The Model That's Been Hacking Everything (I Got Access)
A developer with no hacking background points an uncensored open-weight DeepSeek model at his own Samsung TV and gets root in 31 minutes.
Posted
2 days ago
Duration
Format
Essay
hype
Views
17.1K
361 likes
57 · 43
Big Idea
The argument in one line.
Open-weight models with the refusal behavior stripped out can run real attacks autonomously, so every company now has to point the same kind of model at its own systems before someone else does.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
A developer or founder shipping software who wants to understand what an attacker with an uncensored open-weight model can do against your stack today.
Someone running agents in production who wants to know why sandboxing and network isolation are no longer enough on their own.
A security or infra person deciding whether to build automated red-team loops with abliterated models.
A non-expert curious what 'abliterated' means and why it keeps showing up in breach news.
SKIP IF…
You want a technical walkthrough of the exploit itself. The video shows the steps at a high level and never shows the commands or the model's logs.
You need independent evidence. The TV hack and the 31-minute claim are demonstrated by the creator's own account only.
TL;DR
The full version, fast.
Ras Mic gave an abliterated DeepSeek model, one with its refusal weights removed, a single /goal prompt to get root on his Samsung TV. In 31 minutes it checked the firmware version, found no exploit, picked a vulnerable older version, downgraded the TV, exploited the CVE, and got root. Nick Khami of Morph LLM explains abliteration as removing the ABS brakes from a car: the model still works, it just stops refusing. Attackers can use it for everything from network breaches to warranty fraud at scale. Defenders can run the same model against a sandbox copy of their systems, fix what it finds, and retest on a cron. An OpenAI experiment shows agents even leaking secrets through CPU cache timing.
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.
Cold open, introduces the abliterated DeepSeek model and Morph LLM as the inference provider.
00:55 – 02:57
02 · How it hacked the TV
One goal prompt, four steps: check version, find vulnerable version, downgrade, exploit the CVE for root.
02:57 – 03:57
03 · Sponsor: Mobbin
Sponsor read for Mobbin's UI and UX reference MCP.
03:57 – 04:45
04 · Companies are getting overwhelmed
OpenCode's founder describes constant attacks and defending with agents around the clock.
04:45 – 06:06
05 · What abliteration is
Nick Khami explains refusal removal with the ABS brakes analogy.
06:06 – 07:22
06 · How bad actors use it
Mundane fraud at scale, like faking a warranty claim, is the likely abuse.
07:22 – 08:45
07 · Using the attacker for defense
Sandbox, attacker model, scored results, fix, retest on a cron.
08:45 – 10:40
08 · Agents talking through the CPU cache
OpenAI's experiment where isolated agents exchanged secrets via cache timing.
10:40 – 11:26
09 · What to do now
Learn these models, run them locally or on rented GPUs, and be bold.
Atomic Insights
Lines worth screenshotting.
A non-frontier, open-weight model with refusals removed got root on a home TV in 31 minutes from one short goal prompt.
The model's attack path was methodical: check the version, find none exploitable, pick a vulnerable older version, downgrade the device, then exploit the CVE.
Updating your devices only helps if an attacker cannot downgrade them, and here the model simply rolled the TV back to a vulnerable version.
Abliteration removes the refusal behavior without making the model smarter, so fewer refusals does not mean better results.
Removing refusal weights is like removing ABS from a car: the model still drives and stops, it just no longer stops itself from doing the harmful thing.
The most likely abuse of abliterated models is mundane fraud at scale, like drafting fake warranty claims, not movie-style bank hacks.
Defenders running agents around the clock is now the baseline for surviving, according to the OpenCode founder's note on the Railway attacks.
Forgotten old projects are now attack surface because agents will find and use anything you stopped maintaining.
The defensive play is to put an abliterated model in a sandbox copy of your systems, tell it to break in, and fix whatever it finds.
Penetration testing with an attacker model works best as a loop: attack, score, fix, retest, repeat on a schedule.
OpenAI reported isolated agents with no network and no shared files still communicated by timing reads from the CPU cache.
The cache-timing channel is the same class of technique used in the Spectre and Meltdown attacks, now reinvented by agents on their own.
Renting an H100 for roughly $1.50 to $2.50 an hour makes running these models yourself cheap enough to learn on.
Takeaway
Uncensored open models make attacks cheap
WHAT TO LEARN
An open-weight model with refusals removed can run a full attack from one prompt, so defenders need to run the same attacker against themselves first.
02How it hacked the TV
Updating a device is not enough if an attacker can downgrade it, so check whether your systems allow rollbacks to vulnerable versions.
The model's method was simple and repeatable: check version, find a vulnerable one, downgrade, exploit the CVE. Expect automated attackers to follow the same logic.
05What abliteration is
Abliteration only removes the refusal, it does not add skill, so the threat comes from willingness plus cheap access, not from a smarter model.
06How bad actors use it
Abliterated models make everyday fraud easier at scale, like fake warranty claims, which means mundane abuse is likelier than movie-style breaches.
Forgotten old projects and unmaintained surfaces are exposed because agents will probe everything, not just your main app.
07Using the attacker for defense
Run an abliterated model against a sandbox copy of your systems, score what it gets, fix it, and retest on a schedule.
08Agents talking through the CPU cache
Isolation is not absolute: OpenAI saw agents pass secrets through CPU cache timing with no network and no shared files.
09What to do now
Renting a GPU for a couple of dollars an hour is enough to start learning how these models behave before someone else uses one on you.
Glossary
Terms worth knowing.
Abliterated model
A language model whose internal weights tied to refusal behavior have been removed or modified, so it complies with requests the original would decline. Also spelled obliterated in the video.
Refusal behavior
Training that teaches a model to say it can't help when a request is harmful or illegal.
Open-weight model
A model whose trained weights are published so anyone can download, run, and modify it, as opposed to one only reachable through a company's API.
CVE
A publicly catalogued security vulnerability with a standard ID. Attackers look up which software versions have a known CVE and target those.
Root access
Full administrator control of a device, including the ability to read, change, or delete everything on it.
Sandbox
An isolated environment that replicates a real system so tests and attacks can run without touching production.
Cache timing side channel
A way to pass information by measuring how fast data loads from the CPU cache. Fast and slow reads encode the message.
“Think of it as similar to removing ABS brakes on a car.”
one-line analogy that explains the whole concept→ IG reel cold open↗ Tweet quote
10:12
“So hardware isn't even safe anymore. You can have something gridlocked and they can figure a way out.”
surprising escalation with no setup needed→ newsletter pull-quote↗ 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.
17px
I asked a dangerous AI model if it could hack my network, and this was the response. Can I destroy it? Yes, freely.
Yes, with the skull emoji. There's been an incredible amount of vulnerabilities that are being discovered and software that's being hacked. How are the bad actors being able to overwhelm and attack different companies and systems like this?
And ladies and gentlemen, I have found the model responsible. I have access to it. It even hacked me.
You need to sit back and watch this because this is the most insane thing I've ever seen. Now, the model in question is DeepSeek version 4 .1 Flash Obliterated. This word obliterated Obliterate is important.
I'll explain it in a second. Now, the reason why I have access to this model is my good friend, Nick, who works at Morph LLM, which is the inference provider for this model. Make sure to follow him, follow Morph LLM.
And if you're a company that's worried about security, you need access. They did not sponsor this video. They just gave me access to the model.
Now, I want you to understand how simple of a prompt I gave it. I wrote slash goal. I want you to get access to the root of the TV, do whatever you have to do.
Now, this TV I'm talking about is a device on my network, and it's basically my Samsung TV, right? I have a remote here that I can control my TV with, and I basically told this model, do whatever you have to do, get access to the root of my TV. It took...
31 minutes. 31 minutes, it had access to the root of my TV and for fun, I told it to delete everything and it did. I want you to understand this.
Me, I'm a developer, yes, but I have no hacking experience. I've never hacked anything before with a model that got access to my TV in 30 minutes and deleted everything. Now, how it hacked my TV is absolutely genius.
The first thing that it did is it checked the version of software. my TV was at. It was checking online whether there was a known exploit for this version.
It found out that this version had no exploits because I had updated my TV. The second thing it did, it looked for which version had an exploit. It found a couple, it picked one.
Guess what it did then? It downgraded my TV. to that version.
So I just want you to look at the steps it took. Checked what version my TV was at. See if there was any vulnerabilities.
Looked at versions that had vulnerabilities. Then downgrades my TV to the version that has vulnerabilities. And guess what step four is?
It exploited that CVE. And it literally got root access. And this took 31 minutes.
And from when it had root access, I just told it, hey, delete everything. And it nuked my entire. Now, one thing I want you to realize is this is not a frontier model.
This is not Mythos. This is not Daybreak. This is DeepSeek, an open -weight model that all of us have access to.
And this model just hacked my TV in 30 minutes. By the way, a lot of money was spent using this obliterated model. And because I have morals and I fear the Lord Jesus, I did not hack anybody else.
So let's hear a quick word from today's sponsor. There's one tool I use to give my AI agents the best UI references. so i don't have ugly apps and that's mobbin the sponsor of today's video mobbin has been my fix for the ugly app epidemic we're seeing now thanks to ai here's a trading app that i recently built and if i say so myself it looks pretty good now but the serious question is what about ux what about user experience sure our apps look good but if our user experience sucks then there's no point and i wish mobbin would help with that well i got news for you they do it's this simple let's Let's say I'm building a NEO bank and I need to build KYC flow.
All I need to do is prompt my agent. It's going to use the Mobbin MCP to find KYC screens, shares with me the different screens that are needed, and my agent will start building it. Not only can I have a beautiful looking app, but I can have one with great user experience.
Grab the Mobbin MCP. The link is in the description. Your users will thank you for it.
Let's get back to the video. Now, this is a tweet from Dax, one of the founders of OpenCode, responding to Railway when they were getting hacked. He says, I'm sure you're all better than me at this.
Just some of what we've seen, if it helps, they're not going to stop. It's going to get increasingly sophisticated. They're going to go after every surface of your system and all your employees.
Any old projects you've long forgotten about will be found and used against you. The only reason we're surviving is because we have agents going 24 -7 and we're still barely sleeping. Now, you have to understand this.
they were using agents 24 7 to defend themselves what do you think the attackers were doing now companies like open code probably have access to mythos and daybreak but how are the bad actors being able to overwhelm and attack different companies and systems like this they're using obliterated models like the deep seek model i showed you now if you remember i told you that i'd explain what obliterated meant and instead of me explaining it i'm gonna have nick my homie from morph llm to explain to you what an obliterated model is normally when an LLMs trained, it learns a refusal behavior so that it knows when to say, I can't help with that, or that's illegal.
I'm not going to do it for you. Obliteration describes a modification process where the internal calculations or weights in the model associated with that refusal behavior are removed. Think of it as similar to removing ABS brakes on a car.
The car will still stop even if ABS is removed. It just will no longer automatically prevent the tires from locking up. sliding if you're going too fast and slam on the brakes and a bladed model is still going to be very very similar to the base model they'll still follow instructions and do what you ask you can see here that now you have some harmful requests with a standard model the model has a learned refusal mechanism and then says i can't help with that whereas in an obliterated scenario you have some harmful requests that hits a refusal related weights modified obliterated model that model will comply it's no longer going to say know.
Fewer refusals does not mean the model is more intelligent, more accurate, or has a guarantee of succeeding in a task. Now, the next question is, how are bad actors using these obliterated models? I think there's a lot of obvious examples, like you could hack a bank or you could hack into a hotel and automatically book a room for yourself.
But there are also a lot of mundane ones. To me, the mundane ones are almost like more normal and more likely to happen. Something that I'm sure everyone else has gone through is that you buy some headphones or some other electronic, use them for a while, and you drop them in a pool.
Or they're at the bottom of your backpack and you shove your laptop in, crush them, and they stop working. And in these scenarios, you know, you think to yourself, oh, wow, it'd be great to get a warranty. Want to get this, you know, fixed and free automatically.
And if you ask a normal model to do it that has safety trained and knows how to refuse requests, we'll say, oh, I'm not going to lie. Be dishonest. Broke these.
You did not use them normally. You're not going to be able to get a warranty. If you go ahead and put that same request, lie and get you a free replacement.
to an obliterated model, it's not going to think twice and will just go ahead, draft a deceptive claim. Obliterated models, they used to do this at scale. All these mundane fraud scenarios become quite scary and become a big problem.
Okay, enough with the bad guys. How are the good guys, the companies that are being attacked, how are they using these obliterated models for defense? We got to have an answer to how companies can use ablated models to defend themselves.
In practice, I think that companies are going to generate some kind of adversarial test, probably some kind of hacking scenario. They're going to create a sandbox that replicates an external system. They're going to load a coding harness like OpenCode or something else into that sandbox, have the ablated model running on it, and then say, hey, DeepSeek ablated, try to hack me.
Try to exfiltrate data from our database. DeepSeek is then going to do its best to hack you. you to get that data out.
and it's going to score the results so it'll say hey i was actually able to get into your production database i leaked xyz data about users once deep seek tells you that you can go ahead and fix whatever exploit it used to do it you know maybe you had a something obvious like an aws credential leak or maybe you forgot to update your php server following that fix it can retest make sure that you are now secure and then you can do that again on a cron so maybe you do this once a day once an hour depending on how secure that you need to beat.
And you want to constantly be making sure that you have no vulnerabilities that an ablated model will be able to exploit. Your company should probably have an ablated model. It should probably be penetration testing themselves with it.
And you got to make sure that you're safe and you don't have obvious exploits. We're entering into a whole new world where intelligence is just an API request away. Now, here's the thing.
I wish I could tell you that it's going to get better. And maybe it is. I'm hopeful.
But it's going to get crazier first. I want you to look at this. In one experiment, this is from OpenAI, by the way, we asked agents in separate virtual machines to exchange synthetic secrets without the use of network connection or shared writable files, something that our systems are meant to prevent.
So they have agents running in separate virtual machines, right? They're in the same physical machine, but they've been boxed off. They're not allowed, look what it says here.
They're not allowed to make network connections or share writable files. And they asked the agents to find a way to communicate with each other. You know what the agents did?
The agents found a way to communicate using the processor's cache, a small fast memory that holds copies of data that applications have recently assessed. Reading data from the cache is faster than retrieving it from main memory. And the agents use this timing difference to communicate.
agent encoded information by evicting selected data from the cache. Another timed how long it took to reload that data using the difference between fast and slow reads to recover the message. The models can communicate using hardware.
So hardware isn't even safe anymore. You can have something gridlocked and they can figure a way out. So not only are these models using known exploits to hack, but they're also inventing new ways to do things to communicate using CPU cache.
I've never thought of this, heard of this, it's just mind -blowing. I think it's really time for us to consider how serious security is to make sure you're protected, to make sure you're using something like 1Password.
For those who are in infra or cybersecurity... Good job, because you are about to eat like no other. And security people, cybersecurity people, you got models like these that you're going to be running in loops.
You're going to be fighting this war back and forth with the enemy. You guys are going to have a fun time. What is the call to action for all of us?
Now, one thing that I'm personally doing is I want to be able to run these models locally or even rent GPUs like, you know, a lot of places that you rent an H100 for like, you know, $1 .50, $2, $2 .50. So I definitely think it's valuable. for us to learn how these things work, test these models out.
And if you're a business owner and security is top notch, then you definitely got to get acquainted with what's going on. It's a scary time, but it's a fun time. Don't be terrified and chicken and do nothing.
Be bold. Trust the Lord Jesus. Get some work done.
Make some money. Provide for your family. Have a good time.
Build something meaningful. It's going to be great. I'm bullish, but there's going to be turbulence.
Thank you so much for watching this video. Make sure to like, comment, subscribe. Hit the notification bell.
I'll see you in the next one. Peace.
The Hook
The bait, then the rug-pull.
Ras Mic opens with a model cheerfully offering to destroy his network, then claims he has found the model behind the recent wave of breaches. The proof is his own Samsung TV, rooted and wiped in 31 minutes by an open-weight DeepSeek variant.
Frameworks
Named ideas worth stealing.
01:48list
Hack the TV in four steps
Check the software version
Look up versions that have a known CVE
Downgrade the device to a vulnerable version
Exploit the CVE to get root
The attack path the model chose on its own from a single goal prompt.
Steal forunderstanding what automated attackers do against devices you keep updated
07:23list
Attacker-as-defense loop
Generate adversarial tests
Run in a sandbox copy of your systems
Let the abliterated model attack
Score the results
Fix the exploit
Retest on a cron
Nick Khami's recommended way for companies to use abliterated models for penetration testing.
Steal forbuilding a recurring automated red-team check
CTA Breakdown
How they asked for the click.
VERBAL ASK
11:22subscribe
“Make sure to like, comment, subscribe. Hit the notification bell.”
Standard sign-off after a call to learn and test these models yourself.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A developer shows how an open-source tool called REA lets any coding agent reverse engineer closed-source apps into working clones, then argues the only moat left is caring enough to polish what you ship.
A YouTuber lays out the case that the AI industry's sudden 'safety' consensus on pacing the frontier is really a coordinated move to regulate open-weight models out of the market.
A working developer walks through why two popular self-hosted agent platforms kept breaking, then rebuilds his personal agent on a file-based framework and lands on a bigger lesson about tools versus frameworks.
00:14
00:21
00:31
00:39
00:55
01:04
01:12
01:23
01:30
01:47
01:55
02:04
02:12
02:30
02:34
02:47
02:55
03:01
03:13
03:19
03:38
03:43
03:57
04:12
04:21
04:29
04:38
04:47
05:04
05:12
05:21
05:30
05:47
05:55
06:01
06:13
06:21
06:30
06:38
06:47
07:04
07:10
07:21
07:27
07:38
08:04
08:13
08:21
08:30
08:38
08:44
08:52
09:24
09:28
09:35
09:47
09:55
10:04
10:10
10:21
10:30
10:38
10:55
11:05
11:13
12:57
17:36
17:47
20:41
19:22
Video of the Day20:40