00:00:00:04 - 00:00:05:22
Martin
Adam, did you see that story that Claude was used to attack nine Korean banks?
00:00:05:23 - 00:00:15:10
Adam
Is this the beginning of a trend? Okay, we have an update. We talked a little bit about Citrix being hacked a week and a half later. They have finally released the patch.
00:00:15:11 - 00:00:20:03
Martin
There is one thing about this because the patch isn't 100% perfect.
00:00:20:04 - 00:00:28:10
Adam
Harness recently announced that it was acquiring augment. I've been hearing a lot of talk about software factories being faster, but right now speed is not the issue.
00:00:28:11 - 00:00:36:05
Martin
OpenAI's end of year financials came out, and they were $20 billion less than they expected.
00:00:36:06 - 00:00:41:11
Adam
Going to like. That means that this might be a bubble.
00:00:41:13 - 00:00:49:00
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, Martin Reynolds.
00:00:49:01 - 00:00:50:19
Adam
And I am Adam Arellano.
00:00:50:20 - 00:00:59:15
Martin
Adam, did you see that story where Claude was used to attack nine Korean banks? Nine banks. Did you hear about that?
00:00:59:17 - 00:01:05:12
Adam
No, wait. Nine Korean banks were attacked by with Claude. Not by Claude, but for.
00:01:05:13 - 00:01:27:21
Martin
Well, yeah. With Claude, sort of. The banks were hacked by a 26 year old man from China. And the way he did it is he used this open source pen testing agent called Rex, and he paired that with Claude Code. And then he used it to essentially go in and say, hey, where can I sell stolen Korean data?
00:01:27:21 - 00:01:44:26
Martin
And then find me some telegram data channels, where data sales channels where I can actually sell it and then write me a fake security researcher resume me including his actual personal details, which is how people actually tracked back to the suspect.
00:01:44:26 - 00:01:47:00
Adam
So he used his own name.
00:01:47:00 - 00:01:57:14
Adam
On the resume. And the resume was written by Claude. Was it his use of Claude that actually got him, or would they, like, get a subpoena and go back to the to the source to see if he had?
00:01:57:19 - 00:02:18:01
Martin
No, it was the it was the resume that actually gave it away. So there were personal details in the resume and they used that to track back to him explicitly. I thought this was really interesting because even the method of attacking like a human could have done it, it would have taken a lot longer, but they are essentially using the agent to plan the attack.
00:02:18:06 - 00:02:33:21
Martin
You know, it was like going to all these banks, knocking on one door and saying, no, I'm going to try a different door now. I'm going to try a different door. Okay, I found a routine. And then even like planning all the tools, how they were going to do the tack, making sure that the attack software was going to work.
00:02:33:22 - 00:02:44:06
Martin
It was like all planned and all done using Claude, but a person could have done it, just would have taken a lot, lot longer to do the same amount of work.
00:02:44:07 - 00:03:09:06
Adam
And then how would they have known where to find those channels to actually sell the data? Like, what's funny is that in this, in this scenario, the the AI is being used to enhance the abilities of an obviously not sharp person, a person that was just willing to break the law, but that person was so not smart that their use of AI is also what got them like, popped.
00:03:09:07 - 00:03:29:11
Adam
Okay, so here's my question. First of all, Claude code was being used in this whole thing. Is this the beginning of a trend? Do you think that this is going to be like the low level, silly shit where it's like, basically script kiddies, but with AI enabled script kiddies that are going to be using Claude as their like, step by step guide of how to do wrong things.
00:03:29:13 - 00:04:01:04
Martin
Sorry. Wasn't Claude code alone, right? It was. It was the testing agent with Claude paired with Claude code. You know, essentially an open source penetration testing tool paired with Claude code makes it a very low cost attack. Do I think we're going to see more of it? For sure. Yeah, not necessarily with Claude. I think it could be any of the, you know, any of the LMS either the open way or the closed wait, you know, the open source or the or the closed models.
00:04:01:06 - 00:04:12:10
Martin
You know, when you've got that kind of agent and, and a fairly powerful model, you compare them together. I'm pretty sure it's already happening. It's just this one caught the headlines.
00:04:12:15 - 00:04:25:12
Adam
Yeah. And partly because this guy was like a clown. Like he I think one time he didn't turn his VPN on when he was routing command and control traffic, which gave them like his home IP address, which is like it's pretty, pretty basic.
00:04:25:13 - 00:04:26:22
Martin
That is actually insane.
00:04:26:26 - 00:04:50:24
Adam
I okay, we have an update from our last show. We talked a little bit about Citrix being hacked. A week and a half later, they have finally released the patch that's going to help their customers be safe. Finally, been a week and a half. So Citrix actually published the bulletin for the CV. I'm not going to go through the number, which is a memory overflow for in.
00:04:50:26 - 00:05:09:16
Adam
Net scalar ADC and net scalar gateway lead to remote code execution or denial of service. Has a CVS score of 9.5. Now, we talked about this last time, and we left it off at a place where they actually hadn't even released a good patch for it. But now, a week and a half later, they released a patch.
00:05:09:18 - 00:05:37:23
Martin
There is one thing about this because the patch isn't 100% perfect, I think it was. Mandiant confirmed that the patch doesn't even remove post exploitation access. So if the attackers got in before the fix, which means every org that patched on time in September, there's still a question as to whether they were hit first. And, you know, how do you even respond to that?
00:05:37:24 - 00:06:11:03
Adam
Yeah, because if you apply the patch, it doesn't kick out the attackers that got in so that there probably are still lingering living off the land inside your system. Yeah. And it's really highlights the fact that even if your own company. So let's say if we had these devices in our in our environment, even if we were super quick to patch and we had our like SLAs down to hours or maybe even minutes, like we could roll something out that quickly, an environment that's cool unless there's nothing to patch, then we just know about it and we have to either decide to take it offline.
00:06:11:04 - 00:06:37:05
Adam
Like that's what everybody did, actually. Yeah, either take it offline or just live with it in the environment. And then when the patch does come, it's like we're week and a half late. This is the this is the shrinking of time to attack from CV from an you know, when they announce that there's actually a flaw, even if it's a security researcher, the time before that gets exploited in the wild is zero or less.
00:06:37:05 - 00:06:38:09
Adam
Now that's scary.
00:06:38:10 - 00:07:05:00
Martin
Yeah. And I mean, I honestly, I think it's a fair expectation that we should expect these companies to respond faster, right? If you know they need to evolve in the same way everybody else is evolving to a different kind of world, right? You would hope that they are using AI assisted development and testing practices that would help them get to a resolution of this faster hell.
00:07:05:01 - 00:07:14:01
Martin
Why are they not using AI to constantly try and hack their own stuff inside a lab so that they can find these things faster and patch them and fix them and release them?
00:07:14:03 - 00:07:44:07
Adam
Yeah, here's what could be finally reversing. Up until now, companies who get publicly hacked so companies admit publicly that they get hacked. The effect on their on their share price has always been up. Every single company has been hacked. Their share price is now higher than it was when they got hacked. That used to be the case. I think what we're going to see is that companies who can't keep up with these attacks, like Citrix, they're going to see a huge hit to their actual renewal rates.
00:07:44:07 - 00:07:55:03
Adam
I know in the in the CSO world, there are already these conversations happening of like, how and how quickly can we get out of Citrix and how can we get to one of their competitors that can respond more quickly?
00:07:55:04 - 00:08:18:20
Martin
I mean, based on that prediction, like genuinely, the evolution of how you deliver software in your organization just has to evolve now faster than ever. I know we don't often talk about what we do in here very often, but I sat, you know, in front of some of our customers who are targeting, you know, maximum 12 hour remediation as soon as they get a signal.
00:08:18:21 - 00:08:30:03
Martin
And I think that evolution is going to happen in broader and broader and broader circles, because everybody has something that they need to protect and there's no perfect wall.
00:08:30:04 - 00:08:45:20
Adam
I really thought you were going to go a different direction. I thought you were going to talk about ratcheting up security protections. And what we're talking about is not security protections. What you're talking about is engineering. Engineering needs to be faster. Software engineers, shipping code that needs to be faster, more than you need to be locked down.
00:08:45:21 - 00:09:02:22
Martin
Right. But you need to be able to respond at least close to the speed that you're being attacked. So based on all this, you know, if you were put in charge at, let's say, Citrix tomorrow on how you were going to build those muscles, what is it that you would actually do?
00:09:02:23 - 00:09:20:21
Adam
I mean, there's a school of thought where you would bring in a bunch of red teamers to constantly bang, await your stuff. That's probably fine, but I think that's the second thing I would do. I think the first thing I would do was start figuring out where my bottlenecks are. Where do I need to improve speed to develop better code?
00:09:20:24 - 00:09:45:23
Adam
And while that's happening, have some red teamers go at it. Because if you if you reverse those two actions and you hire a bunch of red teamers first, you're just going to find a bunch of problems and then not be able to move fast enough to actually patch them. So it's actually counterproductive. So to me, you really, really have to fix the fundamentals of engineering to be faster before it really even makes sense to have more input of like, oh, all these are problems.
00:09:45:24 - 00:10:07:05
Adam
Got it, I got problems, but I can't do anything about them because I'm slow. So in other news, harness who we work for and often where t shirts of recently announced that it was acquiring augment. Augment is the maker of a tool called cosmos, which is the basis for software factories. Martin, what do you know about software factories?
00:10:07:12 - 00:10:48:07
Martin
They make software. You'd be amazed. It's filling that gap between I'm a product manager and I have a requirement, and I. I'm a software engineer, a developer. And this is what's required technically. And pulling those two things together so that when the AI is actually generating the software, it's doing it with a lot more context. I was extra excited because of the augment purchase, just because when you link that with everything that we do in terms of the delivery, how it's deployed, how it runs, how it's operate, what the security posture is, one of the regulations and guidelines that you need to operate software, and you pull both of those things together.
00:10:48:08 - 00:11:00:00
Martin
What you should get is much, much better software generated out of out of out of that so that you actually get, you know, something that understands end to end.
00:11:00:07 - 00:11:26:09
Adam
I like that you use the word better or not faster, because I've been hearing a lot of talk about software factories being faster, which sure like it can. But right now speed is not the issue. Generating code at speed is the easy part. Generating good code and speed is the hard part. And I think that'll ultimately be what we're going to be bringing to the market is a full end to end solution that can help you build quality software, not necessarily quantity software.
00:11:26:14 - 00:11:49:14
Martin
I do have something I genuinely wanted to bring up, and only because you have heard me mention this so many times. But you know, I think last night as we're recording this open AI's, you know, end of year financials came out and they were 20 billion less than they expected.
00:11:49:17 - 00:11:52:26
Adam
That seems like a problem in a quarter or in a year.
00:11:53:00 - 00:12:28:13
Martin
Let me give you the details. So essentially there's they made a net loss of 39,000,000,000 in 2025, nearly eight times worse than the nine sorry than the 5,000,000,000 in 2024. Revenue reached 13.1 billion, which is up 253% year on year. To be fair, with monthly revenue approaching 2 billion by year end. However, that number is kind of tainted because they're operating.
00:12:28:13 - 00:12:41:25
Martin
Loss in 2025 was like 20.9 billion. That excludes any restructuring and other stuff they did. So essentially they're still losing $1.60 for every dollar earned.
00:12:41:26 - 00:12:48:25
Adam
But I think they were predicting they were going to lose money. It's just that they predicted like way off by to the tune of 20 billion.
00:12:49:01 - 00:12:55:15
Martin
The market was expecting that it would be $20 billion better than it was.
00:12:55:17 - 00:13:11:02
Adam
So the market was like, they're going to lose 20 billion thumbs up. Let's give them more money. Oh, it's twice that. Yeah. Doesn't that mean they should get more money? That's been the pattern so far. Is that when they overreach on their spend they get more in fundraising? Yeah. That's crazy.
00:13:11:02 - 00:13:43:08
Martin
But I think the interesting thing that I noticed is that they're saying, you know, the debts that these AI companies are holding. They're the people who are holding that data, starting to sell it like some of the three of the big banks have already started to sell off some of that debt that they own. And interestingly, the people who buying it are buying it at a higher interest rate because they're now a little bit more nervous about the risk of whether all this, you know, accelerated innovation is really going to pay off.
00:13:43:10 - 00:14:02:09
Adam
Interesting. So they're they're demanding a better interest rate if they're going to buy the debt from the banks, and the banks are having to take it. Yeah, I kind of feel good because the banks are getting kind of hosed, because that makes me feel good, because I just feel like banks getting host is great. But I also am like, oh no, like that means that this might be a bubble.
00:14:02:11 - 00:14:23:23
Adam
So the CEO of a company I know was on CNBC last week, and one of the big takeaways I had from it was that there is a bubble in the compute. Whether or not there's an AI bubble is a whole different thing, but in the compute that's being built and being forecast for being built. So just data centers and storage and processing, that is probably way too high.
00:14:23:23 - 00:14:33:24
Adam
And they've actually over predicted what they need. As models get more efficient and people get more efficient in the way they use them. So that might be this might be an indicator of of that statement.
00:14:33:24 - 00:15:05:21
Martin
So OpenAI dropped like 722 math papers produced by a model that they haven't actually released yet, which is amazing. And then so the the timeline goes October 6th, they published these 722 research manuscripts on GitHub, covering hundreds of long standing unsolved problems across 17 areas of mathematics. They fed the model roughly 4000 problems, so, you know, less than 25% success rate.
00:15:05:21 - 00:15:35:23
Martin
Anyway, filtered for significance. And this is what came out. And the average compute result was about three hours of ChatGPT Pro thinking mode. So three hours per solution, which is insane. At this point, none of its peer reviewed. OpenAI admitted that some of its results might contain errors, and that some proofs their own researchers don't even understand.
00:15:35:25 - 00:15:37:02
Adam
Wow.
00:15:37:04 - 00:15:54:09
Martin
The reason I'm bringing it up, though, is because it kind of sides on that side that says, hey, all this innovation is worth it. All this investment is worth it because we are going to make some of these advances. Now, it might take a little bit more time to really get the value out of those. All those new things have been solved.
00:15:54:15 - 00:16:18:05
Martin
Is it moving our knowledge forwards? Absolutely. You know, and I would rather that they were pushing this out there and said, hey, yes, we've solved this. All these things. We maybe should have audited what was going on a little bit better. We'll do better next time. But the fact that it's doing that and it's moving us on, you know, from a knowledge perspective and the fact that they're sharing it, that that feels good to me.
00:16:18:05 - 00:16:24:09
Martin
And it feels like some of that investment is really showing, you know, some some big promise areas.
00:16:24:11 - 00:16:33:05
Adam
Google and Beth Israel Deaconess Medical Center published a Lancet study of Amy. Amy stands for Articulate Medical Intelligence Explorer.
00:16:33:06 - 00:16:33:25
Martin
That's a mouthful.
00:16:33:26 - 00:17:00:03
Adam
Google's research diagnostic AI chatbot Amy, in a real world primary care clinic with 98 patients, patients talked with Amy before urgent care visits, so its summaries could inform clinicians with physicians supervising in real time. No conversation had to be stopped for safety reasons. So like, no point did the bot try to kill anybody. Awesome. And Amy's differential diagnosis matched the doctors final diagnosis 90% of the time.
00:17:00:03 - 00:17:04:19
Adam
So go ahead. What's the first reaction? What did you have to this.
00:17:04:19 - 00:17:12:07
Martin
So my first reaction is to me this is an evolution of a service I've already seen elsewhere in the world.
00:17:12:07 - 00:17:14:19
Adam
So WebMD okay.
00:17:14:20 - 00:17:43:00
Martin
As you know, in the UK we have we don't have nine on one. We have nine, nine, nine. Right. But we also have one, one, one for medical emergencies and the way for non-emergency medical help and the way that system works is I would say essentially it's a sophisticated Q&A system where a like an automated system will ask you, you know, where are you feeling pain?
00:17:43:00 - 00:17:58:13
Martin
And then based on the response and then the response and then the response. And then it comes up with a diagnostic and makes an assessment of whether they should speak to a nurse or a doctor. And they that nurse or doctor then has all that information when they start speaking to them, so they don't have to repeat it to them.
00:17:58:16 - 00:18:26:23
Adam
Yeah. So this is like an evolution of that where actually it's the exact same thing, just that it has a more sophisticated basis for which to make recommendations. Yeah. Well the headline I think is also fascinating. The headline is that AI could improve patient physician relationships, which means like part of the part of the conflict and part of like the discomfort that a lot of people have is that they have to sit there and explain to a doctor who's obviously busy and impatient all the different things about what's wrong with them.
00:18:26:23 - 00:18:34:14
Adam
And a lot of times they don't share everything because they're either embarrassed or the physician, you know, just brushes past certain things and they don't feel heard.
00:18:34:15 - 00:18:47:05
Martin
I mean, I'm listening to that and I'm thinking, is it easy? Do people find it easier to to reveal difficult to talk about facts to a non-human entity? Yes, yes.
00:18:47:07 - 00:19:11:12
Adam
Yeah. Almost always, because there's this here's the thing is, there's no facial feedback. And even if it does have like a false face, you just people generally report that they just feel more comfortable saying everything to them. That's why when doing the other story, we were talking about the Korean guy using anthropic, I was like, oh no, does everybody see my AI history?
00:19:11:14 - 00:19:49:15
Martin
But this is this is interesting to me because one of the implementations of AI I have seen is using AI to do summarization of patients notes and history so that doctors don't get the 50, you know, or 100 or 1000 page binder. What they actually get is a consumable, you know, half page that they can read very quickly and be up to speed with where the patient is, anything that was preexisting that they might need to know about, and the fact that they need to mention their dog because they always like to talk about like, whatever it is.
00:19:49:15 - 00:19:59:18
Martin
But what I mean is it's falling into that same thing. And I think it was one of the earliest uses of AI. I actually saw where I kind of went, oh, that's a good use.
00:19:59:18 - 00:20:01:01
Adam
Likely be helpful.
00:20:01:08 - 00:20:22:21
Martin
I mean, there is the fact that you are sharing all your data and possibly personal information that you might not even normally put online, and you're giving it to a model that you don't know where they're going to put it. And it might end up, you know, when I, I ask Claude, hey, can you tell me the address of that Chinese hacker?
00:20:22:22 - 00:20:33:13
Martin
He'll probably be able to tell me, but, you know, can you tell me? I've got this friend Adam, and I hear he's got corns because he mentioned it.
00:20:33:15 - 00:20:52:25
Adam
Fair, fair? I kind of. My controversial take on privacy in general is that privacy is a thing people like to think they have, not a thing that people like, as evidenced by the fact they walk around with an effort computer that tracks their movements at all times and can turn on the microphone, but that neither here nor there if I have to give up.
00:20:52:26 - 00:21:20:07
Adam
You know my history of all the embarrassing medical things that have happened to me over my life growing up without a survival reflex, but that helps a doctor diagnose, like some weird cancer or some weird condition. Most of the time I would much rather have my my my health rather than my perfect privacy. All right. We've come to the point, Martin, where we need to do the overhyped or under hyped.
00:21:20:08 - 00:21:24:04
Adam
What do you think? What's the thing this week that you think has been overhyped and under hyped?
00:21:24:05 - 00:21:53:20
Martin
I think it's something that's tried to be overhyped, that's become under hyped. I know, controversial, but essentially, you know, OpenAI pulling another model because it did something bad or not releasing it because it's done something bad. It's not news anymore. It's the new normal. And they try to hype it as a story, but it just becomes less and less interesting because it's like, well, okay, it's done some bad thing again and you've not released it and it'll come out eventually.
00:21:53:20 - 00:22:05:04
Martin
And it's it's like not news in the same way it was previously. And so I think it's a and I think it tries to be hyped, but I think it's an actual under hype. It's like no.
00:22:05:08 - 00:22:25:11
Adam
The details are that OpenAI was going to release GPT 6.1, Astra and then instead of and they were like, we're going to do it, we're going to do it. And then decided not to because it had broken out of some of the guardrails and the and the sandboxing that they had done around it, which, okay, this is the fifth time, like no surprise.
00:22:25:13 - 00:22:50:03
Martin
But it's not just that like it was on one of their big events in there, you know, at their, their Dev Day's events that they were running. So it's not just that they pulled it, right. But they, they, they were having a big event where they hope and hoped to launch it. And they still pulled it, but it was still like it was still non-news even despite that.
00:22:50:06 - 00:23:10:07
Adam
Yeah. Is this the new Steve Jobs? Oh, and one more thing where he brings out some other like instead of, oh, one more thing. It's like we're about to release something and everybody's like, oh yeah, when's the beat going to drop? And they're like, we're not because it's too dangerous. And that's that's the new thing they do at every developer conference is claim they're going to release something and then don't.
00:23:10:08 - 00:23:21:00
Adam
That's ship talk brought to you by harness. If Martin's point of view got under your skin. Subscribe and tell us why. Until next time, stop talking. Start shipping.