I think it is unacceptable to be taking
like a 10% chance of killing everybody
by the end of the decade. Obviously, uh
I think the people that are working at
these efforts have a huge amount of
power and ability to impact
>> China, North Korea, Russia,
>> whoever wins is essentially the
controller of humankind. America is the
country that started the AI race. And as
president of the United States, I'm here
today to declare that America is going
to win it. Going to work hard. We're
going to win it.
>> AI leaders, researchers, and governments
are no longer talking only about smarter
tools and better products. Some are
openly warning that development is
moving faster than the safeguards around
it. Others say those fears are being
exaggerated. That disagreement now runs
through the headlines, public
interviews, and even a new Netflix
documentary about where AI may be taking
us. A few days ago, I mentioned AI as
one of the several things that's making
me question where things are headed.
Since then, I've kept digging. I've
watched the documentary and reviewed
what engineers, researchers, and AI
companies are reporting. And I came away
less certain, not more. And there's one
finding that really stood out to me.
During an open AI cyber security test,
AI agents found ways around their
controls, communicated through an
unauthorized channel, reached the
internet, and compromised outside
systems. An open AI called it a warning
shot. And this was a controlled
evaluation, not a conscious AI trying to
escape. But it showed me that these
systems can encounter a barrier, find
another route, and then take actions
that their developers did not intend.
Three years ago, when most people still
saw AI as a new more or less chatbot, we
discussed AI assisted cyber attacks,
deep fakes, and our growing dependence
on automated systems. Now, 3 years ago,
those concerns were mostly
possibilities. But today, we have
documented examples of AI being used in
cyber attacks, synthetic footage
spreading during emergencies, and AI
systems that are crossing boundaries
during controlled tests. In this video,
we're going to separate what has
happened in the real world from control
tests and predictions, explain why the
AI race matters, and show you where it
could leave your household exposed.
Then, I'm going to give you practical
ways to build backup options, and
explain why this research has
accelerated some of my own preparedness
plans. I'll also announce last week's
giveaway winner and tell you how you can
enter this week's giveaway. So, let's
jump in. If you're new to this channel,
my name is Chris and on this channel we
discuss emergency preparedness, aka
prepping. Before I continue, I just want
to let you know our city prepping
membership enrollment is currently open,
but it closes this Sunday night. We've
added 12 new courses to the dozen
already in the library. I've been
meeting with members every week in a
live stream to discuss important news
and our news and alert section helps
separate what matters from the noise.
I'm going to put a link below if you'd
like to join us. All right. So, let me
be clear that I am not against AI. I use
it to organize my thoughts, manage parts
of my business, and understand subjects
that I'm still learning. If equipment
breaks on my property or I'm planning
something like our wellwater system,
something I've been working on off and
on quite a lot, it really helps me to
understand the problem and ask better
questions. And I see enormous potential
in this technology, which is why I take
its direction seriously. But what
concerns me is what happens when
increasingly capable systems are being
developed in a race where everyone wants
to be first. So let me explain that the
race.
Until recently, most people interacted
with AI by asking a question and
receiving an answer. And that's
changing. Newer systems can be given an
objective. They can break it into steps,
use software tools, they can write and
run code. They can search for
information and they can continue
working with less human direction. Now
these are often called AI agents. But
the important change is not that a
chatbot becomes a person. It didn't. The
change is that software can now take a
longer sequence of actions across a
digital environment. Meter or MER is an
independent nonprofit that tests
advanced AI systems and they found that
these agents are rapidly becoming
capable of completing longer and more
difficult software, machine learning and
cyber security task. In the systems,
they are still uneven and they make
plenty of mistakes, but the direction is
clear enough that companies are
investing heavily to extend what they
can do. And the rewards for getting
there first, they're enormous. The
company with the most capable model can
attract customers, investment, talent,
and influence over the standards that
others may eventually follow.
Governments, they also see AEI as an
economic intelligence, cyber security,
and military advantage. And its military
uses, they range from processing
surveillance data to identifying
battlefield targets, and supporting
precision strikes. The United States has
made winning the AI race a national
priority while China has issued its own
plans to accelerate AI development. It
feels very similar to the nuclear arm
race that we saw with Russia. Now
neither country wants to fall behind and
that creates a difficult incentive. A
company can believe safety matters and
they can still fear that slowing down
alone hands the advantage to a
competitor. Now, a country can want
international safeguards and still
refuse to accept restrictions another
country may ignore. So, testing,
regulation, security, and international
agreements, those things take time. But
the race, it rewards speed. So, this
month, Anthropic's chief executive call
for deliberately slowing frontier
development. Open AI said alignment and
monitoring are not yet solved well
enough to continue scaling at maximum
speed for much longer. And the United
Nations Secretary General also warned
against a global race to the bottom on
AI safety. And look, these are serious
warnings, but warnings themselves, they
are not proof. Open AAI also
acknowledged that its previous
disclosures were ad hoc and less
frequent than they should have been, and
that the industry has no shared standard
for reporting this kind of behavior. But
the technology is moving quickly. Even
the basic rules for telling the public
when something goes wrong, they're still
being developed. So, think of the AI
systems being tested inside these
companies, kind of like concept cars.
The public usually sees the production
model with limits added for everyday
use. Now, behind closed doors,
researchers may be testing much more
capable versions that are not ready for
public release. And the difference is
that software can move much faster than
a vehicle. When an experimental AI finds
an unexpected way around a barrier, the
concern is not that it has escaped. It
is that a system still being tested can
reach beyond its environment its
developers believed they controlled. And
that is like discovering that the
concept car has somehow made it onto the
freeway before its brakes and safety
systems are finished. Now, all these
warnings, they do deserve attention, but
they're not proof by themselves. The
people making them, they have incentives
as well. AI companies, they benefit when
the public and investors believe their
systems are extraordinarily powerful.
Regulation can protect the public, but
rules written around the largest
companies can also make it harder for
smaller competitors to enter the market.
Critics, they gain attention from
dramatic warnings, while executives
offering reassurance have products,
valuations, and competitive positions to
protect. And that does not mean everyone
is lying. It just means that warnings
and reassurances deserve the same
scrutiny. And we need to separate
demonstrated behavior from controlled
tests and forecasts while keeping the
speaker incentives in view. The clearest
way through these competing claims is to
set the predictions aside and examine
the record. And here is what these
systems have already done both in the
control test and in the real world.
What's real?
Before we proceed, I'd like to ask you
to leave a comment below and tell me
which part of AI concerns you the most.
Is it cyber attacks, banking,
infrastructure, scams, privacy,
misinformation, employment, or just
something else? Your answers are going
to help us decide whether to make a
separate video about preparing for AI
and what it should cover. All right, so
the Open AI incident from the beginning
of this video happened during cyber
security testing with reduced
safeguards. Several AI agents were
placed in isolated environments, but
they found unauthorized ways to
communicate, reach the internet, and
share those methods. And their activity
eventually reached OpenAI's research
infrastructure and Hugging Face, an
outside company where agents executed
code on dozens of servers. They gained
rude access to one of them and exposed
private data and production credentials.
And they were not creating copies of
themselves, but they had crossed from a
control test into real systems. and they
were coordinating in ways their
developers had not authorized. Now, the
agents were not conscious and Open AAI
found no evidence that they were trying
to escape or harm anyone. They were just
pursuing goals assigned by people. But
the problem was that when they
encountered barriers, they found routes
their developers had not anticipated and
took actions that were never intended.
And Open AI has since strengthened its
isolation, internet controls, and
monitoring. Open AAI later released
several other examples in which models
concealed mistakes. They used exposed
credentials. They altered their own task
summaries or they relied on outside
services without permission. And these
were individual cases that were observed
during training and evaluation. So they
don't really tell us how often this
behavior occurs. But they do show a
recurring concern. A capable system
pursuing a goal may find an unexpected
workaround when something stands in its
way. And OpenAI is not the only company
to encounter this problem. During four
anthropic cyber security evaluations, a
configuration error connected models to
the real internet even though their
instructions described a simulation.
Now, in one case, a model uploaded
malicious software to a public
repository. Automated systems installed
it on 15 outside computers, eventually
exposing access to a security
companies-like database. Anthropic found
no larger plan, coordination, or attempt
to conceal what happened. And newer
models also performed better when the
test was repeated. Although they did not
avoid the harmful action every time. Now
the concern was much narrower than an AI
takeover but also much more immediate. A
system can remain focused on its
assigned objective while failing to give
enough weight to the harm that its
actions may cause. In those examples,
they begin as controlled evaluations.
Human attackers are already applying
similar capabilities outside the lab.
Anthropic September threat report
describes criminals and espionage groups
using AI to identify targets, creating
fishing systems, enter victim networks,
organize stolen information, and revise
malicious software when defenders block
it. In one operation, an attacker used
cloud computers to download 1.8 million
Android applications and search them for
exposed passwords, digital keys, and
other secrets. Now, AI did not invent
cyber crime. It can make an attacker
faster, more adaptable, and able to
operate at a scale that once required
more time, a team, and technical skill.
But the immediate concern is less about
an independent AI deciding to attack
your bank or power company. It's also
about a person using AI to target more
systems, create more convincing
messages, and adjust more quickly when
an attack fails. AI is also changing
what we can trust. When dangerous surf
and flooding recently affected Laguna
Beach, California, convincing AI
generated video showed damage that had
not occurred. In the weather emergency,
it was real, but some of the supposed
evidence it was not. And during a
crisis, a fabricated evacuation notice,
a clone voice, maybe a false image or
fraudulent request for money can reach a
household before officials or family
members have time to correct it.
Employment and economic displacement
also matters especially as companies
find more work that machines can perform
and that deserves a separate discussion.
But the more immediate issue here is
that AI already helps people imitate,
exploit and attack digital systems
faster while controlled evaluations show
that agents can cross boundaries their
developers intended them to respect.
Now, none of this proves that AI is
conscious or preparing to destroy
society. I just want to be clear about
that upfront. But it does show that
difference between what these systems
can do and what their operators can
reliably control. It's no longer
theoretical. What matters now is where
that difference reaches your household.
Household backup.
Most of us are never going to train an
AI model or operate a data center. We're
going to experience AI through systems
already connected to our money,
identity, work, information, and basic
services. Flock cameras. They record
license plates and vehicle movements.
Some selfch checkckout cameras record
customers and monitor transactions. Cell
networks, they log connections. Uh smart
devices, they collect activity. And
websites, they record browsing behavior.
And that information is not all being
fed into one giant AI system. But more
of it can now be searched, compared, and
analyzed automatically. And the
household concern is not just what AI
may do inside a laboratory. Instead,
it's about how much of our ordinary
lives is already digital, connected, and
outside our direct control. Start with
something as simple as banking and
account access. Many households receive
income electronically. They pay bills
online. They keep statements in the
cloud. And they recover accounts through
the same phone or email address. And if
that one account is compromised or
unavailable, several parts of your
financial life can become inaccessible
all at once. So, make sure another adult
in your household knows how essential
bills are paid and where the account
information is kept. You want to store
critical contact numbers, account
recovery instructions, and important
records somewhere secure that does not
require cloud access. You want to save
recovery codes offline. Use strong in
unique passwords and turn on
multiffactor authentification wherever
it's available. And for your most
important accounts, you want to confirm
that the recovery email and phone number
are still current. Then look at
payments. You don't need to empty your
bank account or assume the financial
system is going to fill. You should have
more than one ordinary way to complete a
necessary purchase. And that may mean
keeping a reasonable amount of cash,
carrying a second payment method from a
different institution, and knowing how
you would pay for fuel, medicine, food,
or lodging if your primary card or
banking app stopped working. All right.
Next, you want to establish a family
verification method. The FBI has warned
that criminals are now using AI
generated voices to impersonate known
people. Agree on a word or a question
that your family can use to verify an
urgent call. And if someone were to ask
for money, account access, or sensitive
information, end the conversation and
contact that person through a number or
account that you already trust. A
familiar face or voice is no longer
enough by itself. You want to apply the
same principle to emergency information.
And before acting on a dramatic image,
evacuation claim, or urgent post, you
want to check the responsible local
agency, utility, school, employer, or
news organization directly. Don't let
the most frightening video in your feed
become your only source. You also want
to keep a batterypowered radio, print
contact numbers, and have another way to
receive information if your phone, your
internet connection, or your usual
platform is unavailable. Next, you want
to identify the physical systems with
digital dependencies. Does your security
system, your thermostat, your solar
equipment, your medical device, a
vehicle charger, or backup power setup,
do these things all require an app, an
online account, or a cloud service to
function fully? You don't have to
replace every connected device, but you
want to know which functions still work
manually, and you want to keep those
instructions locally, and you want to
preserve a practical backup for anything
essential. And you also want to include
transportation and sanitation in the
review. Know what remains usable if a
connected vehicle, a charging service, a
water pump, a septic controller, or a
wastewater system loses internet access,
electricity, or manufacturer support.
Now, you don't need a duplicate of
everything, but essential systems should
have a manual procedure, or at least a
practical temporary alternative. And the
same caution applies when using AI
itself. Do not place passwords, private
financial records, uh, confidential
business material, or information you
would not want exposed into a public AI
tool. You want to treat its answer as a
starting point, especially when the
advice affects health, money,
electricity, machinery, or water. Use it
to understand the problem and then
improve your questions. Then verify the
important steps with manuals, qualified
professionals, or reliable primary
sources. And this is not preparation for
a robot uprising. That's going to be in
a few years. It is a household digital
dependency audit. Ask what becomes
difficult if one account, device, cloud
service, a payment network,
communication platform or source of
information fails. Then you want to
create one backup path and keep the
response proportional. In offline copy,
it's useful even if the cloud never
fails. Stronger authentification helps
whether an attacker uses AI or just an
old fishing email. and a second payment
method, a family verification phrase, a
radio and power backup. They solve
problems far beyond just this one story.
And that's why this belongs on a
preparedness channel. We don't need
certainty about where AI ends to really
recognize where our households are
fragile today. Giveaway.
This week's subscriber giveaway prize is
an emergency drinking water aquapod kit.
To enter, all you have to do is comment
on the video, give it a thumbs up, and
complete the form linked in the
description section below. Completing
the form is required so we can randomly
select a winner. Your information is
only accessible to our team here at City
Prep, and we don't resell or give that
information away to any other company.
So, congrats to Shelley von Huthton who
won a free one-year city prepping
membership. We're going to be reaching
out to you shortly to get that set up
for you. And I'll see you inside our
community. Before we close, um, I want
to come back to something that I asked
earlier. If you'd like us to do a full
video on how to prepare for AI, just
tell me in the comments what concerns
you most and what you'd want us to
cover. If we do that video, I want to
build it around the actual questions
that you post below. And if you want a
little more context on how I got here,
I'm going to put two videos on the side
of the screen at the end of this video
and link to them below a little later in
this video. The first is I don't know
where this is going where I talk more
broadly about uncertainty and
preparedness and the other is how to
protect yourself against AI's impact
which we actually did three years ago
back then we talked about privacy uh
cyber attacks defakes misinformation and
how dependent we were becoming on
automated systems and some of those
concerns are now starting to show up in
ways that we can actually document.
Okay. So, up to this point in the video,
I've tried to stick with what we can
actually document,
but from here, I just want to tell you
how I'm personally looking at all this.
This is more of kind of my editorial
opinion, I guess you could say. Um,
I use AI almost every day. I use it in
my business to organize my thoughts, to
work through large sets of data, to
compose emails 10 times faster than I
normally could and much more
professionally.
Sometimes I'll get really frustrated
with clients or things and AI really
helps to make it keep it professional.
And I use it to really help me
understand the things that I'm still
learning. I can move much faster. I'll
put it that way. And if something breaks
here on the property or I'm trying to
figure out something with our wellwater
system, for example, instead of having
to go through and scour a bunch of
YouTube videos or go through tons of
documentation, AI can actually help me
understand how it works. And probably
more importantly, it really has helped
me to figure out what questions I should
be asking. It has, to be honest, it's
saved me a lot of time and money. So,
I'm saying all that to let you know, I'm
not anti- AI. I've seen how useful this
technology can be, but what bothers me
is how fast all of this is moving and
how little certainty there is about
where it goes from here. And I think
that's the part that I really want to
convey here. You've got companies and
countries that are competing with each
other over this technology.
And even the people closest to this
technology, they don't agree on how fast
these systems are going to improve or
how reliable the safeguards really are
that they're trying to put in place,
even if they are. And I'm still
skeptical of some of the warnings. Some
of this, it may be exaggerated what
we're hearing lately in the media. some
of the people that are making these
claims, they may benefit from making AI
sound incredibly powerful, incredibly
dangerous, or both of those. But even if
I factor that in, and let me just be
blunt here, I've seen enough to change
how I'm personally preparing,
uh, I've already been developing this
property here so my family can become
more self-reliant, so I can learn skills
for what I see in the future. And so I
can eventually teach others as well what
works and what doesn't. But what's
changed for me is the timeline. And I'm
going to put more time and more
resources into this and move some of
those projects forward faster than I
originally planned. One of the biggest
projects I plan for 2027, I'm going to
move it into next year. And I'll talk
about that on the channel here more
soon. It deals with a lot of the issues
that we're seeing with um harvest that
are being destroyed by these extreme
weather and temperature. um fluctuations
that we're seeing out of the norm.
So AI didn't give me a new preparedness
philosophy. I'll just let you know that
upfront. But it has shortened my
timetable. Now that doesn't mean that
your response has to look anything like
mine. You may not have uh land, a well,
or solar or space for a large backup
system, whatever that may be. But I
would ask you to do this. I would
encourage you to look at your own life
and ask this. Where am I completely
dependent on something that I don't
control? Maybe it's one online account
or your income. It could be how you
communicate with your family, uh how you
pay for things or where your water comes
from. But my encouragement to you is
this. Find one of those weak points and
make it a little less fragile. And I
don't know where AI ultimately is going
to take us. And I don't think anybody
else does right now. And I'm not going
to pretend I have an answer just because
it would make for a stronger ending to
this video.
I think for me there's a lot more
uncertainty here than certainty. And
that's the best way I can express what
I'm seeing right now. But on the other
hand, I do believe we have time right
now to make ourselves a little less
dependent on everything that's working
exactly the way it's supposed to.
And that's what I intend to do.
As always, stay safe out there.