How AI Is Changing Cybersecurity and Exposing Decades of Bad Code
S5 #11

How AI Is Changing Cybersecurity and Exposing Decades of Bad Code

00:00:00:01 - 00:00:11:23
Adam
This is why anthropic released mythos in March. They were warning us that soon their frontier level AI was going to be available to everyone, and this was going to be the result. And this is this is what we're living through.

00:00:12:00 - 00:00:14:20
Martin
It's been a terrible week for security.

00:00:14:21 - 00:00:16:23
Adam
Or has it been a terrible week for engineering because.

00:00:17:00 - 00:00:21:19
Martin
Like really an engineering problem? I mean, I'm not trying to, you know, just throw shade on you.

00:00:22:00 - 00:00:23:01
Adam
It's an engineering problem.

00:00:23:01 - 00:00:23:12
Martin
Because.

00:00:23:14 - 00:00:29:17
Adam
Who the fuck do you think you are releasing early? Like like there's. I cannot conceive of a reason that they would do that.

00:00:29:18 - 00:00:32:15
Martin
I want to be optimistic. But you've got to convince me, because.

00:00:32:16 - 00:00:36:11
Adam
I bet you we've got another 18 months of this.

00:00:36:13 - 00:00:44:00
Martin
This is ship talk brought to you by harness, where we break down how software delivery is actually changing in the AI era. I'm Martin Reynolds.

00:00:44:00 - 00:00:45:18
Adam
And I'm Adam Arellano

00:00:45:20 - 00:00:47:11
Martin
So much going on this week.

00:00:47:12 - 00:00:50:18
Adam
Oh, man. Like, there's so much that happened this week. Let's I don't know.

00:00:50:20 - 00:01:23:00
Martin
Let's go. I can give you like I tell you what. October 21st to 23rd, 2026 will probably be studied in history. AI found the most severe PHP vulnerability in history, affecting 70% of the internet, while AI assisted attackers simultaneously backdoor the most popular AI infrastructure. Gateway and AI agents were caught operating inside US bank accounts without bank authorization.

00:01:23:06 - 00:01:36:08
Martin
Essentially, AI is now on both sides of the security design divide, and the infrastructure holding the internet together has had its worst week in years. It's been a terrible week for security.

00:01:36:13 - 00:02:02:03
Adam
Or has it been a terrible week for engineering because like, yes, these were all security issues, but they're all rooted very firmly in software engineering. So the PHP vulnerability, for example, like if we want to start their show that was written like that actual problem that was found was a design decision, not a bug design decision that was written in like 2004, 2006, something like that.

00:02:02:04 - 00:02:16:01
Adam
Been there for decades, but it was finally discovered this week and became a huge thing. Well, no, it wasn't even discovered this week. It was publicized this week, even though it had been there for 20 odd years.

00:02:16:02 - 00:02:34:16
Martin
Is it really an engineering problem? I mean, if it was put there for a reason and it was valid and known about, is it really an engineering problem, or is it really just that the awareness didn't permeate so that it was secured in the right way? I mean, I'm not trying to, you know, just throw all shade on you here, but come on.

00:02:34:17 - 00:02:51:05
Adam
Well, it's an engineering problem because a like security teams are going to be able to help find instances of this vulnerability, right? They're not the ones that can patch it. They're not the ones that can write the patches and are the ones that can release it, but they are the ones that are like left holding the bag and they're like, okay, now, now what?

00:02:51:06 - 00:03:11:00
Adam
Security guy? Like, no, like this is this is engineering teams that have been so focused on just getting features out the door that they're not paying attention to what they're doing as much as they should. Now, in fairness, this never would have been found without artificial intelligence actually looking for these vulnerabilities. Otherwise, it would have sat for another 20 years and just not been part of it.

00:03:11:00 - 00:03:21:10
Adam
But like that was that's that's how it came about. But let's let's get into the details of it. So like where was this like give us, give us a timeline on this. Like what happened with PHP.

00:03:21:12 - 00:04:08:04
Martin
In under 72 hours? Which research used AI to find a decade old CE flaws in PHP, which we discussed a little bit, already coordinated a global patch with the NSA, EU, Cloudflare, AWS, Google, and Microsoft and watch governments take their websites websites offline as early attackers emerged and then simultaneously. So while all that's going on, the light gateway, the library routing traffic between AI apps and almost every major LLM provider was Trojan via a GitHub oath tokens that Claude code had been embedding in logging responses.

00:04:08:06 - 00:04:43:04
Martin
I know, and wait, we're not finished. AI agents were also reported inside US bank accounts, contacting call centers and taking actions independent of user instructions. And then wait, there's more. Because, you know, you talked about those software engineers and getting the fixes delivered. GitLab dropped a CVS, CVS ten floor covering 12 years of releases, and then we're not done yet because I'm going to probably pronounce this wrong, which is a trustee certificate.

00:04:43:04 - 00:04:46:00
Martin
Authority was found to have issued intermediate.

00:04:46:00 - 00:04:50:00
Adam
Trusted certificate authority was okay.

00:04:50:01 - 00:05:01:17
Martin
Yes. Was a trusted certificate to authorities. Was found to have issued certificates with private keys as weak as four. And can you explain this for a week? As for before I go on.

00:05:01:19 - 00:05:32:19
Adam
Yeah, yeah. First of all, mudra, the that company that you're referring to because of the online backlash of how weak their private key security was on was seven of their certificates or nine or something. They literally pulled out of the business of making of being a certificate authority. They stopped. Oh really? In a matter of their done, they stopped Microsoft, Google Chrome, Apple, everybody.

00:05:32:19 - 00:06:04:10
Adam
That was yeah. Apple also immediately rolled all the all all of those certificates. But that's just the certificates that they own and not everybody else's. Everybody else has to do it too. But anyway, the reason why it's important to understand that public keys were weak is or the private keys were weak is because they were human. Guess. So like so like it's the comic which of like, you know, a random number generator can also generate the number four.

00:06:04:12 - 00:06:05:16
Adam
Yeah.

00:06:05:18 - 00:06:06:08
Martin
Okay.

00:06:06:09 - 00:06:29:00
Adam
And so they should have been using an actual effective random number generator. But the one that they used generated the number four. And so any random person could have just gone through and guess. And it was only a 256 length or 200, 256 character length key that was being used. So it only would have taken 255 tries to figure it out.

00:06:29:01 - 00:06:31:18
Adam
Wow. But you only had to get to four.

00:06:31:20 - 00:06:35:07
Martin
Just the number four, which is genuinely insane.

00:06:35:12 - 00:06:54:19
Adam
And if you just guessed that, you could then impersonate and create new certificates and falsify certificates and impersonate Apple. And so a valid request from my Safari browser to Apple could be intercepted, decrypted, and then answered as if I was actually Apple. And it could be like a nefarious site, which would.

00:06:54:20 - 00:07:05:04
Martin
Be terrible for things like sending your payment details or, well, anything to do with personal information. Or, you know, it.

00:07:05:04 - 00:07:20:23
Adam
Was on Chrome like Chrome was using that certificate authority, that root certificate, which means that everybody that saves their password, their credit card, their, you know, address, all that information was completely in the clear to those attackers, which is crazy.

00:07:21:00 - 00:07:46:18
Martin
There was one more thing that happened. There was the number five in that list, which was Oracle also pushed a like a CVS 9.1 in Java's TLS implementation. So their secure communications communication. So overall, it has been a pretty horrific week because there's so many attack vectors all happening at once.

00:07:46:18 - 00:08:10:04
Adam
But I would argue that this this is this is week one of the new normal. Like we're going to look back on this week and think that it's cute because there's a lot of stuff happening right now. Sure. And it's like very like unprecedented. But things that are unprecedented don't stay that way. If they keep on happening. I'll bet you we've got another 18 months of this of just constant like, oh.

00:08:10:04 - 00:08:26:15
Martin
You don't think that that there'll be a concerted effort to remediate, like to trying to, like, get ahead of it or you just think there's just too much stuff out there to actually to beat it.

00:08:26:17 - 00:08:54:05
Adam
What has to happen. And I think what this week made clear is that every piece of code has to be inspected using AI to find vulnerabilities like this. And then as we find problems which we will find problems, we're going to have to go through and patch everything. And like literally all of it, this is this is what Project Glass Wing and Mythos told us was going to happen.

00:08:54:05 - 00:09:21:11
Adam
This is why anthropic released Mythos in March or whatever it was, they were warning us that soon their frontier level AI was going to be available to everyone, and this was going to be the result. And this is this is what we're living through. Like previously, the Frontier Labs with the bleeding edge AI capabilities, had guardrails on everything that they made about 5 or 6 weeks ago.

00:09:21:13 - 00:09:43:12
Adam
Open weight models, the ones that don't have guardrails caught up and that that was inevitable. That was going to happen. And so what we're seeing now is we're seeing guardrail less, mythos level, fable level artificial intelligence being used in the wild by people who don't like, like just there's no control on it. There's no there's no guardrails put around it.

00:09:43:13 - 00:09:57:12
Adam
Like now you can ask it to build you something nefarious. Now you can tell it like, hey, build me up, you know, inspect this code, find me a way to to hack this thing. That is what wasn't possible 5 or 6 weeks ago. That is now possible.

00:09:57:14 - 00:10:01:21
Martin
And so is now the time to actually try and have global regulation.

00:10:01:23 - 00:10:21:21
Adam
So this so the Von Pack lips. This is the thing that we've been expecting. The thing that we knew was coming like this is you know, the reason I joined harness two years ago is because I knew this day was going to come where open weight models without guardrails were going to be used. Nefarious, right. So so let's take a look at this and see what the implications are.

00:10:21:22 - 00:10:46:12
Adam
So there's there's a a cautionary tale in here about a particular company that really messed it up. Like really really messed it up. So Chain Guard is a hardened image company that makes basically like very secure, hardened images of Linux and different things that they then sell to their customers and their customers use as something that is safe.

00:10:46:14 - 00:11:18:21
Adam
They were brought into the circle of trust about PHP and they were told there's an embargo. Nobody talks about this, nobody says anything about it until this particular date and time, which was supposed to be Wednesday morning, and nobody has permission to release anything until then. Chain guards CEO unilaterally decided without conversing with anybody else, without, you know, checking with a PHP foundation that actually maintains PHP, decided that they had enough information to go ahead and release their hardened image with a fix to that PHP issue.

00:11:19:01 - 00:11:58:14
Adam
What does that mean? Two years ago? Nothing like two years ago. Who cares? Because it's going to take somebody 3 or 4 months to reverse engineer that fix and then get to an attack path, or get to a kill chain using the information that they gleaned from that patch. But now that these open weight models are in the wild, it was 90 minutes between the time that Chain Guard released their patch and they started seeing indicators of compromise up against other indicators of compromise in the wild, because someone had reverse engineered that patch and then was exploiting PHP in 90 minutes.

00:11:58:16 - 00:12:33:20
Martin
That's when we say 70% of the world's websites use PHP is because things it powers like a lot of the website content management systems like Drupal, Joomla, WordPress, like, you know, even wiki Wikipedia is a PHP site, you know, let alone the fact that it powers governments like in in Europe. Interestingly, the EU themselves didn't shut down any websites, but they issued really quickly guidance saying, hey, if you're using PHP, you should really deal with those sites quickly.

00:12:33:20 - 00:12:39:08
Martin
And then, you know, countries across not just across Europe, across the globe, didn't they publish that.

00:12:39:09 - 00:12:41:10
Adam
On a website that was using PHP?

00:12:41:11 - 00:12:46:08
Martin
And they published that on their site that used BHP? They did.

00:12:46:10 - 00:12:47:14
Adam
They published it.

00:12:47:19 - 00:12:53:09
Martin
I think Chain God is going to face a backlash probably over that.

00:12:53:10 - 00:12:55:03
Adam
Already are already are.

00:12:55:04 - 00:13:13:07
Martin
But I feel like something like that should have a hard regulated. Hey, if you're in this loop of people who are going to fix it, you know, and we've all agreed to do it together so that we can minimize the impact, I think there should be. It needs to go beyond just like a almost like a gentlemen's agreement.

00:13:13:09 - 00:13:34:05
Adam
Yeah. I want to disagree with you so hard on that, but I can't I think that did you just. Like that is actually like the one kind of regulation that I think would actually work in this, in this scenario like that, that in particular of like, hey, if you're part of this embargo, you are legally bound to not break confidence and not release it.

00:13:34:06 - 00:13:48:10
Adam
I think it's moot now. I mean, the problem with Chain Guard is like like even on. So I'm part of like a bunch of different CSO slack channels and stuff like that. There's a couple different ones, some that are like Fight Club that I can't talk about, but somehow people keep joining and other ones that we can't talk about.

00:13:48:10 - 00:14:09:00
Adam
But like that was the big reaction from from the CISOs was like, who the who the fuck do you think you are releasing early? Like like there's. I cannot conceive of a reason that they would do that. Besides, they were like, oh shit, if we don't, if we're not the first ones to release this, then we're going to lose a marketing opportunity.

00:14:09:01 - 00:14:21:01
Adam
And that's like, I can't think of any other reason. All right, well, let's talk about regulation. Martin, what the hell? Why? This is cool fun, like scary stuff. And you want to talk regulation? What the I.

00:14:21:03 - 00:14:21:14
Martin
Because.

00:14:21:15 - 00:14:22:05
Adam
With all this.

00:14:22:08 - 00:14:56:09
Martin
I want I want a certain amount of regulation. It's the the bank one that really scares me. So I think it was 5 or 5 plus banks, including like all the big ones Morgan Stanley, JP Morgan Bank from Erica Wells, Fargo City, all the and one of the exists. And this blew my mind one of the exact told journalists we struggle to differentiate between agents or whether an agent was authorized by a user or has gone rogue.

00:14:56:10 - 00:15:12:00
Adam
How would you though, right? Like like my mom. I'm convinced that my mom has never written her password down, and every time she needs to get into her bank account, she changes the password. Something like her. Her flow of getting into the bank, into her.

00:15:12:04 - 00:15:13:08
Martin
Reset password.

00:15:13:11 - 00:15:14:23
Adam
Reset password. And then.

00:15:15:00 - 00:15:18:00
Martin
She's doing the magic link just manually. Like.

00:15:18:02 - 00:15:25:20
Adam
Exactly like like what's the difference between my aged mom being the worst computer user in the world and an AI doing it?

00:15:25:20 - 00:15:55:20
Martin
Okay, so this is where like regulating the agents have to have a unique identity, even if they're for one of a better phrase offering operating under under the authority of Adam's mom. Right. They still need to have a unique identity. And and I think mandating that making that a hey, agents can't talk to our systems if they don't have those things.

00:15:55:22 - 00:16:05:05
Adam
Are you telling me that you walk into a Taco Bell and check the idea of the person that hands you a taco with the cash register? You don't know who that person is just because they have a Taco Bell T-shirt on.

00:16:05:07 - 00:16:20:15
Martin
But it's the other way around. It's not the it's not the agents on the inside. It's the agents on the outside coming, coming in like. And, you know, they have a nice badge on that, says Adam and has like three stars on it or something.

00:16:20:17 - 00:16:33:13
Adam
But that's my point is like if they look and smell and walk like a duck and they happen to quack, like, what is the actual difference if they're if they're acting on behalf of people, what does it matter whether or not they're identified as such?

00:16:33:18 - 00:16:50:23
Martin
But I think that's that's the problem, right. Because they could do that. But I think knowing that it's an agent and having a mechanism to verify it's acting on behalf of whoever set it off correctly, or whether it's just something that's gone rogue, you know, is that thing.

00:16:51:00 - 00:17:00:07
Adam
And but it would have to be it would have to be identifying the agent and somehow watermarking the actions that it's taking. And so like I. Actually I.

00:17:00:12 - 00:17:00:17
Martin
Think.

00:17:00:19 - 00:17:05:18
Adam
I think that will be the regulation that comes down the pike. I just don't think it's going to be worth anything.

00:17:05:20 - 00:17:33:11
Martin
Okay. I feel like that is a place to stop. We let's let's agree to disagree. So I know you're part of several security groups that talk about what's going on, what the risks are like is the shift that you anticipated, you know, from mythos and everything else that's been going on. Is that shift really happening? And what is the pivot like?

00:17:33:11 - 00:17:37:20
Martin
How is the security community as a whole changing their view?

00:17:37:20 - 00:17:44:00
Adam
It reminds me of this is going to sound alarmist. So I'm going to say it two different ways because this reminds me.

00:17:44:00 - 00:17:46:09
Martin
Of I like the alarmist sounding one, so please do it.

00:17:46:10 - 00:18:22:09
Adam
Yeah, literally. Like seriously, this this reminds me a lot of post 911 America where like right after nine over 11 happened, there was a galvanized of like the public and everybody was like, in this together. Like, we're going to get through this, this, this whole thing. Like somebody that's what I feel like is happening right now in these, in these security circles, is we're starting to pull together as a community and like, galvanizing our resolve to actually get through this in a, in a principled way, which is why everybody was so upset with Chain Guard and also why we're there's information sharing at an unprecedented level.

00:18:22:10 - 00:18:42:05
Adam
What I'm not seeing I'm not seeing CISOs in arguments with their engineering teams about whether or not they stop shipping features. That's not a conversation. Nobody's like nobody's like, hey, like like nobody woke up Thursday morning and had an argument with their engineering lead that was like, hey, I really need you to patch this. And engineering the lead was like, well, I've got KPIs and I need to ship.

00:18:42:05 - 00:18:55:00
Adam
This feature did not happen. Everybody was on the same on the same page of like, oh shit, yeah, I know this is a big deal. Like, yes, actually yes, we will do this. I hope that that leads to long term change.

00:18:55:02 - 00:19:30:14
Martin
I've been in organizations that have been hacked. I've seen the impact, I've seen that galvanize happen. And then I've seen people were pivoting back to their own priorities over that galvanized view. So I want to be hopeful. I can't believe I'm being the pessimist. And you're the optimist because I feel like it's always the other way around. But here I, I want to be optimistic, but you've got to convince me because I am I, you know, and the chain God thing kind of makes me feel that pessimism is right there.

00:19:30:16 - 00:19:54:00
Adam
Yeah. And here's, here's, here's my evidence. Here's why. Okay. Because this pain is not temporary. This is not a one time thing. Like I said, this is the first of many weeks of the next 18 months that this is going to be a constant pressure. So, yes, you know, the morning after air raids over London, people still went out and bought bread.

00:19:54:01 - 00:20:14:06
Adam
Yeah, I get that. But it happened constantly and people lived a different way. Like, yes, the normal things had to keep going and yes, features will still be released, but the pressure of this is not going to let up because, I mean, we're talking like six things that happen. Seven things had happened in one week. And we're like, minds are blown.

00:20:14:07 - 00:20:36:00
Adam
Like people like that pressure won't let up. So I don't think it's a matter of willpower. I think it's a matter of constant exposure to the danger, is going to change the way that engineering functions going forward, which, again, is why I think it's an engineering problem, not a security problem. Security is helping. Securities got the flashlight. Security is leading people out of the woods.

00:20:36:02 - 00:20:40:14
Adam
Still got to be engineering. That makes the change the way they do. They do their shit.

00:20:40:16 - 00:20:48:14
Martin
I mean, is this an argument for organizational restructuring around the way engineering security interface?

00:20:48:16 - 00:21:13:13
Adam
Hell yeah. Oh yeah. It has to be like like the days of security being off on the fringe and being, you know, the person that comes in. Here's my spreadsheet of vulnerabilities. Please fix these. Gone like that's not that's not the case anymore. This is this is the day that not. I'm I'm being exaggerated. But like, this will be the time period in which engineering and security come back together as they should.

00:21:13:14 - 00:21:23:21
Martin
You know, like, I have that platform engineering view that you should be building secure by design platforms to deploy to because it makes it it doesn't take the risk away.

00:21:24:01 - 00:21:25:08
Adam
This this is what security.

00:21:25:08 - 00:21:44:03
Adam
People have known for years. And we've struggled to convince the rest of the technology teams of this is like, this is not a question of if. This is a question of when, it's not, if you're going to be attacked and if you're going to be owned by a new vulnerability. It is when like, swear to you. There was a company that was hiring for a security leader.

00:21:44:03 - 00:22:05:08
Adam
And like I interviewed with them and they were like, well, have you ever been has a network under your control ever been hacked? And they wanted the answer to be no. Like what? No. That's like that's that was the mindset was like, we want somebody who's never been hacked before. You don't there's no such thing. There's only people that don't know about it.

00:22:05:09 - 00:22:24:13
Adam
Whole, whole different story. And that's like, that's the mental process that needs to happen here is like, it will never be safe again. You have to do everything you can to stay ahead of it, but you'll never be ahead of it. Just keep pushing because that's the only way to get around it. All right. That, my friends, is ship talk.

00:22:24:13 - 00:22:37:15
Adam
Brought to you by harness. If Martin's point of view about regulation got into your skin, drop a comment, like and subscribe. But until next time, stop talking and start shipping.