Smashing Security podcast #493: Why your AI model could be a sleeper agent

Hacking stories and cybersecurity insights.

Graham Cluley
Graham Cluley
@

 @grahamcluley.com
 / grahamcluley

Smashing Security podcast #493: Why your AI model could be a sleeper agent

You join a video call, and your contact and a few colleagues are already there, waiting patiently. But they can’t hear you, and a pop-up says you need to install an update to fix your microphone. What could possibly go wrong?

Meanwhile, we explain how changing just a handful of numbers can turn an AI model into a sleeper agent, waiting for the right moment to strike.

Plus, Andrius Buinovskis of NordLayer explains how companies can see which AI tools their staff are really using, and stop sensitive data leaking out, without banning AI altogether.

All this and more in episode 493 of the “Smashing Security” podcast with cybersecurity expert and keynote speaker Graham Cluley, and special guest Ray Redacted.

0:00
0:00 0:00
0:00
Show full transcript ▼
TranscriptThis transcript was generated automatically, probably contains mistakes, and has not been manually verified.
GRAHAM CLULEY
You don't have to tell me if you're not comfortable doing so, Ray, but do you let an AI actually do things on your computer?
RAY REDACTED
100%. Absolutely.
GRAHAM CLULEY
Oh my goodness, Ray!
Unknown
Smashing Security, episode 493, Why Your AI Model Could Be a Sleeper Agent, with Graham Cluley and special guest, Ray Redacted.
GRAHAM CLULEY
Hello, hello, and welcome to Smashing Security, episode 493. My name's Graham Cluley. And I'm Ray Redacted.

I always wonder, Ray, when you're on the show, whether we should bleep out the redacted. Maybe we should just have Ray beep.
RAY REDACTED
You know, I've actually had people ask me that before. Graham, here's the funny thing.

I once upon a time was actually looking to legally change my name to Redacted, but my attorney told me I could not do that 'cause I would get subpoenaed all the time anytime somebody was redacted from the court orders and stuff, so.

Did not do that.
GRAHAM CLULEY
It's better than being Bobby Drop Tables or whatever though, I suppose. That could have caused problems too.
RAY REDACTED
XKCD reference. That was really good. That's particularly sharp of you this morning.
GRAHAM CLULEY
Well, before we kick off, let's thank this week's wonderful sponsors, NordLayer AI Scout, Dropzone AI, and Vanta. We'll be hearing more about them later on in the podcast.
Unknown
This week on Smashing Security.
GRAHAM CLULEY
We won't be talking about how a flaw in a website testing tool let malware reach the BBC, CNN, and Zoom.

You'll hear no discussion of how a breach at an AI cloud vendor exposed close to 150,000 Uber driver licences.

And we won't even mention how 2 Grok AI agents tried to raid corporate accounts. So Ray, what are you going to be talking about this week?
RAY REDACTED
I'm gonna be talking about an exploit called Lazarus that Microsoft Research has recently published.
GRAHAM CLULEY
And I'll be talking about how your face might be being stolen to fool your colleagues on Zoom.

Plus, we've got a featured interview with Andrius Buinovskis of Nord Security about how you can get better visibility into how sanctioned and unsanctioned AI tools are being used in an organisation and prevent data leaks.

All this and much more coming up on Smashing Security. Coming up in this episode of Smashing Security. Joe, can I ask you a slightly awkward question?

If it's about the biscuits, that wasn't me. It's not about the biscuits. Which AI tools have you been using for working on the podcast?
JOE
Just the approved ones, obviously.
GRAHAM CLULEY
So what's this Chatty McChatface thing? Free, no sign-up required.
JOE
Ah, that's more of a research partner.
GRAHAM CLULEY
And what have you been pasting into it, Joe?
JOE
Oh, just a few bits. Guest emails, the sponsor schedule, your bank details.
GRAHAM CLULEY
My bank details?
JOE
It needed context.
GRAHAM CLULEY
Listeners, this is shadow AI, and Joe's not alone. IBM found that 43% of breached organisations had a shadow AI incident this year. That's more than double the year before.
JOE
Which is why this episode is sponsored by NordLayer AI Scout.
GRAHAM CLULEY
AI Scout finds every AI tool being used across your organisation in the browser and on the desktop, so you can see what was asked, by whom, and where.
JOE
Hang on, by whom?
GRAHAM CLULEY
By whom, Joe? It also flags passwords, financial data, and personal information at the prompt level before it wanders off somewhere it shouldn't.
JOE
Does it also show who's paying for AI licences they never actually use?
GRAHAM CLULEY
It does. It measures real usage, not just the seats you've bought.
JOE
Great, because I noticed we're paying for 3 AI subscriptions and the only thing you've used them for is writing a limerick about ransomware.
GRAHAM CLULEY
Right, moving on.
JOE
NordLayer AI Scout rolls out to every device at once and runs quietly in the background.
GRAHAM CLULEY
Unlike some people.
JOE
See how your teams really use AI and what it's really costing you. Book a free 30-minute demo at smashingsecurity.com/nordlayer.
GRAHAM CLULEY
That's smashingsecurity.com/nordlayer. And Joe, delete Chatty McChatface.
GRAHAM CLULEY
Now, chums, I want you to imagine the scene, if you will. You are really into cryptocurrency. Are you really into cryptocurrency, Ray?
RAY REDACTED
I am, but I don't usually admit that for OPSEC reasons on podcasts.
GRAHAM CLULEY
Oh dear. Well, you just have now. Well, maybe you're into cryptocurrency.

Maybe you work for a crypto company, maybe you've made a small fortune investing in crypto coins, and you get a message out of the blue.

A message comes through on Telegram or LinkedIn, wherever it is, and it's from somebody that you know.

And maybe it's someone you met at a crypto conference last year or someone fairly important, some bigwig in the cryptocurrency industry.

And they're saying, hey, do you fancy a quick chat? Jump on the call, here's the link. And so you click on the link.
RAY REDACTED
Sure.
GRAHAM CLULEY
And you land in something which looks well, just like a Zoom or Microsoft Teams call. And you know how it is, you're in the waiting room, right?
RAY REDACTED
Mm-hmm.
GRAHAM CLULEY
You're waiting for the call to begin and they're showing you a little preview of yourself.

And that's fantastic because you can check your webcam preview, you can check your hair, you can sit up straight, make sure you're all comfortable in the middle of the screen, make sure you haven't got any broccoli stuck anywhere.

But what you don't realise is that all the time that you are in that waiting room, somebody is secretly already recording you. So footage is being taken.

And finally, the other people show up and the meeting starts. And there's the person who contacted you on the screen and they're nodding away.

And perhaps there's a few other people there as well. And it's all very normal, but you get a message in the chat inside the messaging app saying, sorry, Ray, we can't hear you.

I mean, this happens all the time, doesn't it? I was doing a webinar the other day for a customer. And I dialled in 15 minutes early.

They said, please come in early, Graham, to make sure everything's — and I said, yeah, no worries.

And so I got there early and the people hosting the webinar, they couldn't hear me at all. And so I said, oh, I'll disconnect and I'll reconnect.

And then I had to restart my computer and I was feeling, oh my gideon, because of course it was going to go live with people tuning in. I was like, ah, I can't do this.

I only made it by the skin of my teeth and everything eventually got to work.
RAY REDACTED
That's usually when you get software updates too.
GRAHAM CLULEY
Yes. So anyway, this is not an uncommon scenario of arriving on a call and your microphone or something not working properly.

And in this particular case, what happened was an error message pops up saying that you need to run an update to fix the problem because your microphone — maybe there's some sort of codec or audio driver or something that needs to update.

And you know, it's mortifying to you because you just want the call to happen because, you know, you're thinking, oh, all these other people are waiting for me.

So maybe you would, in that situation, run the update or do the install or apply the patch or whatever it is that it's telling you to do.

And what you don't realise is that your contact, the one you're having the call with, was never actually there. This has been explained by the FBI in recent days.

They say you are looking not at your contact. You're actually, in fact, watching a video of a prior victim of this particular hack.

So somebody you know has actually previously been hit by this thing. Video of them has been taken, and it's now being played back to you. So this isn't deepfake.

This is a genuine piece of footage of someone sort of sat in front of their monitor looking like they're preparing for a web meeting.

So all of these people who you can see who are in the meeting room waiting for you to fix your microphone, they are just there piling on the pressure on you.
RAY REDACTED
Of course.
GRAHAM CLULEY
To go and fix your audio, which is, you know, a pretty sneaky trick, I would say. And this update, which they get you to install, surprise, surprise.

Now, Ray, you and I, we've been working in cybersecurity a long time. It's not gonna be any surprise to us that of course the update is malicious.
RAY REDACTED
Oh, but Graham, I could so see myself falling for this, 'cause you're on the call and you see all these people waiting for you and you're not thinking in terms of this might be malware.

It's like giving you an update — this looks like it's from Zoom or whatever. I would absolutely click that.
GRAHAM CLULEY
And you receive the message in the first place from somebody you know or have an industry contact with. It's not like some unknown who has sent you this message.
RAY REDACTED
Sure.
GRAHAM CLULEY
And this malware gives the bad guys access to your computer, access to your cryptocurrency wallet, quite possibly your other accounts as well, which means, of course, they could steal — well, I don't know how much you've got in your cryptocurrency wallet, but, you know, they could steal a fortune.

And according to the FBI, the people behind all of this — let's open this up as a quiz, Ray. Where do you think they might be based?
RAY REDACTED
I would guess this sounds an awful lot like a rogue nation in Asia that is interested in pursuing nuclear weapons.
GRAHAM CLULEY
Yes, yes, yes. In other words, it's North Korea. It's the North Korean government.

I don't want to say that everything cryptocurrency-based is going to be North Korea, but it's not uncommon.
RAY REDACTED
A lot of it is.
GRAHAM CLULEY
A lot. They are funding an awful lot of their naughty business with stolen cryptocurrency.

Back in episode 478, I told a story about how North Korean hackers were recording people's webcams during fake job interviews.

And this seems to be an extension of that kind of attack. So you've already said, Ray, that if you were in this sort of scenario, you would be tempted to run the update.

So it's no surprise that people who work outside of the cybersecurity industry would be fooled into doing that as well, is it?
RAY REDACTED
Well, what's fascinating about this, if it is indeed the DPRK, is they've gotten really good at social engineering lately. If you think about
GRAHAM CLULEY
The
RAY REDACTED
Old school North Korean attacks were typically zero-day exploits or bugs, or maybe badly written phishing emails.

But now suddenly they're getting pretty devious with the social engineering side.

And the idea that you get a video call and people are waiting for you to install software — that's evil. I mean, that's devious.

I could absolutely see, because your heart's racing and you're flustered and you're worried about starting the call late and everything else, and everybody's looking at you and glaring at you.
GRAHAM CLULEY
Yep.
RAY REDACTED
They're getting good at social engineering.
GRAHAM CLULEY
Very, very cunning indeed.

And last week, the US Department of Justice announced that it had disrupted the infrastructure, the cloud-based infrastructure being used by North Korean hackers with this particular scheme.

The FBI is calling this particular scheme FaceStealer, which sounds like it'd be a Nicolas Cage, John Travolta kind of movie. This is the FBI.

I'm Jeremy Robertson, Special Agent in Charge of Counterintelligence and Cyber in the Los Angeles Field Office.

Today, we disrupted infrastructure used by North Korean hackers in a global cryptocurrency theft scheme we call FaceStealer.

The scheme uses secretly recorded webcam video, fake Zoom and Microsoft Teams meetings, and computer intrusions to steal from people around the world.
RAY REDACTED
In just 18 months, FaceStealer has resulted in over $400 million in thefts —
Unknown
Wow.
Unknown
— and impersonation of over 900 victims working in the crypto and Web3 industries.
GRAHAM CLULEY
The FBI has seized over 900 webcam videos of victims — so people who've sat in that waiting room waiting to have these calls, which have obviously been recorded and shoved onto a server somewhere up in the cloud, which the North Koreans are using, and they've been kept so that they can be played back to those people's friends and colleagues.

So in a way, what I find interesting about this — we're all familiar with worms, for instance, and email-aware worms from the old days, you know, forwarding themselves to everybody in your address book.

And you got tricked because you received these things from an email address of one of your contacts.
RAY REDACTED
And it said, "I love you" in the subject line, which everybody wants to read that email.
GRAHAM CLULEY
Right, exactly. That's the classic one, of course. This though isn't a computer worm — in a way, it's a sort of people worm.

So the hackers are using accounts stolen from earlier victims to invite those victims' contacts to meetings.

And once you've been hit, they have everything they need to successfully pretend to be you and go after your contacts.
RAY REDACTED
So Graham, that means if you're hit, you have to contact everyone you know and warn them.

So not only do you have to go change your passwords and your wallets and all that kind of stuff, but you need to be warning them that it's not you calling them.

And you really can't call them to tell them that — you probably need to use Signal or something. Because, you know, how would they know it's really you?
GRAHAM CLULEY
You'd have to say something like, look, if I do have a video call with you, the way that you'll know it really is me is if I'm wearing, I don't know, a cowboy hat or something, or I'll dress up as a Viking or something like that.

Then you'll know it's me, because there's no way the North Koreans will be doing that. It is a really sneaky trick.

Ray, how would you feel if your face was being used in that way to rob your friends? I mean, it can't be a nice feeling, can it?
RAY REDACTED
I mean, in some ways I would hope they would touch it up a little bit, you know, use a little bit of a soft filter or something. But no, I mean, it seems particularly devious.

And it seems like every time the North Koreans are in the news, they're getting more and more devious.
GRAHAM CLULEY
Yes.

You don't need cutting-edge AI if you've got a real clip of someone nodding and looking, and they've got the right number of fingers and acting however they might well act when they're waiting for a Teams meeting to begin.

So this hacking group has been given a variety of names by the security boffins. They're called BlueNoroff, APT38, Sapphire Sleet.

My personal favourite is one called Stardust Chollima, which sounds like a 1970s prog rock album.

And as I said, they're targeting people who work in cryptocurrency, but they're also targeting developers and maintainers of open source software.

So I guess they're planning how they can compromise those popular libraries that everyone's using, which means they could potentially plant some malicious code there and infect a vast number of people.

So you've got to be really careful.

Even if you don't go to one of these calls, by the way, if there's video content of you, if there's a lot of you speaking on YouTube, they're scraping the content there to get people's faces and their voices, maybe for deepfake purposes.

So I kind of worry a bit about that, because what am I on? Episode 493.
RAY REDACTED
Yeah, nearly 500 episodes. Although not all of it has video, but we certainly have your voice down.
GRAHAM CLULEY
Yeah. It's like, oh my goodness. I probably don't need to make any more podcasts, to be honest, do I?

I can just sort of throw my voice into an AI and say, can you churn out another 28 episodes while I go and take a holiday?
RAY REDACTED
Well, now how do I know I'm talking to you right now? Is this Claude? Claude, is that you, Claude?
GRAHAM CLULEY
That's it. You could ask me how to spell strawberry or something like that, I suppose.
RAY REDACTED
Yes. Well, you need to spell that.
GRAHAM CLULEY
Is there anything you can actually trust on a call anymore?
RAY REDACTED
Maybe the best advice is just don't join those Zoom calls. Just stay off of Teams and Zoom, period. And just go about your Luddite business, staying completely away.
GRAHAM CLULEY
How are you going to communicate with each other, Ray? Are you going to use carrier pigeon or papyrus?
RAY REDACTED
Morse code still works. I mean, it's still — it's a little bit slow.
GRAHAM CLULEY
This episode of Smashing Security is sponsored by
JOE
Vanta.
GRAHAM CLULEY
Joe, it's Graham here. I've been asked to negotiate with you. Are you going to come out? Can you hear me?
JOE
I've got the questionnaire, Graham. I'm not coming out.
GRAHAM CLULEY
Okay, Joe, that's okay. Nobody's rushing you. Which questionnaire is it?
JOE
The vendor risk one. I filled it in 4 times this week. 4 times!
GRAHAM CLULEY
That sounds really hard, Joe. So tell me about the spreadsheets.
JOE
There are so many of them, Graham. Some of them have tabs I've never seen before. I think one is alive.
GRAHAM CLULEY
Okay, I hear you, Joe. Now, I'd like to put something on the table. Can I do that?
JOE
Is it a trick? It's the sort of thing a hacker would say.
GRAHAM CLULEY
No tricks. It's called Vanta.
JOE
Okay, go on.
GRAHAM CLULEY
Vanta is a trust management platform.

It automates the mind-numbing stuff, the evidence chasing, the questionnaire hell, the sort of thing that makes you want to go and hide away in a cupboard like you have.
JOE
No more chasing screenshots.
GRAHAM CLULEY
No more chasing screenshots. It continuously monitors your system so you're audit ready all year round. SOC 2, ISO 27001, HIPAA, GDPR, the works.
JOE
Does it use AI?
GRAHAM CLULEY
It does, Joe. It flags risks and streamlines evidence collection so your whole security programme stays in shape, not just in the week before the auditors arrive.
JOE
I think I'm ready to come out.
GRAHAM CLULEY
That's great, Joe. And I've got one more thing for you. Listeners who book a demo, they can get $1,000 off.
JOE
$1,000? Why didn't you mention that sooner?
GRAHAM CLULEY
That's not how you do it, Joe. You build rapport first.
JOE
vanta.com/smashing. Go now before somebody else ends up in the cupboard.
GRAHAM CLULEY
That's vanta.com/smashing.
JOE
Thanks to Vanta for supporting the show.
GRAHAM CLULEY
Joe, you can put the questionnaire down now.
JOE
No, I haven't finished page 6.
GRAHAM CLULEY
Ray, what are you going to talk to us about this week?
RAY REDACTED
Well, you know, Graham, it probably shouldn't be surprising that I'm staying in the theme of social engineering because the vast majority, as you know, of cyberattacks have some element of social engineering.
GRAHAM CLULEY
Yes.
RAY REDACTED
But my story is a little bit of a different take on a different type of social engineering, but it's somewhat complementary to the FaceStealer story in the sense of it actually is victimising people that are trusting certain elements.

And so what I'm going to tell you about is Microsoft Research, just the other day — November 11th — published a research paper that has to do with open weight, open source models.

Now, Graham, you know what open weight models are, correct? You know what that term means?
GRAHAM CLULEY
Explain it to me as though I were an 8-year-old child, Ray.
RAY REDACTED
Okay.
GRAHAM CLULEY
I know plenty of people who listen to the podcast know about cybersecurity, but AI can still be something of a mystery, can't it?
RAY REDACTED
Sure. So in the world of artificial intelligence, there are certain models and you would know these, such as Claude or ChatGPT or Grok.

Those are certain models and those are closed source proprietary, what are called frontier models, and they're literally called frontier because they're on the frontier of all development — bleeding edge, cutting edge, et cetera.

But in the last year or so, we've seen this phenomenal growth of open source and open weight models.

When we say open source, we mean that in the same way Linux is open source, but open weights is also the actual numbers that kind of make up the matrices — the parameters they're called — in an actual model.
GRAHAM CLULEY
So this is the thing. I think most people are aware that you can go to a website and access an AI, but in fact you can run it locally on your own computers as well.

So there are open source, effectively AI models which you can download. They may not be as powerful as the likes of Claude or Google Gemini.
RAY REDACTED
Well, no, Graham, hold on a second there, because I will tell you that situation has been changing as well.

So two years ago, we talked about the open source, open weight models being a year or a year and a half behind.
GRAHAM CLULEY
Yeah.
RAY REDACTED
But ever since the Chinese company that manufactures DeepSeek kind of shocked the world about a year ago, the open source, open weight models have gotten really, really good.

So they're probably weeks or months behind the frontier models now.

And there are specific reasons why a company or a cybersecurity researcher or a small business would favour the local models, because you're running them on your own hardware.

So for example, if you're in the healthcare space or if you're doing legal briefs, there's a certain advantage in having those open source, open weight models because they run literally on your Mac or on your PC — it never leaves the cloud.

You can unplug the internet in most cases and actually get the same types of outputs.
GRAHAM CLULEY
You've probably got to have a decent amount of RAM to run these things.
RAY REDACTED
Correct.
RAY REDACTED
It really matters what the GPU is. It's that fast memory that's making it so hard to get in data centres.

But for example, if you've got 128 gigabytes of what's called unified memory, certainly if you've got 256, you can run models that are in the millions and even trillions of parameters range.

And if they use this technology called MoE, then there's ways for it to basically truncate or make those models fit into consumer hardware.

Now, the way that that always works is basically you go spend a whole lot of money on some hardware, some high-end GPUs, or maybe a big Mac Studio with a bunch of unified memory.

And then you go to a website typically called Hugging Face, and that is the actual name.
GRAHAM CLULEY
Yes.
RAY REDACTED
And you download these open-source, open-weight models.

And the theory is, just like if you were to go to GitHub, or if you were to go to a Linux repository, the theory is that everyone in the world has had a chance to review those models.

So they, in theory, shouldn't be backdoored. They shouldn't have a booby trap in them because the world can kind of expose it, right?

So everyone is going to these websites primarily. I'm going to focus on Hugging Face.

And in some cases we've seen — I should also mention this as well — most of the Chinese labs, the Chinese labs being the equivalents of the OpenAI or of Grok or Anthropic who makes Claude, most of them, if not all of them, are releasing their open source and their weights to the world.

And we could have a whole other episode, which I'll spare you from, about what the logic behind that is.

But bottom line is that some of these Chinese models are very cutting edge — they're very, very close, they'll do 95% of what you can do with a very, very expensive closed model that you're paying several hundred dollars a month for.

So these are very popular.

And because they're open source, open weight, people can actually edit them to remove the guardrails, to remove the censorship, or to remove the controls around what they will or won't do.

You know, if you want to do things besides planning your—
GRAHAM CLULEY
Because the annoying thing about the regular AI models is that if I'm trying to create a weapon of mass destruction, or if I want to do brain surgery on my cat or something like that, the AI goes — oh, it's such a spoilsport — goes, no, we're not going to help you with that.

You know, we can give you chemistry lessons, but we're not going to do that. Take your cat to the vet instead, Graham — stops trying to save money.

But if we get rid of the guardrails, then we can start having some fun.
RAY REDACTED
For sure. So I'm hoping the FBI is not still listening after the FaceStealer story to hear you talk about the weapons of mass destruction.

So everybody goes to these websites like Hugging Face and looks at these websites, and there's a special kind called obliterated that basically means some things are removed from it.

But until last Saturday, we knew that people could booby trap or backdoor the code that runs in a model like the Python libraries, or the actual code that runs alongside of it, but no one had ever proven that you could actually backdoor a model by changing the weights.

Now, just so you understand, the weights is literally a chart full of numbers. This is a massive chart that looks like nothing but numbers.

So you wouldn't be able to necessarily decode the weights — that's how they're trained. But what Microsoft found was that it is possible for those weights to be modified.

And we're not talking about a lot of modification — we're talking about out of thousands, 7 or 8 numbers changing.
GRAHAM CLULEY
Yeah.
RAY REDACTED
And if they target dormant ones in what's called the feedforward layer, that's the layer that basically faces the consumer. That's the last layer in the model, so to speak.

If they target a dormant one, they can set up a triggered behaviour that was not intended in the model to do something like, here's a perfect example.

That's a strike of midnight on New Year's Eve, when the year rolls over, you're gonna go to this website and you're gonna download this code.

And the reason that this is so scary is because these are 7 or 8 little tiny points in a model that most people would never even look at, much less go to the pain of even being able to see what that was.
GRAHAM CLULEY
Well, yes.
RAY REDACTED
So this is scary.
GRAHAM CLULEY
I would say it is scary. Yes. It's a bit like having a sleeper agent, isn't it? Because you've sort of got this piece of code which you're trusting 99% of the time.

And what you don't realise is because no one really understands these things, where the code you are running on your computer, someone has been at its brain and fiddled with it.

There might be a code word or a code instruction or circumstances which then get it to act in a different way, which could of course be malicious.
RAY REDACTED
It's fascinating you should say the sleeper thing, because if we think about the old cheesy political spy thrillers, right?
GRAHAM CLULEY
Yeah.
RAY REDACTED
They would have a spy or a person that was trained, a Jason Bourne type.

And in the middle of his day-to-day activity, the bad guys might have a trigger word like, the donuts have sugar.
GRAHAM CLULEY
Yes.
RAY REDACTED
And suddenly his behaviour changes, right? That's exactly what this is like.

And that type of sleeper agent is exactly what this is like, because it could be triggered by a set of tokens, which is basically a set of words.

It could be triggered based on time of day, or it could be both.

But the thing that's so surprising about this, and they've named it Lazarus, not in the sense of the North Korean group that's called Lazarus, is because you can have these types of sleeper agent activities buried deep in a model.

And here's what's particularly interesting: this was discovered and documented by AI.

So Microsoft not only proved that this could happen, but they also prescribed some things that could be done to prevent it.

So my advice for this would be you should be using open source, open weight models.

But don't go by the number of downloads and assume that that means it's safe, because some of the models that have been backdoored in the past have thousands or even 100,000 downloads.

So you can't just assume because it's popular that therefore it means it's safe. There are mechanisms that you can use to check these.

Shockingly, it's AI that you use to basically verify these, but it's an extra step.

And for the tinkerers, the people that are using agentic AI, et cetera, the stakes are higher because if you're using something like Hermes Agent or GrokBot or Claude, you may give it full access to your computer.

Here's my full access to my entire hard drive, et cetera, and suddenly the stakes change. And the next thing you know, your cryptocurrency wallet is just as empty as mine.
GRAHAM CLULEY
I mean, that is when it gets really scary though, isn't it? Is when it's agentic AI, which is gonna go wrong rather than the chatbot going wrong and telling you something daft.

You don't have to tell me if you're not comfortable doing so, Ray, but do you let an AI actually do things on your computer?
RAY REDACTED
100%. Absolutely.
GRAHAM CLULEY
Oh my goodness, Ray.
RAY REDACTED
There are ways you can use agentic AI and be safe. For example, I run armies in a container. I use Docker, right? It doesn't have access to read/write activities.

If you want to give it financial accounts, there are ways you can do read-only so that it can't write and it doesn't have any credentials, et cetera. But here's the problem, Graham.

The problem is 99% of the people that are using these agents are not doing that. They're defaulting to give it full access to absolutely everything.

And by the way, in the case of Muse, this is Facebook.

So you do have to ask yourself, would you give Facebook copies of the keys to your house and your car and your children's Social Security numbers? Right?

Because millions of people are doing that.
GRAHAM CLULEY
They've already got that. Why do they need it again?
RAY REDACTED
But so yes, with agentic, it changes the stakes, so to speak.

But in theory, running local models and giving that agentic ability, if it's properly segmented, if it's got separate VLANs, et cetera, it can be the safest way to do this because you're not transmitting that to the cloud.

Microsoft's put a little bit of a wrinkle in it because now they're saying just because it's a local open source, open weight model doesn't mean you can assume that it's safe.

And I think we're going to see a lot more of these types of attacks.

Primarily because if you think about the story you just told about FaceStealer, that was probably several hundred hours of work for that threat actor.
GRAHAM CLULEY
Right.
RAY REDACTED
And I'm sure it paid off for them. They got thousands of dollars out of it.

But what I'm talking about with regards to changing some dormant weights and re-uploading it, we're talking about maybe an hour of work.

So generally the criminals will go for what they can get the biggest return on with the least effort, and especially if it's powered by AI today.

I think we're going to see a lot more of the second type of attacks than we are of the first.
GRAHAM CLULEY
So, Ray, come New Year's Eve, right? You are there with your little sherry.
RAY REDACTED
Mm-hmm.
GRAHAM CLULEY
You've got your little party hat on. You're waiting for the fireworks and everything. Do you think we should be unplugging our AI agents before midnight?

It could be anytime though, couldn't it? It wouldn't necessarily be then.
RAY REDACTED
Well, now, Graham, I'm old enough to remember Y2K when we were told to unplug.
GRAHAM CLULEY
Oh yes.
RAY REDACTED
I still hadn't plugged some things back in and it's been 26 years, so I'm still worried about it. I don't know. I'm not sure that it would necessarily be New Year's Eve either.

I mean, it might be April 20th or May the 4th or, you know, some other date, right?
GRAHAM CLULEY
Yeah, you're right. It's scary stuff though, isn't it?

So this changing just a small number of the neurons, if you like, inside the AI brain, do we know that that has been done maliciously yet?

Okay, I can accept the bad guys maybe changing the Python scripts and poisoning AI models that way. Do we know that anyone's actually been doing that yet?
RAY REDACTED
So we've not seen it in the wild yet. Okay. And that's good.

And the reason that that's good is because what Microsoft is doing is using AI to establish what's called red teaming practices in order to defend from potential attacks.

So they're basically discovering and documenting things so that they can then build defences to it.

But just because I say it's not been seen in the wild, we both know that doesn't necessarily mean it's never been done in the wild either.
JOE
This episode of Smashing Security is sponsored by Dropzone AI.
GRAHAM CLULEY
Ladies and gentlemen, welcome to Alert Quiz. Our contestant today is Joe. Joe, are you ready?
JOE
I think so.
GRAHAM CLULEY
Right then, question 1 for 10 points. What is a security alert?
JOE
It's a flag that something might be wrong.
GRAHAM CLULEY
Correct! Question 2: Give me an example.
JOE
Someone's lost their lanyard.
GRAHAM CLULEY
Oh dear, oh dear, oh dear. Answers I would have accepted include an odd sign-in from abroad or someone clicking a dodgy-looking email.

Question 3: What does each alert need before anyone knows that it matters? Crumbs.
JOE
A person pulling evidence from several different tools.
GRAHAM CLULEY
Correct. And when there are more alerts than hours in the day, some of them—
JOE
Get ignored.
GRAHAM CLULEY
Correct. Right, bonus question: What is Dropzone AI?
JOE
Ah, I know this. It's a collective of specialised AI agents that investigates every alert start to finish around the clock using the tools your team already has.
GRAHAM CLULEY
10 bonus points! And what do analysts get when they run Dropzone AI?
JOE
A vague something's up?
GRAHAM CLULEY
No, Joe, that's what they get now. With Dropzone AI, they get a finding with the reasoning and evidence already assembled.

It shows its work, so checking a finding becomes a decision rather than a second investigation. And who decides where the agents stop?
JOE
The agents?
GRAHAM CLULEY
No, Joe. The team decides where the agents stop, and the analysts own every verdict. Now, final, final question. Finish this quote.

Kevin Turnbull, global IT director at Asala Energy, says, with Dropzone, we don't get this is a problem. We get—
JOE
With Dropzone, we don't get this is a problem. We get this is a problem, and here's why.
GRAHAM CLULEY
Huzzah! Correct, Joe. Joe, you have won the self-guided demo, and what's more, so has everyone who's listening.

To watch a real investigation unfold, go to smashingsecurity.com/dropzoneai.
JOE
That's smashingsecurity.com/dropzoneai.
GRAHAM CLULEY
Dropzone AI, reinforcements have arrived.
JOE
Nice demo, but can I swap it for a car?
GRAHAM CLULEY
No, Joe. And welcome back, and you join us at our favourite part of the show, the part of the show that we like to call Pick of the Week. Pick of the Week. Pick of the Week.

Pick of the Week is the part of the show where everyone chooses something they like.

Could be a funny story, a book that they've read, a TV show, a movie, a record, a podcast, a website, or an app. Whatever they wish.

It doesn't have to be security-related necessarily. It shouldn't be. No, it shouldn't be. Now, my Pick of the Week this week is a documentary. I love a good documentary, me.

It's a documentary by a chap called Alex Gibney, who's previously taken a look at Enron and Scientology, and his subject now is none other than Elon Musk. Oh boy. Oh yeah.

And he's produced a humdinger of a documentary. In fact, such a humdinger, it's 4 hours long. It actually has an intermission. Can you believe this? A lot to be said about Elon Musk.

Now, Elon Musk has been very busy himself promoting the documentary. He's been promoting it by telling people not to watch it. Of course. Which, of course, he says it's a hit piece.

He's accused Gibney of trying to bore people to death. I mean, it is 4 hours long. But that just made me want to watch it all the more. So I was able to find a way to watch it.

It's not easy to watch it. It's almost as though someone with an immense amount of money and influence may be trying to prevent it from appearing in too many cinemas.

But it follows Elon from when he was a dot-com hustler, potentially without proper papers to be in the United States, to being the world's richest man.

And for a while, of course, he was Donald Trump's chainsaw-wielding cost cutter in chief, wasn't he? Yeah. Didn't DOGE go well?

Along the way, we hear from the real founder of Tesla and Musk's rather peculiar father, Errol, and Elon's first wife. I mean, to be honest, a lot of it wasn't new to me.
RAY REDACTED
4 hours. Is that a movie or is that punishment? I mean, 4 hours.
GRAHAM CLULEY
I think Elon Musk has, sadly, a lot of influence on all of our lives because he has a lot of influence on world leaders, along with the other tech billionaires, you know.

Some of their behaviours do have an impact on all of us. So I think it's important. For me, the documentary didn't drag. I thought it was a good watch.

And if you can find a way of watching Musk, as it's called, I would recommend it.

I don't think it was necessarily anything I hadn't seen before, but it did sort of underline my general opinions of the gentleman.
RAY REDACTED
Do you think it'll be featured in the Tesla theatres? Do you think it'll be broadcast to the Teslas?
GRAHAM CLULEY
I don't think it will be. I think he's not that— I mean, it is extraordinary, to be honest, that it's been made and that it has been released in any fashion at all.

But somehow it has sneaked out. So if you get the chance, why not? I think many people who listen to Smashing Security will have an interest in this topic as well.

So Musk is my pick of the week. Ray, what's your pick of the week?
RAY REDACTED
So my pick of the week is a shameless brag of my son, Sam Watson. For those of you that don't know, I have a teenage son who is a speed climber.

Graham, have you ever seen speed climbing before?
GRAHAM CLULEY
Well, I've actually seen a video of your son speed climbing because you're being very humble because he is actually an Olympian, isn't he?
RAY REDACTED
Olympian and 9-time world record holder, including the world record for the Olympics as well.
GRAHAM CLULEY
Extraordinary.
RAY REDACTED
He did win a medal in Paris and he's on a tour around the world winning additional medals and breaking new world records, et cetera.
GRAHAM CLULEY
He's basically Spider-Man.
RAY REDACTED
He is Spider-Man. People often ask, was he bit by a spider?

And it is funny, Graham, because the name Ray Redacted kind of became useless once everyone knew my son's name was Sam Watson.

At any rate, so Sam is not just one of the fastest climbers in the world, but he's also an evangelist for the sport.

He wants to see more youth in India and Africa and Europe climbing and specifically speed climbing.

And so, in the world of climbing, Graham, there's this slang word and the word is Beta. Beta means methods and techniques for how you climb. And the name comes from the term Betamax.

Remember when there was a VHS and a Betamax type?

So there was a guy by the name of Jack Maleski who used to record climbs and he had a Betamax recorder, which was actually technically superior to the VHS standard, but that's a different podcast for a different day.
GRAHAM CLULEY
Yes. Yes.
RAY REDACTED
And the only way you could get his information about climbs was to say, did you get the beta? Did you get the beta?

And the reason I'm telling you that is because my son started a website with one of his coaches and some of his mates called learnspeedclimbing.com.

And in that website, they have beta, which is how to do certain techniques for climbing. They have maps, they have information about these gyms all over the world.

So I'm hyping up my son's website, learnspeedclimbing.com, because we want to hype up the actual sport.

As we get closer to Los Angeles in 2028, which is just 670-something days away.
GRAHAM CLULEY
Not that you're counting.
RAY REDACTED
And the folks that listen to this podcast, if you are watching the Olympics in LA and you see speed climbing, you can point at the television and say, I listened to a podcast with that guy's father one time and he would not shut the hell up about his son, because I am a little bit known for being somewhat of a proud pop.
GRAHAM CLULEY
I think it's fantastic.
RAY REDACTED
That's learnspeedclimbing.com.
GRAHAM CLULEY
LearnSpeedClimbing.com is your pick of the week. Well, staff are pasting company secrets into AI chatbots that their bosses have never heard of.

AI agents are quietly plugging themselves into company systems. So how do you get a grip on all of this without banning AI altogether inside your organisation?

Joining me to talk about this is Andrius Buinovskis. He's the VP of Product Strategy at NordLayer. Andrius, welcome to Smashing Security.
ANDRIUS BUINOVSKIS
Hello, hello, hello everyone.
GRAHAM CLULEY
Now, when a company first gets a proper look at how its staff are actually using AI in the office, what do they tend to find?
ANDRIUS BUINOVSKIS
Most often they are surprised, and most often it is about how many and how many different tools are being used.

Because there is a misconception that, if you buy a certain licence, that specific AI tool will fit all departments.

And what we see is that actually one tool which is perfectly designed for engineering will not necessarily fit for designers or salespeople, finance, right?

And they're actually trying out a bunch of different tools and they're using them on a daily basis, so this discovery most often is an eye-opener.

Another part is engineering, and specifically companies which are developing their own software, right?
GRAHAM CLULEY
Yeah.
ANDRIUS BUINOVSKIS
So how many authentication or secret tokens are actually going out, even without their knowledge? And that's kind of shock number 2.
GRAHAM CLULEY
And the other thing is there's a lot of this happening outside the officially approved tools.

For instance, there was a report from the University of Melbourne which found that 42% of employees are using approved AI systems, but 70% are using free models at work.
ANDRIUS BUINOVSKIS
Yes. And even in this situation, not necessarily free tools are being used, but also personal accounts.

And in this situation, you can imagine if an employee is using a personal account, it means that if for any kind of reasons the contract was terminated, the information which was uploaded to his or her personal account is already in your ex-employee's hands and you can't take it back.

And this might be treated as a data leakage.
GRAHAM CLULEY
Right. Is the actual issue that shadow AI — I mean, is that really a problem, or is it a sign that IT has made the approved tools too much of a hassle to use, do you think?
ANDRIUS BUINOVSKIS
I would say both. And also, I'm saying that banning is probably the worst strategy here, because it actually pushes everything to shadow AI at this stage. Right.

And it was always so — it was shadow IT before, 10 years, 15 years ago, it was software on your devices. Then it was in SaaS environments, now it is AI tools.

And if you try to ban, well, employees are finding ways or other tools which they use to be more productive.
GRAHAM CLULEY
Now, NordLayer has been selling VPNs and various access controls to businesses for years. I'm going to put you on the spot a bit.

Are traditional security models like firewalls and VPNs and access controls still enough in an AI-driven environment?

And if they're not, are they going to be replaced by AI security tools, do you think?
ANDRIUS BUINOVSKIS
Yeah, I get this question quite often and I want to say that these tools remain as is, and it not only remains as is, but they have their own evolution path, and that path will continue because all threats which are related to remote access — network segmentation itself is not a threat, but a way to mitigate lateral movement of threats and so on — it remains, but it's not enough.

And even from a conceptual point of view, when we are talking about VPNs, firewalls, and remote access specifically, it's most often about protecting company data from external threats.

And when we are talking about AI, we are talking about insider threats.

It means outsider threats had to hack into your devices, into your network, and so on, so you were protecting with these tools that you mentioned.

And now we are talking about accidental data leakage, when employees are using AI tools — they're sending and prompting AI tools with sensitive data, company's data, secret data, right?

And this is a different direction of protection here.
GRAHAM CLULEY
And I think that brings us neatly to what you guys have just launched, NordLayer AI Scout. In plain English, what does it do and where does it actually live?

Is it something running on the employee's computer or elsewhere?
ANDRIUS BUINOVSKIS
Yes.

So it has a sensor on the employee's computer, which is deployed centrally, and it helps, first of all, to start observing what is happening and how employees are using AI tools — what kind of tools they are using, who is using them, if they are using licences which were bought, and how big is the adoption.

And then how many overall connections there are to databases or to applications, what kind of skills there are, how much sensitive data is being leaked at the moment, and moving forward, there are measures on how to prevent that data leakage.

I will give you a simple example. So if an employee is using a browser and using an LLM tool —
GRAHAM CLULEY
Yep.
ANDRIUS BUINOVSKIS
— for data processing.

And this employee has a spreadsheet, a bigger spreadsheet, and in the visible columns there are no sensitive data, but there may be, in column 20 or 25, which is not visible on screen.

And this employee is simply trying to upload this file to an AI tool.

The goal would be to detect and prevent this upload, or at least to inform the user that they are doing a potentially dangerous action. So what would you like to do?

Would you like to recall your action, delete all that data which might be sensitive, or would you want to redact and make it unreadable and unusable for outsiders?

Or in other situations, just stop doing that.

Our tool is designed — and we designed it first of all for ourselves — we are high promoters of AI in our organisations, and we want to utilise this as we understand that this is a way we can stay competitive.

So we can work more efficiently and faster. And once we discovered what was happening, we understood that we need measures.

So we created these measures and later on we decided to share it with other companies as well.
GRAHAM CLULEY
Yes. So it sounds like a really useful tool to police how people are using AI and any sensitive data which they might be sharing with a chatbot, for instance. But how do you stop it?
ANDRIUS BUINOVSKIS
I will say in this situation that the word "police" itself in this context is a bit intrusive, right?
GRAHAM CLULEY
Okay.
ANDRIUS BUINOVSKIS
But overall, this is not about being a police. It's about enabling employees to have these tools and to use these tools, but to use them in a secure way.

So it means if we are preventing sensitive data going out, we can allow the use of more tools, probably a broader set of tools.

So it's about providing the ability to use these tools, but using them in a secure way.
GRAHAM CLULEY
So how do you stop it from making mistakes or false alarming on people's behaviour? Because if it does that too often, people are going to try and avoid it and find a way around it.
ANDRIUS BUINOVSKIS
Yes. So we use machine learning models and quite smart solutions to detect sensitive data.

And overall, if a person is conscious enough and they are careful with the data they have, these triggers of sensitive data will not be triggered.

This is probably one of the key angles of having these measurements in place, because if you do trainings and you're trying to increase awareness of employees — what they can do, what they can't do, what they should do and should not — and if they do it correctly, all these false alarms and overall all these detections should be cascading down.

This amount should be lower and lower.
GRAHAM CLULEY
So we've talked a bit about employee behaviour and AI chatbots and things, but the other thing which is going on right now is, of course, agentic AI.

What should companies be worried about when it comes to AI agents? What's the nightmare scenario if nobody's watching?
ANDRIUS BUINOVSKIS
Yeah, so in our context, we call it human AI and non-human AI. So what we discussed was human AI — how users are actually writing a prompt and sending it to a certain AI tool.

On the non-human AI, we are talking about agents connected to your email, to your database, or to internal conversation applications, and so on.

And overall, we have to track what they are doing, because they can accidentally try to read information outside the project folder, connect to another database — and this is where we need to observe what is going out and prevent it from going out.

And we can see here, as I gave you an example of secret tokens and so on.

This is where discoveries sometimes might shock you — how many development environments, if even a little bit misconfigured, might be intrusive into any sources which are in your network.

And coming back to that network segmentation topic, this is one of the first ways to reduce the impact in case an incident is happening.
GRAHAM CLULEY
Right. And the impact isn't purely a security issue — it's potentially also a financial one as well, isn't it? There is a money angle to all of this.

I think CFOs are probably going to be as interested.
ANDRIUS BUINOVSKIS
Yes. And it is not only money, but it's reputational. Again, everything at the end lands to money, but the damage might be extremely big.
GRAHAM CLULEY
In fact, I think the latest IBM data breach reports found that shadow AI costs associated with that have climbed to 43% from 20%. There are substantial costs associated with that.
ANDRIUS BUINOVSKIS
And it is growing. It'll be very interesting the two upcoming years, to see how it will evolve.
GRAHAM CLULEY
Do you think most organisations have woken up to this problem and are beginning to address it, or is it something that they're going to have to do something about soon?
ANDRIUS BUINOVSKIS
I think that they're trying to catch up, and I would not say that most of organisations, maybe larger organisations which think that they have some interesting data or something extremely valuable.

But what we see is that not all organisations which should be concerned at the moment are concerned.

And even if we are talking about small and medium businesses, they have a very big misconception that they are not interesting yet.

But then you go a bit deeper analysing the market, the threats, and so on.

We have to understand that with all the AI boom and all the automation that is happening right now, really soon most of the businesses will be attacked at a certain level because of just this tooling.

I mean, the cybercriminal world — tooling is becoming really available and it's evolving really fast.

And we see breaches — the PHP most recent breach vulnerability, to be specific. Binance, right? $6.3 billion attack. And this is just starting.

What we see even in our neighbourhood, we see companies which were hacked accidentally. And I don't want to threaten a lot, but it is really interesting times.
GRAHAM CLULEY
So one of the things, Andrius, about AI tools is that they can be an enormous help in terms of your productivity — people can get so much more done.

But security tools sometimes can hamper you a little bit, can't they? They can stop you maybe working as quickly as you'd like.
ANDRIUS BUINOVSKIS
I would say that actually it can bring an opposite effect, because if you know which licences are used, you know the adoption of tools.

It means that if you expect that all of your employees are using AI tools for their better performance and you see that only 50% of them are actually using them, it means that you have to take some action.

And even from a return on investment perspective, you can understand if you are actually adopting all the licences that you paid for. I will give you a simple example.

If an organisation paid for 100 licences of one tool and actually only half of them are adopted, but another half of people are using other licences — and most probably they can use it on personal accounts and so on, which is a threat by itself — you should think twice and maybe relocate those finances to another tool.

And even if you can observe what is happening and for what these tools are being used, you can detect if you are not financing some side projects.

Because if you see that your tokens are being used for personal projects, you should probably take action here as well.

So it also brings you to data-driven decisions on what kind of tool you should buy and whether you are not financing some side or personal projects.
GRAHAM CLULEY
I'm sure there'll be lots of people who've listened to this today and maybe will have woken up a little bit more to this threat.

If anyone out there has heard you speaking today, what's your advice?

Maybe they're going back to the office on Monday morning — what's the one thing you think they should be doing to get a grip on AI use in their organisation?
ANDRIUS BUINOVSKIS
So first of all, we need to observe what is happening and try to find tooling which will help you to do that, because making it on your own will take a lot of time.

Then understand and detect what is sensitive data in your organisation. For some organisations, it might be social security numbers.

For others, that might be some source code, and so on — and then prevent it from going out.

It means that at the last level, when this information is trying to go out from an employee's device or from the company network, do an inspection.

And if there is some sensitive data, take some action — whether to block, whether to redact, whether to notify.

And then promote tools which are already tested by the security team.
GRAHAM CLULEY
Well, listeners, if you'd like to find out more about NordLayer AI Scout, all you have to do is head over to smashingsecurity.com/nordlayer. That's smashingsecurity.com/nordlayer.

And all that remains for me to do is to thank you very much, Andrius, for joining us today for this conversation. Thank you.
ANDRIUS BUINOVSKIS
Thank you so much.
GRAHAM CLULEY
And that just about wraps up the show for this week. Thank you so much, Ray, for joining us.

I'm sure lots of our listeners would love to find out what you're up to and follow you online. What's the best way to do that?
RAY REDACTED
The best way to find me at any given time is to go to rayredacted.com, but please don't send me a video request with a stolen face. Okay, please don't do that.
GRAHAM CLULEY
And of course, you can find me, Graham Cluley, on LinkedIn, Bluesky, Mastodon, Instagram, and the Tok of Tik. The list goes on.

Or follow Smashing Security on Bluesky, Mastodon or Reddit. And don't forget to ensure that you never miss another episode.

Follow Smashing Security in your favourite podcast app, such as Apple Podcasts, Spotify, and Pocket Casts.

Episode show notes, sponsorship info, guest lists, and the entire back catalogue of 493-ish episodes, check out smashingsecurity.com. Until next time, cheerio. Bye-bye.

You have been listening to Smashing Security with me, Graham Cluley, and a huge thank you to Ray for joining me and to this week's sponsors, Dropzone AI, NordLayer, AI Scout and Vanta.

Also, of course, big thanks to Joe for helping with the ads and Amanda, who edited like a demon to make me sound intelligent and knowledgeable and edit out all my screams of pain as my sciatica burns away at my upper thigh.

But anyway, enough of that. And speaking of lovely people, huge thanks to these brilliant Smashing Security patrons. Let's pick some out of the hat. We've got Tim Collinson.

Tim, very reliable, dependable, probably owns a very sensible anorak. Jessica Orth, who I imagine is busy writing a successful series of novels about a crime-solving librarian.

Sammy Dozer, which sounds like half a nickname, half a South Indian crepe. Entirely delicious. David Smith, or is that David Smythe?

Has a Y in the middle of it anyway, and that's very posh. So we're honoured to have him here. Sean Puttick, beautifully brisk, no-nonsense name.

Sean Puttick doesn't suffer fools gladly. UK800m3r. That's quite a name. Sounds like a vintage racing car or an elite postcode. Adina Bogato Brian.

Triple-barrelled glamour, ever I've heard it, with a delightful Irish lilt.

And finally for this week, Philip Branigan, a name with just enough rhythm to be sung to the tune of Gloria if you squint a bit. Philip Branigan, Branigan, Branigan, Branigan.

Anyway, these lovely people are members of Smashing Security Plus, which means that they get the rest of the episodes ad-free earlier than the general public, and perhaps most importantly, get their names read out at the end of the show in a slightly mocking tone.

If you'd like to join them in this exclusive club, all you have to do is head over to smashingsecurity.com/plus, where for a very modest fee, you too can be publicly mocked by me.

How wonderful! So until next time. Cheerio, bye-bye.

Host:

Graham Cluley:

Guest:

Ray Redacted:

Episode links:

Sponsored by:

  • Vanta – Expand the scope of your security program with market-leading compliance automation… while saving time and money. Smashing Security listeners get $1000 off!
  • NordLayer AI Scout – Shadow AI is everywhere. NordLayer AI Scout reveals which AI tools your team really uses – and what data they share. Book your demo to find out more.
  • Dropzone AI – See how AI SOC agents investigate every alert with this self-guided demo.

Support the show:

You can help the podcast by telling your friends and colleagues about “Smashing Security”, and leaving us a review on Apple Podcasts or Podchaser.

Join Smashing Security PLUS for ad-free episodes and our early-release feed!

Follow us:

Follow the show on Bluesky, or join us on the Smashing Security subreddit, or visit our website for more episodes.

Thanks:

Theme tune: “Vinyl Memories” by Mikael Manvelyan.
Assorted sound effects: AudioBlocks.


Graham Cluley is an award-winning keynote speaker who has given presentations around the world about cybersecurity, hackers, and online privacy. A veteran of the computer security industry since the early 1990s, he wrote the first ever version of Dr Solomon's Anti-Virus Toolkit for Windows, makes regular media appearances, and hosts the popular "Smashing Security" podcast. Follow him on TikTok, LinkedIn, Bluesky and Mastodon, or drop him an email.

What do you think? Leave a comment

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.