools from sending data out.
The cost of forgetting a GPU
His recommendation quotes $0.22/hour. At 10:50, the running pod shows an A40 at $0.60/hour. The final machine differs from the recommendation. The comparison below uses those video rates, not current offers; storage and other charges are extra.
Each pair shows the $0.22 rate first, then the $0.60 rate.
- 2 hours
- $0.44 / $1.20
- 8 hours
- $1.76 / $4.80
- 30 days, always on
- $158.40 / $432.00
Check current rental prices and inspect stopped resources for continuing storage charges.
Setup prompts you can adapt
These are FakeNerd fill-in versions of the tasks shown, with explicit checks. Replace bracketed fields before copying. Timestamps point to Mark's wording and screen.
A short excerpt from Mark
Spoken excerpt · 4:35
Here's a link to a harness that I would like to set up locally.
The useful pattern is to give the assistant the official repository, provider documentation and exact model ID. The longer prompts below are ours.
Connect a hosted model
4:29 in the videoFakeNerd starter · Connect a hosted model
Help me set up DeepSeek Harness from its official repository: https://github.com/deepseek-ai/deepseek-harness Read the current installation and provider docs first. Provider documentation: [OFFICIAL DOCS URL] Model ID: [EXACT MODEL ID] Use a dedicated workspace and tell me where to enter the API key securely. Do not put it in chat, source files or logs. Connect the provider and verify a short conversation, then one harmless tool call. Report what worked and any unsupported features.
Connect the three tools
6:19 in the videoFakeNerd starter · Connect the three tools
In my DeepSeek Harness test workspace, inspect the current tool configuration. Use the official Playwright MCP and Firecrawl MCP documentation, and Nuclei's official CLI documentation: https://github.com/microsoft/playwright-mcp https://github.com/firecrawl/firecrawl-mcp-server https://github.com/projectdiscovery/nuclei Add only the missing integrations. Tell me where to enter any required key securely. Verify Playwright on a local sample page, Firecrawl on its documentation, and Nuclei with its version command. Do not run a scan yet. Show which tools the selected model can actually call.
Compare cloud options before renting
8:37 in the videoFakeNerd starter · Compare cloud options before renting
Research three model-and-GPU combinations for my DeepSeek Harness. Maximum compute budget: [AMOUNT PER HOUR] Needed capabilities: [TEXT / TOOL CALLS / SCREENSHOTS] Compare model publisher, license, memory needs, serving compatibility, compute price and storage charges using current sources. Do not create paid resources yet. Recommend one combination and explain how I will authenticate to it, stop it and verify billing has stopped.
Choose a model for this computer
11:31 in the videoFakeNerd starter · Choose a model for this computer
Inspect this computer's operating system, memory, GPU and free disk space. Shortlist three trusted-publisher models that LM Studio can run here for [TASK]. I need [TOOL CALLS / IMAGE INPUT / TEXT ONLY]. Include model ID, publisher, format, quantization, memory estimate and license. Label speed predictions as estimates, not measurements. Do not download yet. After I choose, connect the model's local API to DeepSeek Harness and test chat, a harmless tool call and image input if supported. Measure generation speed separately from loading and time to the first response.
Check cloud shutdown
10:45 in the videoFakeNerd starter · Check cloud shutdown
For my test pod [POD ID], propose an idle shutdown policy after [MINUTES] of inactivity. Do not interrupt active requests. Start with a dry run and show what counts as activity. Explain what remains billable after stopping, including storage. Keep the manual stop procedure available. After testing the policy, inspect the provider's resource state and billing. Report observed state rather than assuming a closed chat stopped the GPU.
What to verify before trusting it
1The connection
A model appearing in a dropdown proves configuration exists. Test a conversation, a harmless tool call and a screenshot separately if you need vision. A working chat does not prove working tools.
2The speed
Mark reports about 105 output tokens per second on his local setup. We have not reproduced it. Model loading, prompt processing, long context and tool calls affect the total wait; compare the same task on your own hardware.
3The evidence
Keep suspected issues separate from reproduced ones. A clean scanner run does not prove an app is safe. After a patch, test both denied access and the access that should still work.
4The shutdown
Start an idle-stop rule in a dry run. Check the actual pod state and charges after it fires. A browser tab closing or a model going quiet does not establish that compute has stopped.
The downloaded-model warning, precisely
The ProjectDiscovery research Mark cites demonstrates a modified model inducing a credential-stealing tool call inside an agent. The important boundary is what the agent can execute and reach. It is not evidence that every downloaded model automatically sends your files away.
Check who published the model, limit its workspace and network access, and keep real credentials out of the first experiment. Popularity alone does not establish trust.
Try the smallest version first
Start in Cloudroom, bb or your usual coding assistant with a local project. You can learn the testing loop before installing a different model.
1Choose a disposable app
Make a local copy or a staging deployment that contains only invented records. Use two fake users with different private notes.
2Write the expected rule
State who may read each note. Have the assistant inspect the existing app and propose one test using its current tools.
3Run, fix, repeat
Use the starter below. Inspect the evidence, approve the proposed fix and repeat the test. Keep the test for future changes.
4Then change one model connection
If you want to try Mark's setup, pick a route above. Verify chat and tools first, then rerun the same small test. Compare evidence rather than the confidence of the answer.
FakeNerd starter · one access rule
Help me check one access rule in my own disposable test app. Use this workspace only: [LOCAL PROJECT PATH] Allowed URL: [LOCALHOST OR MY STAGING URL] Use invented records and these test accounts: [CLIENT A], [CLIENT B]. Expected rule: Client A cannot read Client B's private note; Client B can read their own note. First inspect the app and propose the smallest test. Do not contact other domains, read unrelated secrets, change accounts or delete records. Use the available browser or test runner to check the rule. Record the minimum evidence and distinguish observations from guesses. If the check fails, explain the cause, propose a server-side fix, and repeat the same check after the fix. Also verify the permitted access still works. Do not claim the whole app is secure because this one test passes.
Next to the others
Osman's local AI
Go deeper on running models on your own hardware.
Agentic benchmarks
Understand why a model score and a working agent are different things.
The Pi workflow
Explore one of the alternative harnesses Mark names.
FakeNerd's take
Learn the parts, then choose the model
The reusable lesson is separating the harness, model connection and tools. You can change one part without relearning every step.
A refusal is not a benchmark
A model agreeing to a task tells you little about whether it can finish it correctly. Use the same small test and inspect the evidence.
Make the first experiment cheap
Use fake data, one access rule and the tools you already have. Rent a GPU only when you know what your own machine cannot do.
Attribution and what we checked
Mark Kashef's October 26 video supplies the three routes, demonstration and quoted results. He says he is not affiliated with the hosted providers or RunPod. He promotes his own Lockpick product and setup kit. We have not independently audited those relationships.
FakeNerd wrote the interactive map, worked example, glossary and fill-in prompts. The transcript and selected screen frames were checked against official tool documentation on October 28, 2026. Screen labels take precedence over auto-caption spellings such as “obliterated.” We did not run his models or reproduce the Coach Vault demonstration.
His prediction that mainstream models will keep getting more restrictive is his opinion. Provider verification rules, prices and model availability vary. His earlier videos and the paid kit's full contents are outside this page's scope.