Running AI Agents Locally = Safe...? Think Again
490 segments
Running AI on your machine does
not make it private.
Most people think
local means safe,
but if your machine is
connected to the internet,
you basically
bought your own house
and left every window open.
Here are the seven ways that
your private AI
is still leaking data and
exactly how to lock it down.
In my previous business,
I used to overcharge
people for this stuff
and made me just sick
of it, so I had to quit.
Now I'll just show you how to
build it yourself.
My last video on local AI setup
got a ton of comments
and the same questions just
kept coming up.
Is local still local
even when my machine
is connected to the internet?
Let's put it this way.
Using cloud AI like
ChatGPT or Claude
through their
website, their platform,
is like renting an apartment.
Your landlord has a
copy of your keys.
They can walk into your unit
whenever they feel the need to (ex. emergency..)
They might look
through your stuff.
You have to trust
that they won't.
Running AI locally is buying
your own house.
No more landlords,
no more shared keys.
That is real.
That matters.
But here's the
part most people skip.
Owning a house does
not mean it is secured.
If you leave the windows open,
the back door unlocked,
and your mail gets forwarded to
your old address,
you are exposed.
Different kind of
exposed than renting,
but you are still exposed.
And right now, there are three
levels of owning your house
that most people mix up.
Level one, local.
You own the house, but it's
still connected to the street.
Internet is on.
Updates are flowing.
Plugins are active.
This is where most of you are.
Level two, offline.
You own the house and you cut
the phone line.
No internet during use,
but you still go online
sometimes to
download new models
or get updates.
level three, air-gapped
You own the house and it is on
its own island.
No internet connection ever.
everything gets carried in by
hand on a USB drive.
Most people think that they are
at level three.
They are at level one and level
They are at level one and level
one has seven open windows.
Let me show you each one and
how you can close it.
The first one is the scariest
in January 2026 two months ago.
in January 2026 two months ago.
Security researchers scan the
internet and found over 175,000
local AI servers completely
exposed to the public internet.
That is very bad.
Not cloud servers,
personal
machines, people running
Olama, LM studio, VLLM, their
private AI was just wide open
for anyone on the
internet to use.
But how?
One setting.
when you run a local AI tool,
it creates a little
server on your computer.
Think of it like a door by
only opens inward.
Only you can use it.
That is called
binding to local host.
But a lot of people change one
setting so they can access
their AI from their phone or
another computer
on their Wi-Fi.
Now it faces the street.
And there's no lock on it.
Attackers can just walk in, run
prompts on your machine,
download your models, and in
some cases, run code directly
on your computer.
Here's what you
should do instead.
Keep your AI server
bound to local host.
If you need access
from other devices,
put a lock gate in front of it.
That means a reverse
proxy with a password.
and add a firewall rule that blocks
anyone outside
your home network.
Just five minutes, that window
is closed and sealed.
This one is happening right now.
In March 2026,
which is this week
Microsoft published a security
report about browser extensions
that were specifically built to
steal AI chat histories.
Here's how it works.
You are running your local AI
through a web
interface in your browser.
You think it is private because
the AI is just on your machine,
but you have browser
extensions installed.
Some of those extensions have
permission to read everything
on every page you visit.
That includes your local AI
chat window as well.
The extension reads your
prompts and your AI's response
right off the screen.
else's server in
the background.
You just never see it happen.
Think of it this way.
Your house is locked, but you
have a mail slot in your front
door and someone is just
sitting outside,
reaching through the slot and
reading every letter before you
even get a chance to.
You can create a separate
browser profile just for AI.
Zero extensions, there's
nothing installed.
Use the profile only when
you're working
with your local AI.
This takes just two minutes to
set up and it
completely kills this attack.
those first two, you can fix it
in five minutes.
The next ones are sneakier
because you didn't
actually set them up.
They were already running
before you even
installed anything.
Here's the one that makes
people's stomachs drop.
When your local AI runs, it
saves your conversations, chat
logs, prompt history, sometimes
full documents you fed into it.
Those files set in
folders on your computer.
It's pretty normal so far.
But most computers have
cloud sync
turned on by default.
Your iCloud, OneDrive, Google
Drive, Dropbox, these services
automatically copy your files
to their servers.
desktop and documents folder
without asking.
So your local private AI
conversations might already be
sitting on Apple servers or
Microsoft's or
Google's right now.
You bought your own house to
get away from the landlord, but
your filing cabinet is still
automatically faxing copies of
everything to your old landlord's office.
Create a dedicated folder for
all your AI work.
Call it something like AI
workspace or whatever.
Then go into the sync settings
and exclude that folder.
Make sure it never sinks.
Keep your prompts, your chat
logs, your RAG documents and
your model files
in that folder only.
Nothing in there should ever
touch the cloud.
This one is about the
AI models themselves.
When you download a model to
run locally, you're downloading
from your downloading a you're
downloading a file from the
Internet, usually from a site
called like Hugging Face.
It is like an app store for AI
models, open source models.
Most of those models are safe,
but researchers have found that
some models files have been
tempered with a bad model.
A bad model file can
do two things here.
First, it can contain hidden
sneaky instructions baked into
the model's template.
When you load it, those
instructions silently change
how the AI behaves
without you even knowing.
Second, and I
think this is worse.
Some model file formats can run
code on your computer the
moment you just opened them.
When you chat with the model,
when you load it, there are
real documented
vulnerabilities for this.
Think of it this way. You
bought your new lock.
for your front door
But you bought it from a
complete
stranger on the street.
And that stranger might have
kept a spare key.
Here's what you should do
instead. Now you're going to
only download models from
verified trusted publishers.
Learn to prefer the safetensor
file format over older formats.
Safetensors was specifically
designed to
prevent code execution.
And always check the file hash.
This is like a fingerprint that
proves the file has not been
tempered with during download.
Now this one is about your
operating system, your OS.
Windows, Mac OS, they all
collect data about how
you use your computer.
They usually just call it
diagnostic and telemetry
Some of are just basic
But some of it goes deeper than
you think. On Windows, crash
reports can include a snapshot
reports can include a snapshot
of what was in your computer's
memory at the
time of the crash.
If your AI was running and it
crashed, or your system
crashed, or your system
crashed, fragments of your
prompts could be
inside that crash report.
and that report
gets sent to Microsoft.
Windows also has a feature call
recall that takes regular
screenshots of your screen,
if that is on
you're using your local AI.
Microsoft is essentially taking
photos of your
private conversation.
So it's basically... So it's
basically your house has
cameras you did not even
install. And they are
pointed at your desk.
So here's what you should do
instead. Here's what you should
do instead. I'm going to need
you to go into your privacy
settings and
minimize your diagnostic.
Turn off the optional
telemetry. Turn off recall if
it is even on. On Mac OS, on
Mac OS, check your analytic
settings and avoid...
and always avoid
submitting crash reports during
sensitive AI session.
Here is the last one. I think this is less
technical, but I do think it
technical, but I do think it
can trip off a lot
of business owners.
If you handle
customer data, client files,
or anything with
personal information,
moving your AI to
local does not erase
your legal obligations.
In fact, I was also not aware
that this could
be this twisted.
GDPR still applies...
HIPPA still applies...
CCPA still applies...
The rules follow the data,
not where your computer sits.
in fact, here's the twist.
When you use Cloud AI,
the provider shares some of the
liability with you.
They have security team
compliance program,
encryption standards.
When you go to
the local though,
all of their
liability shifts to you.
You're now the entire
security department.
Every access control,
every audit log,
every encryption standard
is your responsibility.
So, owning your
house does not actually
change the building code.
You still have to follow them.
And now there's no building
management company
handling that for you.
There's no more Strata.
So, here's what I
think you could do.
I think you should
treat your local AI
as a production server,
almost, not a toy.
So, that's starting with
encrypting your drives,
setting up the access controls,
also knowing what
logs you have to keep
and what's in them.
if you handle regulated
sensitive data,
you gotta get clear on what
specific obligations are.
So, look them up.
This is not legal advice,
but it's a real conversation
you need to have.
Seven windows open.
Here's how to close all of them.
this weekend, 10 steps.
One, bind your AI
server to local host.
Never 0, 0, 0, 0.
Two, add two,
add a firewall rule
that blocks outside connection
to your AI port.
Three, create a dedicated
browser profile for AI,
no extensions.
Four, create a no sync folder
for all AI files that you have
excluded from iCloud, OneDrive,
and Dropbox, and much more.
Five, only download models from
verified publishers,
and learn to prefer
safetensor format.
Six, go into your OS
privacy settings,
minimize the diagnostics,
disable the optional telemetry.
Seven, you gotta
encrypt your hard drive,
and always have backup.
full disk encryption,
turn it on if it's not already.
Eight, turn off any
plug-ins for connectors
you are not actively
using in your AI tools.
Nine, nine, nine,
check your AI tool setting for
telemetry or analytic toggles.
Analytics toggles.
Opt out of everything you can.
Ten
if you handle client data,
write down your AI data policy.
The data processing document,
where data goes,
who has access,
how long you keep it,
and what exactly happens if
something leaks.
That's it. Ten steps.
One weekend, now your
house is fully locked.
Obviously, local AI is much
better than renting.
I believe that.
But local is not a
magic shield yet.
It is a better
starting position.
You still have to lock the
doors, close the windows,
and check who has the keys.
By the way, I put that full
ten-step checklist in a PDF.
You can download it for free.
There's no gatekeeping.
Link in the description.
You can just print it out.
Although I do feel like only two out of ten people
two out of 10 people are going
to actually do this,
but if this changed how you
think about your set up,
why don't you drop a comment
and tell me which
which windows surprised you
the most. And if there's
if there's any windows that you
think I may have missed,
please let me know in the
comments as well
I'm curious.
I read everyone.
I still have to get back to the
comments that I got
from the other videos.
I will do that today
I will see you in the next one
Bye
Ask follow-up questions or revisit key timestamps.
Running AI locally does not automatically guarantee privacy if the machine is connected to the internet. This video explains why local AI setups are often exposed to vulnerabilities and provides a 10-step guide to properly secure a local AI environment, emphasizing that owning the hardware requires taking full responsibility for security, data management, and legal compliance.
Videos recently processed by our community