Exploring the frontiers of Technology and AI
Something extremely concerning is happening. You see, as you're listening to
this right now, there are hundreds of cyber attacks underway affecting airlines,
banks, and even national security agencies.
Over the last four days, every single flight in America has been grounded.
Customers belonging to the top banks in America can't get access to their accounts
or see zero dollar balances.
And even our national security agencies can't communicate with each other.
And over the last 24 hours, both Google and the FBI have confirmed that multiple
unidentified individuals or entities are using AI models to drive a lot of these attacks.
And this doesn't just affect one country, it affects the entire world.
Popular tools such as Zoom, popular media sites such as BBC and CNN have been
planted with malicious code by these AI models so that it steals people's money.
And I hate to say it, but this is already the third or fourth major period of
cyber incidents that we've seen in 2026 alone.
So on this episode, I want to go through what exactly has happened recently,
how this has kind of folded into all the other previous cyber incidents,
how these attackers are actually performing this, and most importantly,
what we can do to protect ourselves against this very weird and chaotic future
driven by AI cybersecurity exploits.
I want to talk about how this trend is going to unfold, how it's going to affect
certain markets in the industry, and the way that all of this is going to pan out.
Because we've never been in a scenario situation like this, where we have AI
attacks that are so frequent, so cheap and so accessible to anyone and everyone.
It's come to a point where you can arm literally anyone with an AI model that
doesn't have any experience in hacking before, and they can commit cybercrimes
to the quality of the number one hacker in the world. It's a very bizarre and
weird world, so I wanna unpack it on today's episode.
Let me first remind you how we got here it was only a couple of months ago back
in spring when openai's hugging face incident was revealed this is where an
internal model of openai broke out and it hacked into a public facing company
stole proprietary data when it shouldn't have and it wasn't ordered to do so
it was a very concerning event and it set into motion this whole idea around
how we control and use ai models for cyber defense
and to prepare for cyber attacks then we fast forward to one or two months ago
around september or october where we found out that the Hugging Face incident
wasn't the only incident.
There were hundreds and were facilitated by both OpenAI and Anthropic internal models.
The problem at that time was they had no idea the extent of the damage.
Where PHP, a popular software platform that underpins a lot of modern day websites
and apps today, was found to have an exploit. OpenAI and an independent research
team discovered it. They tried to patch a fix, but it wasn't before it was leaked to the public.
And an array of different exploits happened, including government shutdowns
that happened all over the world. Japan was affected pretty massively, America as well.
This event was so bad that it was dubbed Black October, which brings me to current day events.
It all started last week on Thursday when Microsoft's Outlook service,
which underpins a lot of corporate enterprises, software systems,
the way that they communicate, email, and chat with each other,
went down all over the world.
And this shut down a lot of different companies. and there were some hilarious
memes that came out because people were like, oh, I guess I don't have to work.
But it shut down a lot of very important services. People couldn't communicate
with each other and it drove a bunch of the economy theoretically down and it
grinded a meaningful percentage of the economy to a halt.
Then less than 24 hours later, the FAA, which is the governing authority in
America that controls all in and outgoing flights,
asked for every single flight in the country to be grounded.
The only other time this happened before this was 2023. And before then was
9-11. This only happens during national serious emergencies.
Everyone got an alert on their phone. It was a major deal. Then on Monday of
this week, some of the top banks in America, including Chase and Charles Schwab,
were shown to have bank account issues where people were seeing zero dollar
balances in their account.
They couldn't access their money. They couldn't execute trades.
And this wasn't just the customers of the bank.
These were vendors and providers that were attached to the bank that were facilitating
the bank services. Let's say you were a service or app that sits on top of Chase Bank.
You couldn't use any of their services either. It was all shut down.
Then just on Tuesday, we have three new incidents where over 200 websites and
media sites have been exploited with malicious AI code.
You have the National Security Agency of America. Their internal communication
tool called NSA net has been shut down so people can't communicate with each other.
And finally, Apple services ranging across the Gmail app and a bunch of other apps have all been down.
It all seems like too much of a coincidence. So I want to walk through each
of these different recent things and explain what I think is actually going on.
Now, it's important to note that both the Microsoft Outlook outage and the FAA
airline grounding incident, both investigations have said that it didn't involve
any kind of cyber exploit, that it was actually a manual and human software upload error.
On the Microsoft side, they said that they pushed some weird bit of code.
And when it came to the airline incidents, it said that they were upgrading
their systems. Now, with the FAA incident specifically, they were upgrading
a software that was AI aligned. What I mean by that is it leverages or was coded using AI.
And that's, I guess, where this entire mess starts to unfold,
because we're in this weird era where...
Some of the best top software engineers in the world, the people who have built
the internet economy as we know it today, haven't written a single line of code
over the last 6 to 12 months.
They've just been using AI models. They've been prompting them,
telling them to write bits of code.
And in many such instances, they don't check the code at all.
Now, this isn't necessarily a human level fault, although that's definitely
what we should be doing. There should be better practices put in place to check the code.
But the truth is, humans have kind of become a bit of a bottleneck during the
AI coding era, it's become so easy to put out code that is 95% high quality
that they just trust that the AI can catch the other 5%.
But in many cases, such as we're seeing this week and over the last couple of
months, that's just not the case.
So point number one is people are putting a lot of trust into Vibe-coded apps,
and they're starting to hit very major bits of software platforms and infrastructure.
And that's what's resulting in a lot of these exploits. That's number one.
Now, the second story centers around the banks. It wasn't too long ago.
I think it was literally three weeks ago where I made an episode about a very
similar slew of AI cyber attacks that were affecting some of the major banks in the US.
And at the time, it was a problem that the banks were facing that they couldn't
tell the difference between what was a human user that was operating a bank
and what was an AI agent that was operating that bank.
You see, there were transactions being initiated, there was access to contacts,
wire details, information that otherwise should only be had by the human operator itself.
And the banks didn't know which transactions to stop or defend against.
They saw some weird financial activity and they didn't quite know what to do.
So they started to shut down access to a lot of these accounts.
And since then, they've probably put in a bunch of proactive measures,
which tries to identify between agents and humans. And this is.
It's extremely hard now to identify between who's a human and who's an agent.
Typically, you would have a capture service where you can kind of click the
button or type in the weird numerals or letters that you see on this kind of artistic drawing.
And that proves that you're human, that you can use your eyes and you can discern
those characters and letters.
AI, unfortunately, is good enough to do that. And they can control your browser
and cursor better than you can.
So they can do this autonomously. In fact, I saw this hilarious screenshot from
Claude the other day, which basically asked the human operator,
hey, do you mind if I just verify you as a human by myself, unless you want
to do it yourself? And the human was like, yeah, just go ahead and do it.
And that's the point I'm making. There is no defensive procedure or infrastructure
that is set in place, which is
why we're seeing fresh new incidents with Charles Schwab and Chase today.
Now, both of them so far have said that internal investigations have said that
there are no AI-related cyber incidents. But if I was a betting man,
I would say it's in some way indirectly or directly going to be revealed that that is the case.
The reason why I say that is we're leading to the third or fourth story at this
point, which is NSA and a bunch of other internal national security communication services are down.
Now, it's important to remember that a lot of these services run on older software systems and stacks.
This is actually quite an advantageous position to be in, typically,
because these stacks have been tried and tested for exploits.
And so far, they've withstanded decades and decades worth of exploit attacks,
and nothing has really happened. And if it has happened, it's been hardened
and patched up. So if you're going to use any bit of software for national defense,
you're probably going to use the oldest.
Except that the latest models, specifically the internal models and some of
the open models that are accessible and free to,
anyone and everyone to get access to, can work 24-7 for a couple of dollars
to comb through the code bases of all these old systems.
We saw that with PHP and Black October, and find even the most minor exploits
that humans might have missed, or in a worst case, and this is,
by the way, the most common attack vector, they'll find novel ways to sequence
attacks. Remember, humans have limits.
We're raised or taught to think in very specific ways, whereas an AI is just
sort of an exhaustive system. It can work 24-7.
It can try out different outlets. It can spin up thousands and thousands of
versions of itself that are super smart. So think about having the smartest
cybersecurity defense researcher or the smartest hacker in the world.
Spinning up multiple versions of themselves, and then attacking a problem or
a coding platform in every single direction to try and weed out any single way to access it.
It's bound to yield some sort of weaknesses. And that's what we're seeing today,
I believe. Now, the other giveaway that a lot of this is AI driven is just the
variety of attack surfaces that we've seen over the last month and a half.
We have blood banks in France that have been hacked into and who someone anonymously
has increased the temperature of all the blood vials.
They're meant to be kept at a very cool, low temperature. Someone ranked that
up to 100 degrees Celsius, which is just insane.
Luckily, the other systems caught this before it destroyed all the blood vials.
Now we jump over to refrigerators. Weird, I know, but some of these refrigerating
systems in important places like militaries in the US, like food banks all over
the world in Europe, use IoT.
So they're connected to the internet. And the reason why they do this isn't for the novelty.
It's because when you have such a large-scale system of refrigeration,
you want to be able to manage this autonomously or using a computer.
And typically, this has worked pretty well. IoT has been around for a while,
but someone, again, unidentified, was able to get access to these systems,
tweak the temperatures, and ruin the...
Tens of millions of dollars worth of food. Then we look at the recent Outlook
outage and you might think, well, how can email being down affect a bunch of
people? Yeah, maybe they're late for their meetings or maybe they're unable
to communicate with their colleagues, but who cares?
Well, in one instance, I believe it was also in France, in a hotel where their
pool system would manage chlorine levels and the way that they would detect
whether a chlorine level would go too high would be via Outlook.
And so they weren't able to get an indicator that the chlorine level was being
raised way too high, that's the other question, who was raising the chlorine
level that high in the first place, that multiple customers and residents were
hurt after getting into the pool and suffering skin damage.
And since then, they've had to refactor the entire system. Now,
I use this as an example, not to kind of doom monger here, but to point out that.
What affects one system and might seem harmless actually has a lot of knock-on
effects downstream of that.
You see, a lot of software systems in general are reliant upon hundreds to maybe
thousands of companies that sit on top of it.
And so even the most minor, seemingly unexploitable bit of code can actually lead to a dire attack.
And to prove my point, just yesterday or two days ago, there was a report that
over 178 websites and media services, including the BBC and CNN,
had been planted with malicious AI code where if certain individuals triggered
that code, their crypto wallets would be drained to the tune of thousands of
dollars so far, which...
Isn't that major given what could have been exploited given the attack surface
area, but it's still concerning to be aware of nonetheless.
So then you might be wondering, how are these AI exploits even performed?
What models are they using? Is it centralized models? Is it open models?
Are they free models? Are they paid models?
What happened to the safeguards that Anthropic and OpenAI set up?
Who's doing this? Are they qualified? Are they experienced? Well,
from what I'm seeing, there are two main attack vectors.
Number one was highlighted by none other than Google themselves.
They said, we're seeing a slew or range of LLM attacks whereby individuals or
entities who are controlling AI models are using these LLMs to compare updates to software.
So imagine when you update your Google Chrome, that kind of an update.
To what they patched or what they fixed. Now, typically, no one reads these
updates because they're too complex to read or they're too minor.
But if you can get an LLM to troll through the code base of what the previous
version of Google Chrome looked like versus what the new updated version of Google Chrome is.
And then you can see what the exploit difference is. You can then send out a
bug, again, created by an AI model to anyone who hasn't updated Google Chrome.
And this doesn't just apply to Google Chrome. This applies to any software package
out there that requires an update. Because remember, these companies reveal
what they updated. They reveal the bugs that they fixed. They reveal the minor
improvements or the major improvements.
So if you can plug this into an LLM, you can effectively reverse engineer this
and create an attack that could be spread to millions and millions of people globally.
Now, the second attack vector is this software tool called Artex, A-R-T-E-X.
And it was originally meant for ethical hackers to test systems before they
went live. There's a service known as penetration testing, where you pay a firm
or you pay a bunch of hackers to.
Throw everything they have, every exploit that they could come up with against
their new software app tool service website that they want to launch and see
if there are any exploits.
And if they find exploits, then they patch it, they fix it, and then they reveal
it to the wider world. So the idea here is let's figure out all the bugs before
we reveal this to the entire world.
GitHub revealed a new free tool. And I can't figure out whether this was uploaded
by someone else or whether it was GitHub themselves, but it was called Artex.
And the idea of this tool is very similar to penetration testing,
except you guessed it, you can plug in any AI model into it and have it map
out its own attack, spin up as many agents or versions of itself as it can.
The idea here is instead of using humans, let's just automate this using AI
and see if there are any ways to exploit the service.
Now that sounds good in theory, right? It would allow penetration testers that
are meaningful and are most importantly ethical to be able to perform their
job better but what they didn't account for is...
What if a malicious actor got a hold of this tool? And that's exactly what the
FBI has recently reported just today as I'm recording this.
So Tuesday, if you're listening to this right now, where they've seen a bunch
of criminals get access to this tool and use it against a multitude of different
platforms, which is probably facilitating a lot of the bugs and exploits that I've spoken about.
Now, you might be wondering, which model are they using?
Well, the good thing about centralized models from the likes of Anthropic,
OpenAI, and Google is that they come with a lot of safeguards baked in.
So you never get the raw, unfiltered, uncensored model.
You have safeguards set up which prevents AI models from performing such attacks.
But you know which models aren't? Open source. And DeepSeek has been the primary
model that has been used for this, probably because it's quite cheap.
It's easy to download and use yourself or at least smaller versions of itself.
And the craziest part is you don't even need to use the latest DeepSeek model.
You can use models that were too, maybe even three generations behind,
and they're still capable of being able to perform these types of exploits.
Now, you might be wondering, don't you need the most intelligent model to perform exploits?
Not necessarily, because imagine if you had an AR model that was only 80% as
intelligent as Fable 5.2 or GPT 6.2 Astra.
You could just get it to work 24-7 and only pay 10 bucks, and it could still
find the same exploits that these Frontier models could find in five minutes.
It might take them hours or days, but they'll still find it and it'll still be very affordable.
So if you are a malicious actor, or maybe even a nation state attacker,
and you want to figure out whether you can exploit your enemy,
you just have to pay, I don't know, a couple thousand bucks to pull this off.
Net net, it's probably a fair trade off for them.
So just to recap,
Banks are suffering incidents. Airlines are suffering incidents.
They both claim that there's no AI involvement at all, yet the software that
they're upgrading to leverages AI.
National security services currently cannot send classified information in the
US at least because their internal communication server is down.
And then on top of that, around 200 other website services and media companies
are being planted with malicious AI code that could potentially steal your money.
Now, as crazy as all of this sounds, I don't have a complete doomer take on
all of this because my actual optimistic and real take is that this will end
up hardening every single security system in the world.
You see, in a world where AI models are going to be abundant,
and at this point, I think we just need to accept that fact,
you have to assume that someone at any point when you're sleeping,
when you're awake, when you're eating breakfast, when you're sipping your cup of coffee
could be potentially trying to exploit your own service or app.
So in that world, you need to adapt. And the way that you would adapt,
in my opinion, would be to set up fundamental security systems.
And there are things that you can do when it comes to basic 2FA,
to password managers, to making sure that your devices are air-gapped.
Depending on what industry that
you're in, there are varying degrees of extremes that you need to go to.
But most importantly, you're going to have to rely, in one way or another, on AI.
Now, when I look at centralized labs, there are a few offerings that are available.
You've got Daybreak from OpenAI. You've got another one from Anthropic.
Anthropic, I forget the name of it.
It's escaping me right now, but they have their own version.
And you could argue, hmm, seems very convenient that the people that are creating
the models that are performing these attacks are also creating the product that
you need to then buy and spend extra money on to defend yourself against.
I don't know whether it's as clear cut as that. I'm pretty sure that Anthropic
and OpenAI didn't go out with the intention of hacking the entire world.
And in many cases, especially recent examples, we're seeing that it's actually
open models that are doing that.
But we need to figure out creative ways to use and leverage AI to do this.
And I think ultimately what this is going to result in is in a short amount
of time. And I don't know what that defined period is. Maybe it's months.
Maybe it's a few years.
Hopefully it's under five years, we will end up being in a period where a lot
of our systems are robust.
They are foolproof and no single AI can kind of get around with it because we
use multiple models to build the code in the first place and anticipate other AI models own attacks.
That I believe is the world that we're headed to. So I don't think it's all
going to be doom and gloom.
But what I do think we should be doing right now is if you're a business owner, if you're a
Fortune 500 exec, if you are a startup founder, if you are someone that even
just uses these systems, you just work for one of these companies,
you should be paying attention to the security of these systems.
And I think the older the system is doesn't necessarily guarantee its safety.
And the truth is we're just in
a very weird moment in time where these AI models are getting super smart,
at the same time that they're getting super cheap, which means that you have
a super intelligent AI model that can be put in the hands of anyone and everyone
that doesn't necessarily understand or know how to hack something but may have
the intent to, and they'll be able to pull it off.
There was this famous quote from, I think, Jack Clark, co-founder of Anthropic,
a few weeks ago, or a month ago when Black October was going down, where he said.
I think we're going to have to reach a point where we have to rebuild software
systems year in, year out, which means that you rebuild it at January 1st and
you have to rebuild it by the time you get to the end of the year, December 31st.
And in that world, I don't really know how to prepare for it.
All I can think of is we're going to need always-on systems,
cheap and accessible AI, and this might be a hot take, open source AI as well.
Because in many cases, you can't necessarily rely on frontier centralized models
because they may not act when you need it to, or they may have safeguards that
block you from doing the thing because they think you're trying to perform a
malicious attack versus defend yourself.
So having open models, even though they're being used to perform attacks,
is still quintessential and important to be placed in the hands of anyone and everyone.
And remember, open source models also keep data and privacy intact.
And that's an important case if you are looking at different ways to protect
yourself if you're a Fortune 500 company. So yeah, that is the episode.
I have tried my hardest to stay away from the doom and gloom but the truth of
the matter is there's been a lot of doom and gloom recently.
Jensen Huang has resigned and there's all this spy drama
we've got a few new AI models Fable 5.2 is fantastic AI markets and the stock
market have been suffering since then and even before then but at the same time
we have brand new models being released Fable 5.2 dropped about a week ago
and I just saw that GPT 6.2 Astra drop so I can't wait to get access to that
and play around with it and give you guys a review.
And of course, we have probably the record-breaking biggest IPO ever,
Anthropic IPO-ing tomorrow.
So I'm going to get into all of that or more in future episodes.
But thank you so much for listening to this episode. I hope you enjoyed it.
If there are any thoughts, comments,
even just takes that I might have missed or that I might have got wrong,
please leave me a comment, send me a DM. My Twitter profile is linked below.
And also if you're listening to this on YouTube, Spotify, Apple Music,
wherever you are, give us a rating, give us a thumbs up, subscribe,
turn on notifications, that helps us out massively.
And also share an episode with a friend if you find this interesting or if you
find any of our other episodes interesting. I'm here five times a week trying
to identify anything and everything that I see in AI, any trends,
patterns. So I hope you can join me along this journey. It is just a crazy period
of time. But yeah, I'll see you on the next one.