So I'm going to start off, what's the one thing you wish people would take away from your talk? That you have to have dedicated people to deal with that. It's not something you can do as in, you know, after the fact or offhanded. You need somebody who's taking care of your agentic infrastructure, who is reading about the latest attacks, the latest tools, the latest protection. Who manages the migration of your company from whatever it is right now into somewhere on the scale of AI first, right? It can be for certain processes or everything, depends on you.
But it's not something you can visit once a quarter in some status meeting. This is something that requires constant attention. And when I say constant, that's almost on a daily basis. I think we're moving, I'd say we're moving away from quarterly account reviews, permission reviews, what have you. And I think it's going to be interesting because I think a lot of IGA departments that have kind of invested in the good old bad old way are going to find themselves challenged. If anything, what we're seeing in the last about a month or so is the understanding that attackers are using AI as well.
So for example, the time it now takes to create an exploit chain or to find vulnerabilities. And okay, it's still an open source, at least that's what we're being told, but that happens super fast. So it's not going to take, I mean, over the years, we've learned that using all the security mitigations that the companies did from ASLR to other things, it raised the price and extended the time required to reach an exploitable vulnerability. We're back to the starting line. Now things happen this fast. Attackers use AI to develop tools, to scan your organization, to attack it.
I mentioned that Anthropic posted a whole analysis of an attack group using cloud to attack. It's super fast. I've seen the thing that Gadi Evron has produced, Raptor and so on. How do you think it's like? Well... Even that in a simple proof of concept form. So Gadi is a dear friend and he's always been a very community minded person. So he's doing a lot of things for the community. And it's good because I think that a large part of our problem is the awareness.
Because yes, everybody in this room and probably this conference is in some way, shape or form a security professional, but not everybody is as aware to how serious the things are and how easy it is for the other side to do things. And releasing these tools kind of puts it in your face and you can't ignore it anymore. This is the brave new world in which we live. And going back to something you said about, you know, incidents have changed. We all know the model. We're all calm and happy. And then something happens and we wheel into place our perfectly planned, rehearsed... Why are you laughing?
Our perfectly planned and rehearsed incident response plan. And we restore normality and it's all calm and happy again. We drink a cup of tea or coffee or whatever we like. That's no longer the world we live in. Now it's incident, incident, incident, incident. Faster than our rate of response and recovery. Not only that, but as I mentioned in my talk, supply chain attacks have peaked. So now the incidents are not anymore in your system.
You wake up one day and you realize that the NPM library you've been using or the tool you've been using has been infected anything from yesterday to two years ago. People find 18 years old vulnerabilities in the Linux kernel. And it's like somebody else's code is being exploited and you did everything right, but you were using their code. So it shifts the balance back into supervision and visibility. And we always said that prevention is never going to happen. It's about detection.
So you got to have all the infrastructure, including in the age of AI, where a lot of the conversation takes place in areas that are not easily supervised. What the conversations are, what the tools are being used for, what is the outcome? I can say turning to AI, you made me laugh earlier when we were planning this talk.
You said, you know, the plans we had three weeks ago, the world's changed. Yeah, I mentioned it as I started talking. The new world. What things have you noticed that have changed much? And without repeating yourself, obviously. So obviously coding agents. That was not a thing when we first discussed having this session. And maybe it's just a matter of focus.
But again, supply chain or infected modules. That happened a lot in the last two weeks. And we've seen a number of published attacks. These are just the published ones, right? You don't always know about everything. So in a very short period of time, things that used to either extend over a longer time or just take longer to happen. Everything is just like happening. It's every day you wake up and there's some other bad news or every other day. And it creates on us what we used to say about the socks, alert fatigue.
Like, okay, what am I going to do about this? I'm surrounded. What can I do? And of course, the answer you're going to get is use AI to process all the volume of alerts. And that's going to help for a while. But it's still a lot. We're just surrounded. I think the danger with a just use AI is we're putting another level of abstraction between us and the presenting problems. And we are trusting that the AI will do the right thing. But everyone has heard the word non-deterministic more times than they want to just in this conference alone.
But there is no guarantee that AI will do the right thing. We have an expectation, but no guarantee.
So, first of all, I don't think I know that we've reached a point in time where the job description of a developer changes from an actual developer to sort of a team leader of agents. Because people no longer read the code or even write it. If it's working, if it passes the tests, if at least on the surface of things it does what it's expected to do, you don't really know what's inside. And that's the trust thing that you mentioned, right? But also you mentioned the non-determinism.
And I mentioned in my talk, because LLMs are non-deterministic, it's very hard to measure their performance or score them or make sure that they do things the way you like or permit. So, the only way to do that is impose a deterministic framework or even just a reporting mechanism that you can inspect and then make sure that the LLM behaves as it does. Because even if everything looks fine now, then Sonnet is going to come out next version in three weeks. And they change the model and now things don't work the same. Maybe your code looks different. Maybe the tests don't even work anymore.
There's another risk actually. I mean, this is looking at it from a different angle. You mentioned, yes, Sonnet comes out in a different model. But also we're seeing the rules under which we can use these models arbitrarily changed.
So, for example, Anthropic said, we no longer want you to use third-party harnesses. But that was a financial decision because OpenClaw really hit them hard. Sorry? OpenClaw really hit them hard because I skipped that part because I ran out of time. But I was going to show you guys what's the best way to install OpenClaw without destroying your business or your own computer. And when I wrote that, I was using the $100 a month plan at home, right? This is personal. This is not a work thing. But then Anthropic woke up and said, whoa, everybody's using all these tokens.
This is killing our computer. And of course, they would sell it as we want to give you a better product for the paying people. So we're not going to allow that.
But yeah, all of a sudden, something that you had working for you, you wake up in the morning and it's no longer relevant. And it's all changing super fast. Yeah. And as you say, it's a financial, not a technical risk.
But to me, it's something to consider. Your supply chain that you perhaps didn't even think about has suddenly the ground and moved under your feet. And as I say, if you've bet some important parts of your business on I can use OpenClaw with Anthropic, then you are in trouble. I would say even something more difficult and problematic than that. This is something for you guys to consider if you are using LLMs in your business critical processes. Just like AWS, you can wake up one morning and cloud is going to be down. And for some models, you can create a DRP, a disaster recovery plan.
You can use local models and maybe spend $20,000 on a very big Mac mini. And you can compensate for some of the things. But if there is any downtime, for example, if a missile hits the Amazon data center in Dubai and your entire business process is based on that, there's nothing you can do unless you somehow created this disaster recovery plan. But we're used to thinking about that in terms of AWS because we've seen it happen. But it's going to happen with those as well because cloud doesn't give you the newest models to run offline. So you got to use other models.
And then all of a sudden, whatever tools and infrastructures and harnesses that you had on your production might not work with this alternative quen engine or whatever. So it's a problem. Absolutely. And it's something which I've puzzled over. I'm going to ask you this question. Why do you think that we have this huge growth, as you say, quen and deep seek?
I mean, there's another one called Quickie, I think, coming out of China. And we obviously have, as you say, Claude. We have ChatGPT. We know all I see in Europe is Mistral. Why do you think that is? So there has been a lot of criticism regarding the European regulation, which is hindering either innovation or quick progress. I don't live in Europe, so it's hard for me to say. I'm just reading what others are saying. If it helps, I do think that we Europeans, I live in Portugal, so we Europeans do tend to regulate first and innovate later. The innovation is bypassing the regulation? Yes. OK.
So that is that. And it might just be a thing of resources. Yeah. Because this stuff's expensive. It's expensive. And the people that innovate, they cost a lot of money. And Europe doesn't have as much in terms of this super technical population as, let's say, Silicon Valley. People also mention Israel and definitely China, because they do send their people all over the world to learn from the best. But this is like a macroeconomics question. I can only speculate. It is an interesting one. I think I heard one of the keynotes yesterday.
I think Jason said 93% of our data is held on servers in the U.S. Yes, but you're Europeans, and for you, it's an issue. But for the rest of the world, not as much, and therefore maybe that is the source. Because if you're using Gmail, think about that. If you are using Gmail, anybody who has ever corresponded with you, Gmail knows half of what they're saying. They don't even need to use Gmail, right? But Google knows their side of the conversation because they read the incoming email.
This thing about making it super important where your data is because we want to be able to force our rules on you, that is fine. It's a legitimate decision when you understand the implications. And of course, I don't live in Europe, so I wasn't part of the discussion. And maybe I would decide differently, but that's not for me to say.
But yet, there are implications that maybe what we're seeing now in the AI world might be giving an incentive to reconsider some of these principles. Because maybe, and I'm sorry to say that here, but maybe privacy is no longer the number one priority. Maybe staying in the game is number one, or maybe not completely losing the ability to compete in any way is number one. I don't know.
Again, it's a macro question. I'm curious. Those of you in the audience who have heard Inbal's presentation and our little chat, do you have any questions? No? Yes? But let's have some little good news would be my request. There must be something good about to happen in terms of security and governance. That's an interesting thought, yeah. Let's hit some of those. Yeah. We spent 20 minutes making you depressed. What good is going to come in the world of AI, cyber, IAM, that intersection?
In times of dire, I like to mention the story that came out, I think it was last month, about this guy whose dog had a very particular type of cancer. He dedicated a part of his life to using AI. He sequenced the genes of the dog and then collaborated with drug companies to create a first-ever drug targeted at that particular genetics of that particular dog and cancer, and the dog is no longer sick.
Of course, we would like that to happen at scale because people get cancer too. I think one of the problems in the drug business is that it is a business, and many times drug companies choose what would give them more money. But if all of a sudden, with the AI, the cost of creating a new drug is identical, whether it is for cancer or pneumonia or whatever, MS, then all of a sudden, we might just see more medication that wouldn't otherwise come to the market, I'm hoping.
I think similarly in the world of cyber security, we talked about attackers using AI and they're able to do faster, better attacks using vulnerabilities. Everyone's got a zero-day horde now. On the other side, as defenders, we can respond faster. We can isolate and we can eradicate these attackers faster and without needing that adrenaline charge that we would have had in the good old days.
I agree with that, and I think that it's not even related necessarily to AI per se, because AI is the evolution of ML, and the big promise that ML brought was the ability to process huge amounts of data in a short time. Now, AI can do it in a context and to learn from that, but at the base of things, at the core, it's the ability to do things fast and at scale. I've quoted before in this conference, the late Professor Ross Anderson, who said, security is really easy in the lab. It's really easy. It's small. It's really hard as you scale up, and I think, as you say, AI will allow us to scale up.
We can see faster responses. We can see, hopefully, better human escalations rather than us having to figure everything. We only cover maybe 1% of the incidents that are coming in.
Yeah, in this world where attacks are becoming business as usual, then our responses can likewise become business as usual. I can even say a utopic vision.
If a large part of the market is going to be using LLM for their security and incident response, and they're going to be using SaaS models and not local ones, then the SaaS companies will have the potential of recognizing something happening at one place and immediately spreading the knowledge and how to defeat that to the rest of their customers, which is something that doesn't happen as much today because of the siloing of security vendors. Those who use this vendor get their news.
Those who use the other vendor get their protection, but everybody, regardless of their security vendors, are going to be using one or two or three AI vendors, so that's a better division of the market. That's true. We're almost out of time, but maybe there's one more question. What's something that you would advise a CISO or anyone in this room to take back and say, like, we should immediately start doing A, B, or C to get ready for this? The number one thing would be to make sure of two things.
One, that they understand that they can't prevent or avoid AI. It's a lost battle because people will find a way, and then you're just not going to be able to see that and control it.
And two, understanding that without AI, your company is going to lose the edge and just fade away. You have to make sure that you are the enabling CISO and not the one turning everybody down. So instead of saying no, find a way to say yes, but or yes, and, and use your users as your own incentive to force a better adoption of AI in the organization. Because now it's down to you. Super. I think we're out of time. Thank you very much. Thank you.
