# ABN AMRO and Mistral: Transforming Threat Detection and Response With Gen AI (Presented by Mistral AI)

## Abstract

Key takeaways Cut SOC response times by automating threat triage with Mistral’s agentic workflows. Streamline alert classification by integrating AI-driven reasoning with existing systems, such as Microsoft Sentinel, Defender, ServiceNow, and more. Enhance compliance and audit trails with self-hosted deployments designed to meet regulatory requirements for financial services.

## AI Summary

_No AI summary available._

## Transcript

**[00:00:04 – 00:00:09]** So what I would like to talk about is about transformation.

**[00:00:09 – 00:00:14]** So here you should see, you see it now, a steam engine in a fabric.

**[00:00:14 – 00:00:24]** And in the 80s, electric motors arrived in this world

**[00:00:24 – 00:00:25]** and in factories.

**[00:00:25 – 00:00:28]** And manufacturers did one thing that they should do;

**[00:00:28 – 00:00:34]** they just replaced the steam engine by electric motors.

**[00:00:34 – 00:00:39]** They swept that, but they kept everything the same, as

**[00:00:39 – 00:00:45]** you can see, the belts, the shafts, and even the same floor plan.

**[00:00:45 – 00:00:49]** And then nothing happened for almost 20 years.

**[00:00:49 – 00:00:54]** So they were totally confused, the poor manufacturers, but

**[00:00:54 – 00:00:58]** also the economists, because the technology was clearly

**[00:00:58 – 00:01:02]** better, but nothing really changed.

**[00:01:02 – 00:01:07]** And then it took historians a while, two decades, to find

**[00:01:07 – 00:01:08]** out why that was the case.

**[00:01:09 – 00:01:12]** And the case is, if you just replace one thing by the other,

**[00:01:12 – 00:01:16]** but you don't think what you really have as an advantage

**[00:01:16 – 00:01:22]** with the new technology, you only get marginal gains at best.

**[00:01:22 – 00:01:26]** And this is when you look at what electronic power brings

**[00:01:26 – 00:01:30]** you, is that you can steer machines simultaneously at every

**[00:01:30 – 00:01:37]** time, and that allows you to build around that a totally new foreplan.

**[00:01:37 – 00:01:44]** And this took until 1920s, and then after 40 years in

**[00:01:44 – 00:01:47]** all the fabricants, all the plants, you could see that.

**[00:01:48 – 00:01:51]** And we have the same right now also with AI.

**[00:01:51 – 00:01:55]** If you look at technology, technology is not only cool,

**[00:01:55 – 00:01:57]** but it's super powerful.

**[00:01:57 – 00:02:01]** You see, every benchmark is better than the one before.

**[00:02:01 – 00:02:05]** So the technology exists, and all the companies that

**[00:02:05 – 00:02:11]** you represent already installed it, I guess so, at least, but

**[00:02:11 – 00:02:13]** nobody really organized around it.

**[00:02:13 – 00:02:18]** I heard the number, I think, a couple of days ago, only

**[00:02:18 – 00:02:21]** 6% of all the companies in the world would call themselves

**[00:02:21 – 00:02:24]** reorganized around AI.

**[00:02:24 – 00:02:28]** In Germany, it's only 4%.

**[00:02:28 – 00:02:32]** It's not as bad as you could tell, but only 4% really

**[00:02:32 – 00:02:34]** organized around that.

**[00:02:34 – 00:02:39]** And the example that I choose today now is cybersecurity as one

**[00:02:39 – 00:02:44]** of the areas where you can organize around AI and where you can

**[00:02:44 – 00:02:47]** transform how you operate things.

**[00:02:47 – 00:02:51]** Because if you look at not only the number of 4% of the

**[00:02:52 – 00:02:56]** companies that really reorganized, you can see that when board

**[00:02:56 – 00:02:59]** meetings happen, they ask themselves, why did we invest so

**[00:02:59 – 00:03:01]** much and we don't get any return?

**[00:03:01 – 00:03:05]** And when they say no return, it's exactly about this 88%

**[00:03:05 – 00:03:08]** of all the projects that never run into production, and there's

**[00:03:08 – 00:03:12]** another number, that 95% of these projects, so even more of that,

**[00:03:12 – 00:03:16]** don't have any impact on your P&L.

**[00:03:16 – 00:03:21]** And part of that is fine because prototypes are built to fail,

**[00:03:21 – 00:03:25]** some of them at least, but the majority of these projects

**[00:03:25 – 00:03:29]** really aim for production, and it's on us to explain what that means.

**[00:03:29 – 00:03:32]** And again, the problem is not the model.

**[00:03:32 – 00:03:34]** The models are good enough.

**[00:03:34 – 00:03:36]** They are capable.

**[00:03:36 – 00:03:38]** They are extremely powerful.

**[00:03:38 – 00:03:42]** It's more the problem how you can bring these capabilities of that

**[00:03:42 – 00:03:47]** models into operation and how you can bring them into an end-to-end

**[00:03:47 – 00:03:51]** process so that they really sit in the system that you built.

**[00:03:51 – 00:03:54]** And this becomes even harder for cybersecurity because

**[00:03:54 – 00:03:57]** there you need to operate these systems in a bank.

**[00:03:57 – 00:04:01]** So the example that I brought a bit today, you need to work with

**[00:04:01 – 00:04:03]** the sensitive data that they have.

**[00:04:03 – 00:04:07]** You need to work with the existing tool that they already have, or

**[00:04:07 – 00:04:09]** at least to want to integrate them.

**[00:04:09 – 00:04:13]** And you need to follow very strict security requirements on that.

**[00:04:13 – 00:04:17]** And if you look at the possibilities of generic

**[00:04:17 – 00:04:24]** AI, then generic AI understands, in general, cyber security.

**[00:04:24 – 00:04:28]** Also, our models that we run, so the frontier models that

**[00:04:28 – 00:04:31]** we have, understand cyber security in general.

**[00:04:31 – 00:04:34]** But it does not automatically understand your environment.

**[00:04:34 – 00:04:37]** It does not automatically understand your edit history,

**[00:04:37 – 00:04:40]** security policies that you run, and the escalation process

**[00:04:40 – 00:04:43]** that you need to have in place.

**[00:04:43 – 00:04:46]** And this enterprise context is super relevant to

**[00:04:46 – 00:04:50]** really bring something end-to-end into production.

**[00:04:50 – 00:04:54]** So the difference is between AI that understands a generic

**[00:04:54 – 00:04:58]** task and AI that can run in your environment.

**[00:04:58 – 00:05:03]** And to look at the, like, the spectrum that you can use

**[00:05:03 – 00:05:07]** for that is, on the one hand side, you have these productivity tools,

**[00:05:07 – 00:05:09]** so the chatbots of the world.

**[00:05:09 – 00:05:14]** So they are super useful, all of us are using it, they are measurable,

**[00:05:14 – 00:05:16]** but they are also replaceable.

**[00:05:17 – 00:05:21]** So what we see on the other side of the spectrum are iconic

**[00:05:21 – 00:05:26]** use cases, so something that you custom build, that you really

**[00:05:26 – 00:05:31]** look at, where can I use my model, maybe even fine-tune my model?

**[00:05:31 – 00:05:33]** Where can I give context to my model?

**[00:05:33 – 00:05:37]** Where can I build the end-to-end process to my model so that

**[00:05:37 – 00:05:40]** I can deploy it in my own sovereign environment, so

**[00:05:40 – 00:05:42]** that I'm secured on this?

**[00:05:42 – 00:05:47]** And security operations sits toward this.

**[00:05:48 – 00:05:52]** It's right side for you, so on the right side of this spectrum.

**[00:05:52 – 00:05:56]** And there is also where the value then is created.

**[00:05:56 – 00:06:01]** So the more this AI understands your enterprise context, and

**[00:06:01 – 00:06:05]** the more it's integrated in your operational workflow,

**[00:06:05 – 00:06:07]** the more value you can create.

**[00:06:07 – 00:06:12]** And if you think again about this electronic story, you

**[00:06:12 – 00:06:16]** see that in the 1920s, the winners were not the ones that used

**[00:06:16 – 00:06:20]** the most powerful engine or model.

**[00:06:21 – 00:06:24]** The winners were the ones that started to build everything

**[00:06:24 – 00:06:29]** around the real value creation

**[00:06:30 – 00:06:32]** and that redesign around that.

**[00:06:32 – 00:06:36]** And the example that I bring today is how we together with

**[00:06:36 – 00:06:41]** ABN AMRO transformed this exactly for cyber security.

**[00:06:41 – 00:06:46]** So ABN AMRO, who here is from Dutch, who is Dutch?

**[00:06:46 – 00:06:52]** We have, okay, so you know it, is the third biggest Dutch

**[00:06:52 – 00:06:58]** bank, and they have to operate some complex challenges.

**[00:06:58 – 00:07:02]** And if you just remember what happened, I think, two days ago

**[00:07:02 – 00:07:07]** with the PHB2 shell vulnerability that we all appear, this is

**[00:07:07 – 00:07:11]** something that this bank has every day, every minute, every

**[00:07:11 – 00:07:17]** second, that they get like this, ATT&CKs, and they have

**[00:07:17 – 00:07:23]** to build a team around that, so a security operations center team,

**[00:07:23 – 00:07:25]** and they are operating on that.

**[00:07:25 – 00:07:29]** And we look together with them, so how can we make that process?

**[00:07:29 – 00:07:34]** How can we transform that process by using AI and our technology?

**[00:07:34 – 00:07:38]** And this is a typical workflow that you can see here now that

**[00:07:38 – 00:07:42]** is taking some security signals.

**[00:07:42 – 00:07:45]** Then you have a manual triage.

**[00:07:45 – 00:07:48]** So you have an agent that has to enrich the data, have

**[00:07:48 – 00:07:53]** to find context around the security incident, and then

**[00:07:53 – 00:07:56]** they have to make a decision.

**[00:07:56 – 00:07:58]** Is it important?

**[00:07:58 – 00:07:59]** Is it a relevant alert?

**[00:07:59 – 00:08:01]** Do you need to act on it?

**[00:08:01 – 00:08:04]** What is the level of risk that you are facing?

**[00:08:05 – 00:08:08]** And does it require any escalation that you want to do?

**[00:08:08 – 00:08:12]** And then in the end, what you then see is the action.

**[00:08:12 – 00:08:16]** So only after this time span, you really can react.

**[00:08:16 – 00:08:19]** So you're always a little bit too late because you have

**[00:08:19 – 00:08:23]** to make this human triage on every single incident.

**[00:08:23 – 00:08:28]** So a lot of that is repetitive, so something where we know that AI

**[00:08:28 – 00:08:31]** maybe is better suited than humans.

**[00:08:31 – 00:08:36]** And a lot of that is information dense, again, something where

**[00:08:36 – 00:08:42]** AI is called to be rather good, and that's why we then changed

**[00:08:42 – 00:08:44]** that process together with them.

**[00:08:44 – 00:08:46]** Now we rethink the whole.

**[00:08:47 – 00:08:53]** By the detect still happens, of course, but this is something

**[00:08:53 – 00:08:57]** that is then reported through tools like Sentinel or Defender.

**[00:08:57 – 00:09:03]** And AI can already enrich in this phase, and collecting

**[00:09:04 – 00:09:08]** and structuring the information that is provided automatically.

**[00:09:08 – 00:09:13]** And then Mistral does the reasoning on that, so that

**[00:09:13 – 00:09:16]** not the human itself has to do the reasoning in the first place.

**[00:09:16 – 00:09:20]** But of course, it's not that we take the human out of the

**[00:09:20 – 00:09:22]** loop, it's still in the loop.

**[00:09:22 – 00:09:25]** It's just that some of that incidents are triaged out

**[00:09:25 – 00:09:28]** and then automated by, in this case, ServiceNow.

**[00:09:28 – 00:09:33]** And still the human is in control and then still can

**[00:09:33 – 00:09:37]** analyse or can react, but it's just that they don't

**[00:09:37 – 00:09:40]** have to do it on all the cases.

**[00:09:40 – 00:09:43]** They're just building everything around the new technology

**[00:09:43 – 00:09:46]** that helped them and put that into the middle.

**[00:09:46 – 00:09:50]** So far, this is just a story that you can hear always

**[00:09:50 – 00:09:53]** because it's just a prototype, it's just the pilot.

**[00:09:53 – 00:09:59]** The reality is that we are only building use cases like this

**[00:09:59 – 00:10:01]** that we also bring into production.

**[00:10:02 – 00:10:05]** And this is where we need the partnership, and we are

**[00:10:05 – 00:10:11]** happy that we have the partnership with NVIDIA because combining our

**[00:10:11 – 00:10:16]** models and our whole product suite that we have together with our

**[00:10:16 – 00:10:21]** forward deployed engineers and the NVIDIA AI Factory Infrastructure,

**[00:10:21 – 00:10:24]** we get three results out of that.

**[00:10:24 – 00:10:27]** First of all, we come to the performance that we need so

**[00:10:28 – 00:10:31]** that we really are better than the original process.

**[00:10:31 – 00:10:33]** This is the first thing that you need.

**[00:10:33 – 00:10:38]** The second then is that you can scale that on a bank level so that

**[00:10:38 – 00:10:42]** you can productionize it and you're not just running it as a pilot.

**[00:10:42 – 00:10:46]** And the last is that you can control it so that AI runs

**[00:10:46 – 00:10:48]** in your environment.

**[00:10:48 – 00:10:52]** So what is your environment, and why is that so important?

**[00:10:52 – 00:10:58]** So first of all, AI becomes embedded in your core process, and

**[00:10:58 – 00:11:01]** you should own your core process.

**[00:11:01 – 00:11:04]** Not only on the controls, so that you understand what

**[00:11:04 – 00:11:09]** you share with the model, but also how to access always

**[00:11:09 – 00:11:12]** the model that runs your business.

**[00:11:12 – 00:11:14]** But let's go back to the one that you share.

**[00:11:14 – 00:11:17]** You would say all the vendors are saying that they don't

**[00:11:18 – 00:11:19]** share information.

**[00:11:19 – 00:11:23]** And you are right, all of that is promising that.

**[00:11:23 – 00:11:27]** And some of you that may have studied mathematics or are

**[00:11:27 – 00:11:31]** just very interested in, who knows what that is?

**[00:11:31 – 00:11:34]** You're allowed to yell.

**[00:11:35 – 00:11:36]** Yes, exactly.

**[00:11:36 – 00:11:38]** Thank you very much. So it's Navia Stokes.

**[00:11:38 – 00:11:45]** It's one of the seven World Millennium Prize problems.

**[00:11:45 – 00:11:47]** And it's not about claiming that somebody solved it.

**[00:11:47 – 00:11:49]** Who solved it?

**[00:11:49 – 00:11:53]** Yes, AI agents, thousands of them solved it, and there were

**[00:11:53 – 00:11:58]** two mathematicians that also tried to solve it, and AI was faster.

**[00:11:58 – 00:12:02]** So it's not about to decide who wins or not.

**[00:12:02 – 00:12:06]** It's more about that the model provider, they worked on a

**[00:12:06 – 00:12:10]** model for a year, and then they had an argument with

**[00:12:10 – 00:12:12]** them, who is responsible?

**[00:12:12 – 00:12:14]** Did they leak the data?

**[00:12:14 – 00:12:16]** Did they reuse the data?

**[00:12:16 – 00:12:19]** And I totally trust that they said, no, we did not use in

**[00:12:19 – 00:12:21]** the first place your No data.

**[00:12:21 – 00:12:24]** But there was a sentence in what they gave as a response that

**[00:12:24 – 00:12:32]** I find astonishing, and that is, we cannot rule out that de-identified

**[00:12:32 – 00:12:39]** data derived from the uses of our products helped improve our models.

**[00:12:40 – 00:12:47]** So if you take that into context for your data, for your security,

**[00:12:47 – 00:12:52]** you don't want to have that sentence as part of your strategy.

**[00:12:53 – 00:12:56]** You want to have something that you really own.

**[00:12:56 – 00:12:59]** Because for the mathematicians, there was only their fame

**[00:12:59 – 00:13:03]** maybe on stake, enterprise of one million dollars, but for you it

**[00:13:03 – 00:13:06]** will be the future of your company.

**[00:13:06 – 00:13:10]** So you don't want to have that risk, and that is why

**[00:13:10 – 00:13:15]** serenity is not only a Decision of your infrastructure, but

**[00:13:15 – 00:13:20]** it's a decision of the AI strategy that you want to do.

**[00:13:20 – 00:13:25]** So, to recap, to bring something like this into production,

**[00:13:25 – 00:13:29]** you first of all need models.

**[00:13:29 – 00:13:32]** Yeah, that's true, but maybe they are only 10 to 50%

**[00:13:32 – 00:13:33]** of what you need.

**[00:13:33 – 00:13:35]** Then you need some context.

**[00:13:35 – 00:13:38]** It can be enough that you just take all the context

**[00:13:38 – 00:13:42]** that you have, you can fine-tune the model, you can do a lot to

**[00:13:42 – 00:13:45]** make that model fit to your needs.

**[00:13:45 – 00:13:50]** And finally, you also need infrastructure to run it on

**[00:13:50 – 00:13:54]** so that you can operate that securely and reliably.

**[00:13:54 – 00:13:57]** So these are the ingredients that you need.

**[00:13:57 – 00:14:01]** And if you put all these together, you're not just running an

**[00:14:01 – 00:14:06]** AI assistant, but you can put really something in the middle like

**[00:14:06 – 00:14:12]** AI and transform your organization, your security piece around that.

**[00:14:12 – 00:14:17]** And then you can operate that on the enterprise.

**[00:14:18 – 00:14:21]** So, let's bring that back to the story that I shared

**[00:14:21 – 00:14:24]** in the beginning.

**[00:14:25 – 00:14:31]** The value of adding AI to the enterprise is not like

**[00:14:31 – 00:14:34]** an electricity, to just adding it.

**[00:14:34 – 00:14:38]** You really have to build around it because, if you look at

**[00:14:38 – 00:14:44]** the companies that really get the most push out of electricity, where

**[00:14:44 – 00:14:48]** the companies not with the most powerful motors, not the one that

**[00:14:48 – 00:14:53]** started first, but the one that built that in the middle, because

**[00:14:53 – 00:14:57]** this effect compounds over time.

**[00:14:57 – 00:15:02]** This means you should rebuild your enterprise around AI.

**[00:15:02 – 00:15:08]** Be the next 4% or 6% that are out there in the world.

**[00:15:08 – 00:15:11]** And this window is open now.

**[00:15:11 – 00:15:15]** Technology is here, it's open now in Germany, in Europe.

**[00:15:15 – 00:15:17]** You just have to do it.

**[00:15:17 – 00:15:18]** Thank you very much.

