Gili Raanan
speaker
558 appearances
2 recordings
2 series
first heard Mar 2025
last heard 28 Mar
Gili Raanan’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in Mar 2026 with 1.
Appearances
I think that what we see today is that there's a global powerhouse conflict that drives major forces into cybersecurity. It's the Ukrainian conflict. The Gazan conflict, it's all over the world. Offensive cybersecurity became a little weapon or even strategic weapon for state and powerhouses involved in those conflicts.
So that's one element that's making the cybersecurity threat vector so real, so dangerous, so available. And there's a constant drift from state-level cybersecurity weapons to criminal organization and then just creep kiddies. So we live in a very dangerous world, way more dangerous than it was just a few years ago. So that's one. The second element is technology.
I think that we've seen the iPhone moment for AI in 2023 or maybe even 2024. AI is not going to improve cybersecurity. It's going to redefine it. The same technology, the same LLMs, the same AI agents that you'd use to predict attacks and prevent them would be used against you to deliver attacks. A scale and level of sophistication we haven't seen before.
And the methods that we developed maybe for the past 80 years starting in World War II would be useless because the speed, scale, and sophistication of those attacks would be something that we haven't seen before. The world is a way more dangerous place today.
Think about modern warfare and the use of technology. It started with better analysis, better decision-making process. Then it moves to human augmentation. Think about the introduction of the first tanks in the First World War or the introduction of the Kalachnikov. It made humans more lethal. But with AI, there are additional steps. It's the workflow automation.
And then the final step where AI takes control. And if you think about a modern warfare, you think about LLMs that quickly put together attack plans that you didn't practice for, you didn't plan for, you don't have a B plan and a C plan for. They're already using AI agents to take control and launch those attacks. On the different side, that means that human augmentation would be an improvement.
For time being, it might be enough, but it won't be a great solution for the long term. For the long term, you'd have to use LLMs to predict attack vectors in the very same way and give AI agents the control so that you have a chance to respond to those attacks in a timely manner. That's the future we are going to. There's no question about it.
If you've been a fan of the Arnold Schwarzenegger Terminator movies, that's been a science fiction future.
The best offensive tactics are always ROI driven. How do I apply the highest damage to you in the smallest effort, lowest cost, fastest way? And what damage is always a contextual question. Depends who you are. But we've seen examples of so many different ways to inflict pain on you on a target.
And some of those attacks had very, very simple measures that you use, but the outcome was unbelievable. Could you give an example? If you like to take off a country and put it on its knees, that takes, using conventional methods, huge amount of resources, times, and coordinated efforts by a lot of people. If you do that through cybersecurity,
You can use very simple ways like taking down name servers for their power stations, and you leave that country without power for a few days. It's not a highly sophisticated attack. You don't have to have an army of people doing that. And it can happen to any country any day.
Now, think about an AI agent that's able to produce thousands of scenarios like that and execute all of them simultaneously. That's the technology today. This is not the future thing.
The deduction model would always be as good as the data provided. Take the analogy in the drugs development industry. If you provide enough information about... a certain disease. You build an LLM that can take that patient data, that can take a lot of research and offer recipes for drugs that can, for instance, heal diseases that are considered unhealable. Let's say Alzheimer's.
You can use the same LLM and ask for recipes for drugs that can kill people. It's the same system. It's the same method. It's the same software. Now, think about LLM that's fed on data around the software infrastructure of a bank and an AI agent that you use to ask questions like, what are the potential breaches in my infrastructure and how should I protect myself? It's a very useful agent.
The same agent you can ask directly or trick to answer the question, how would an attacker breach the defenses of that bank and take that bank offline? The same technologies that we use for the defense are the technology that would be used on the offense. You've seen that even if you put a lot of guardrails into the model, there are many ways to bypass those guardrails.
And that's when AI is in a walled garden. Now you have open source models, and DeepSeek is just one example of it, where AI is not walled garden anymore. It's in the wild, and anyone can download the model and modify the model.
AI would redefine cybersecurity. It would replace the old ways cybersecurity solutions were architected. The old systems of building rule-based systems and behavior-based systems. So the good guys can fix misconfigurations or patch buggy software systems. Those days are gone.
And founders would have to create cybersecurity companies that are AI-first and AI-native just to have a chance to build lasting, important cybersecurity companies.
And that's the reason you see existing players, major players, like our own portfolio companies, you see the level of effort that a major cybersecurity platform like Wiz or Sayera or Island, the level of effort they put in AI, understanding and learning AI and applying AI into their platform, that's not an additive capability to the bot, that's
re-architecting the products to make sure they are useful and effective when the attacker is not human and is not just human augmented.
Showing 421–440 of 558 · page 22 of 28
← Previous
Next →