# Miro CISO on AI Security Strategy

**Podcast:** HMZE
**Published:** 2026-09-03

## Transcript

Everyone is just waiting for the flattening out.
And then they're thinking about the bubble, of course, the AI bubble.
But does it flatten out?
And even if it flattens out, most organizations have at least two or three years worth of backlog to actually just adopt.
Welcome to another episode of our new season of Beyond Vibe Coding, partnering with Impala Search.
The go-to tech and executive search agency in Germany.
In this podcast, we explore the transformational change in software engineering and knowledge work in general.
I am Sebastian Heidemeyer zu Erten, CTO at Ecosia.
And I'm Andrei, CTPO at Trusted Shops.
Great to have you back.
In our episode with Franzi from Miro, she praised the security approach at the company.
Und guess what?
Today we are talking to the person responsible for information security at Miro.
And I need to admit, I'm so excited.
Yes, so am I.
And for good reason.
Marc is very experienced.
He has seen so many changes in his career and is giving a very calm and pragmatic perspective.
And we really enjoyed the episode.
We hope you enjoy the discussion as much as we do.
So, yeah, take a listen.
Welcome to the episode, Marc.
We're super happy to have you.
It's been a long time in the making and I'm really, really glad that now we can finally speak to you.
We learned a little bit already about the security stance and posture at Miro from Francie a couple of episodes back, but now we're talking to the man who's responsible for everything.
And yeah, like usual, please introduce yourself briefly to our listeners.
Marc Strand, und ich bin CISO für Miro für drei Jahre jetzt.
Vor dem, ich war CISO für Klarna für sieben Jahre und eine andere Bank, die auch eine German subsidiary hatte.
Ich habe ja, ich habe ja, ich habe ja, ich habe ja, ich habe ja, ja, ich habe ja, ja, ja, ja, Always doing startups or scale-ups.
So the younger generation of technologies.
I started in 1996, so I've been doing this for 30 years and started with taking stock market trading online.
So I'm sorry to say to Trade Republic that they weren't first, but they certainly are coming hard into the market.
Awesome.
Das ist ein langes Zeit.
Und dann, ja, du hast es alles gesehen.
So die ersten Videos von der ersten Sprechung auf die Web auf die Web für alle.
Alles, von wirklichem Internet in deinem Home, weil ich begann, wenn du noch immer noch ein Dial-Up Modem hast, und wir haben alle die Abenteuerung auf die Broadband.
Aber auch die Home PC.
was kind of under during the years that I worked.
So it has gone full spectrum.
It was the dot-com era, dot-com boom, dot-com crash, financial crash.
So there's many different phases to this.
And it's interesting how some of these patterns kind of come back and repeat and how they are now again.
basically repeating into the next paradigm shift.
We need to save that for later.
Really looking forward for these kind of patterns.
We just had a recording where we also talked about past events and how we can use that to maybe not just propose what is going to happen or predict what's going to happen, but also to create some kind of safety.
But would really like to hear your wie du schon erwähnt hast.
Ich würde ein note nehmen.
Ja, ich würde ein note nehmen.
Let's double click on that later, für sicher.
Ja, eine Frage, kurz vorhin.
Du hast in den letzten 90ern, du hast du stock trading online.
Es war eigentlich ein website, wo end-consumers konnte, basically, trade stocks?
Ja, so wir haben die Nordics-First Platform for online trading.
There was one actor in Sweden that did it before, but they had basically a mail to form that sent the transaction to a trader.
But we had something as strange as a Java applet back in the day running against a Java server whilst Java was still in beta producing real-time data or stock market feeds.
und die Transaktionen all die durch, nach dem Stockmarkt.
Das waren die Zeiten, die Leute noch immer sehr große Phones mit kleinen Schrainen hatten, aber sie haben ihre Stockbroker geholfen und sie haben sie zu tun.
Ich erinnere mich sehr, wenn die Traders, die wir da waren für unsere ersten Transaktionen, waren sie Drinking champagne, but jokingly saying, regular people trading on the stock market.
Yeah, that will be the day.
And well, it wasn't more than 10 years later, they were almost basically out of a job.
So it's interesting how these things happen.
And they go very, very quickly.
But they also...
in a very smooth transition that you don't even really feel it until you kind of realize, hey, everything just changed.
Yes, yes.
That's pretty much exactly what we also discussed, that it takes time for people to adapt, right?
But the wave is coming.
It's just like more smoothly than you probably predict on the short term.
How many years was it that you did your You had to go to your bank office to do most of your financial transactions and do most of the stuff that you had to deal with.
And now you have open banking, you have these APIs that are available, you have so many more capabilities and you just don't realize that, hey, it wasn't that long ago that that wasn't available.
Absolutely.
Yeah, and I vividly remember phone banking because Phone banking was the intermediate step, right?
You didn't have to go to the bank office, but you could do it from home via the phone.
I still keep those brown envelopes that we used to take the bills that we need to pay and the little green stamp that I had that I had to stamp it and sign and send it into the bank.
I still kept that one because it's kind of, I came into that era with Das wirklich macht das Paradigm, das ich mit dem Online-Denken-Tag schaffe.
Fassinant.
Ja, so wir könnten uns vielleicht über diese verschiedenen Zeiten und die Erfahrungen über diese Zeiten und die Erfahrungen über diese Zeiten.
Aber wir werden uns vielleicht zu etwas, was André noch ein paar Punkte später in der Episode.
Aber jetzt kommen wir uns zu unserem übrigen Segment, die Status Quo.
So, wie Sie arbeiten in der Zeit, was Sie Ihre Tech-Stack, was Sie arbeiten?
LLMs or Agents are you using?
Well, my status quo, my tech stack currently is it's pretty predominantly being on Claude most of the time.
Cursor, OpenAI is still like it's coming back into my stack again.
It left for a while.
Gemini is also for some use cases.
And then there's many other specialty models that I didn't use to have that many different specialty tools for different use cases.
But now I have much more of a palette of different things.
But other than that, and then I usually use Miro to kind of tie everything together.
Das ist die Platz wir arbeiten.
Das macht es einfacher für mich.
Ich lege mehrere Teams, so für uns die Collaboration space ist eine Quintessenthalte, die vorhanden Teil des Spiels.
Das macht viel Sinn.
Ich habe eine similar Art von Tools, so ich de-installed Cursor.
Ich habe es nicht mehr benutzt, ich habe es nicht mehr benutzt.
Ich benutze Antigravity als mein IDE jetzt.
Aber ich benutze Cloud, ich benutze OpenAI also, wieder, nachdem ich es für ein paar Jahre alt war.
Aber ich liebe Codex.
Ich bin ja auch so, wie Sie entscheiden, was Sie die verschiedenen Tools für benutzen?
Oder ist es einfach so, was Sie kommt?
Es ist heuristisch, mostly listening to others who are experiencing or testing something out, and then they share their experiences.
Es ist nicht einfach an easy thing to disseminate the knowledge and how to do efficient work.
And it keeps changing so fast.
And I can tell you that over the last year or so, I had to basically go and reset my thinking several times.
Because the things that I knew last summer wasn't applicable by basically December.
And because I didn't have time to kind of focus on kind of retooling my own and putting effort into my own efficiencies and the day-to-days.
I had to then basically take a step back and take several days to reset and recalibrate.
And participating in another company's hackathon was actually one of those things that I did, was to actually learn their harness, their ways of working and how to accomplish.
And that really helped me to reset how I was doing my work.
And then I could bring that back to my team.
There's so many things that keep on happening.
And even the things that I did in January, some of them I have to refactor now in June when I did a hackathon with my own team.
Because it wasn't applicable anymore.
There were new things that made it much more efficient.
Very interesting approach.
Was it the same company or a different company that I got it right?
It was a different company.
Of course, I participate in the things that we do internally within our company.
I want to be part of as many spaces within our engineering as possible.
So basically talking both kind of the infrastructure side of things as well as the application side and so on.
So having those multiple and even working with the data people.
Because I have that diversity, I kind of have the opportunity of actually seeing multiple different.
But it does constrain you a little bit, because every company has some kind of boundaries of what tooling you bring in and what ways of working.
So going outside of this and working and doing a hackathon with a completely different company is actually quite useful.
And I think that we also bring that back into I do some executive sponsorships, so where I meet other executives and tell them about how we're working and sharing that.
And I think that this sharing is one of those key elements that we as an industry need to have because it's moved so fast and we can't listen to all the marketing BS, to be honest, right?
So we have to really see what really works.
Und was ist das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, was das, Ja, wir haben die gleiche gleiche sagen in Sweden.
Es ist ein sehr genauer und ziemliches Thema, das in der industrie richtig gut passt.
Wo kann ich auf den Hackathon sein?
Not to the hackathons for the exchange you just mentioned.
I thought so often about these kind of models and I also gave it a try some years back.
And I realized it's super hard because you need not just finding companies.
I think you get companies.
You need to have NDAs and all that stuff in place.
And so it gets so complicated.
So I really like the approach.
We always want one of the things is like we do this on our home machines in our free time is the fact that we can't do this with other companies.
Like this is like the podcast is the best thing we could thought of.
But this is such a great idea.
So that's why I asked where I can sign up.
But yeah, I would say.
You have a great startup mentality and startup culture and startup scene in Berlin, right?
You have a lot of different small companies and going to them.
And I know that legal will hate my ass for saying this, but having a good friend somewhere and then asking, hey, can I just join your hackathon and guest?
And maybe you need to sign an NDA as a...
oder so.
Aber das ist die einfachste Weg.
Wir haben eine sehr starke Start-up-Seine in Stockholm, so es ist ziemlich easy für mich zu gehen in die Stadt und essen mit jemandem und dann irgendwie, okay, komm nächste Woche und wir machen eine Hackathon.
Und es ist immer ein bisschen disconnected, wie du es auf eine bestimmte Task ist.
Das ist eine gute Weise für eine gewisse Nachfrage.
I can tell you that sitting inside of these CISO communities, I tried to sit in several of them where we were speaking about the work that we were doing.
And I noticed the same thing was applied to engineering as well.
We were coming into meetings and talking about the work, not really doing the work.
And it was more of like this...
internal club for mutual admiration, where everyone was kind of trying to show off and better off what they're doing.
And it didn't bring you forward.
You didn't learn anything, really.
And of course, you have these hacker conferences, like DEF CON was actually amazing, and so on.
So there's many of these.
where you can actually learn from others.
Awesome, yeah.
Let's just dive one level deeper into your tech stack before we then move to the main topic of this episode.
You mentioned different kinds of providers that you're using, right?
Also, do you use some skills or agents that you use for regular things on a daily basis or something?
Yes.
Most of them I have built myself to kind of define based on knowledge and experience that I have.
And trying to kind of identify how I do a certain task is kind of quintessential to that.
Of like, how do I break down making a business case?
Or how do I break down evaluating a business case?
Most of these tasks that you do, being able to create a workflow of it and trying to describe what is important in it.
And then there are certain skills and things that I definitely have asked AI to research for me.
And to be honest, with the newest models, it becomes significantly much better.
But you also have to give it some good tests so you know that you're getting a quality out there.
And then finally, there's now starting to become a marketplace for skills.
I mostly am fascinated by the legal skills, to be honest.
I'm very much waiting for someone to publish more accounting skills.
Mein Personal Accounting und das Dinge, ich liebe das, so finden gute Wäste zu tun, und effiziently, das ist wirklich etwas, was ich gerne mehr zu haben, aber ich habe noch Zeit, zu über.
All right, so du dann wahrscheinlich hast du deine Skills in deinem Cloud set, und dann bist du eigentlich die Skills that du need, wenn du sie need.
I also then separate them into the agents so that they can operate independently and have their own context window so that you kind of get that not adversarial approach, but at least two opposite sides so that you get the segregation of duties between them.
Yeah, but I assume as a CISO you never experience Ich habe auch schon experimentiert mit OpenClaw oder so.
Oh, ich habe.
So, ja, ich habe.
Aber ich finde es nicht so gut, weil es die Dynamik von dem Ding ist, es consumes so viel Energie und so viele Tokens, um zu kommen zurück mit etwas, das eigentlich nicht so gut ist.
Und dann hat es...
Es lacks a lot of constraints that I like to have.
So, to be honest, I'm still looking for the companies that are operating on top of OpenClaw and say that they are successful, not just by doing one thing, but actually consistently being successful and doing the same thing, not going from one industry into another one.
Ich kann mich auch nicht mehr so sagen.
Ich persönlich habe es aber auch immer noch immer noch.
Ich habe diese gewohnt-skills- und-workflows-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein-ein- Das ist für sehr specific cases.
Und ich denke, Hermes Agent und da sind auch andere frameworks, die sind vielleicht besser suited für die Unternehmen.
Und vielleicht nicht even Hermes Agent, aber ich denke, dass andere frameworks sind, in order zu properly runen eine Unternehmen, wo man eher, ich weiß, haben eine Meta-Harness und dann verschiedene Agende, die sind von bestimmten Events und was auch nicht.
Wir haben einen Einwohner, oder ein Team-Members, die sich selbst für das, um das, um das, um das, um das, um das, um das, um das, um das, um das, um das, um das Plastik zu verabschieden.
Aber es war mehr ein security-focused type, um das OpenClaw-Aproch.
Aber die Herausforderung ist, dass es It can go after opportunity, but it doesn't seem very proficient at performing a consistent task.
And a lot of things that we have in our work is repeated consistent tasks.
I definitely want to give out work to do research, but now the models are pretty good to actually go do the research without having a large...
Harness of multiple agents to actually achieve it.
So I'm pretty happy with what you get out of both OpenAI and definitely Anthropics.
For sure.
The harness doesn't matter as much as the skill files, basically.
The markdown files that describe exactly what it should do, what it shouldn't do, what it should look at.
Tools that it uses in order to validate that the links for the showed notes are actually working.
So things like that, right?
That is more important.
It's not the harness itself.
It's just a legacy, if you wish.
Anyhow, we dove pretty deep into the tech this time.
Maybe we can even cut out certain parts of it because I'm not sure if everything of that is relevant for our listeners.
But anyway, now we should...
really segway into the main section of this episode.
And yes, as I mentioned already in the beginning, that our episode with Francie really sparked our interest because she was speaking so enthusiastically about your security posture.
dass wir uns so glücklich haben, dass wir uns so glücklich haben, die Sie so glücklich haben, die Sie kommen von den CISO haben, dass wir uns zu sprechen, dass wir uns zu sprechen, und das ist die Chance, wir haben jetzt.
So, ja, bitte, wir uns wissen, was Ihre thoughts waren, um die Prozess zu Ende des Jahres?
Du hast schon gesagt, dass du bestimmte Dinge in January introduced hast.
Und vielleicht kannst du uns was, was du oder musst du schon jetzt schon mal wieder auf den Refaktor?
Half a year, right?
So things move fast, so we need to change.
Well, one of the things that we started off with as we were moving into kind of how to create checks and balances, especially when Claude Code was developing and new enterprise features were coming out, there wasn't much guardrails that you could implement, right?
Most of the things that you get out of the box wasn't that powerful in the very beginning.
And figuring out how to kind of combine being on the bleeding edge whilst at the same time getting some levels of controls.
And I think that the best way to look at it is to kind of the combination of both working with base layer controls that Give you foundation.
Call it a backstop.
Most confuse guardrails with a bunch of different things.
I consider guardrails to be that backstop.
They are not immovable rules.
Guardrails can't change.
Otherwise, they wouldn't be guardrails, right?
And then you want to add other softer.
Just like you do with humans.
You train humans, create awareness, you have all of these soft things like policies that you then expect people to kind of live by.
But agents and AI don't know this.
They don't have that context.
They don't have that information unless you actually package it up for them.
und es ist, weil die Kontext-Windel der Human-Mind ist, dass die Mind-Blöhne ist, während der AI ist relativ limited.
Es ist sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr, sehr They are not necessarily hard rules, but things that should be considered when you're doing your work.
So this was kind of like the first things that we started implementing.
And we came pretty quickly to kind of 11 rules for our engineering teams that we wanted to kind of have.
And like the last one we had to kind of add to it was...
Okay, we don't want AI to actually work on private repositories.
But there are cases where we want them to work on private repositories because our engineers are contributing into other open source projects or we have open source projects and so on.
So we want to be able to contribute as well.
So we can't make that a hard stop.
We have to make it more of a soft like.
Did you consider this before you take that action?
And that was kind of like the first step.
How do you provide these 11 rules?
So some of them seem to be a bit stricter.
Some of them, like it's soft, you should consider them.
But is there some kind of packaging?
Like do you package that and put them into the...
Any kind of harness you have at Miro?
Or how does it work?
Because I can imagine your engineering organization is pretty large.
So ensuring a certain consistency in achieving standards is not as simple.
No, and we also want to kind of allow for a lot of experimentation and differences within the engineering org.
Whilst you still want to move to some level of harmonization, just like you have most engineering organizations, you want to have a software development lifecycle that is fairly homogenous, that the common denominators are kind of shared across the organization.
But then they might change for certain projects in certain areas, right?
True.
So the way that we...
ist, die enterprise capabilities, wie wir mit dem Md-Filesen und sogar mit unseren Device-Management-Systemen zu den Endpoints haben.
Wir haben unsere eigene Sandbox, die wir in Haus haben, mit einem gewissen Wissen.
Wir haben jetzt auch eine Agent.
that we talk about on our blog, where we can run a lot of these standardized requirements into them.
And then there is the more hard rules of like, okay, how do you control all your MCPs and all your integrations?
Because that's usually where you really are trying to kind of really want to control things.
And then you have the Also, the really hard rules of your pipelines, which can basically instill requirements into your development lifecycle as well, where you can have tests that break the pipeline or so on.
So you have to combine all of these together.
And it's not just one solution for everything.
It's like layering this all on top of each other.
On top of this, you add a real-time detection capability that understands AI.
We have been working with one company that basically gives us that capability and they are very quick to adopt to the new technologies that arise on the market, which is quintessential because you can't sit and wait.
Because sometimes we want to be using some newer technology that is really bleeding edge and it might be in a research preview.
But you want to still use it in contained environments.
So you want to have control over that and not just block it until it's mature.
So it's again the...
Swiss cheese, pretty much layering approach, right?
Where not all the holes would be directly in line.
Visible.
Exactly.
It's an onioning of security layers that you have to put together and it's all holistic.
But as soon as you're...
The warning sign goes off for me when you just have one layer that you're dependent on.
It's only access controls, or it's only this, or it's only this data classification.
It's not only, it's how you add all of these capabilities together.
And I think this will become even more important as we kind of move forward, because I'll give you another example, which is the one that I find most interesting and most challenging right now.
And that is...
Identities and entitlements, for example, for agents.
We all work in organizations that have access management and almost every organization suffers from overprivileged roles.
Because you don't want to have too many roles to administer, hence you overprivilege users and then you trust the user to impose some kind of Segregation of duties or not combining different things or being the soft layer in between the different process steps, right?
When you take AI into it, then you can't do that.
This paradigm basically falls apart.
You don't want to over-entitle an AI, so then you have to kind of 1,000x or 100x the amount of entitlements or roles that...
you needed to have for humans.
So you need to have a new approach of automatically assigning them for the specific use cases because we need to separate different privileges from avoiding different toxic combinations.
A good example is the agent that receives your invoices shouldn't be the same one that basically does the accounting.
die Transaktion in die Finanzierung und dann geht es auf und dann geht es und dann geht es und dann geht es und geht es um die Bank.
Du wahrscheinlich wollen hier eine Humanität, in all of das.
Aber du hast nicht einen Agenten auf multipleen dieser.
Du willst sie auf dem Segregated, so dass es Segregation in zwischen.
Und dann Because you want to both protect yourself against the malicious, like the prompt injections.
The simplicity of that AI just confuses data with its instructions.
It's the power of the AI.
It's the same thing as the Achilles heel.
It's the same as an SQL injection.
It can't distinguish one from the other.
So prompt injections haven't fixed, but even if you don't look at the malicious ones, you have all of the non-malicious ones, all the baked-in assumptions that you have in your head, which you give to the AI or the agent to execute on.
Whilst you didn't explain all of the pieces that were still stuck in your brain, and now it does something different.
And how many of us have sat there and was like, oh no, but that's not what I wanted.
And have to go back and kind of redirect or restart.
And my principle is usually I want to restart it because then I can actually make it much better.
Because I can figure out what was missing in my mental model.
Und natürlich, hallucinations, aber wir alle wissen, aber sie sind besser geworden.
So, mostly, die intuitive Dinge oder die Assumptions, die ich habe, die ich habe, habe ich nicht in meinen Agents, sind es missing.
So, das ist, warum ich wirklich haben, diese Segregation.
Und das, ich denke, ist, dass es wirklich schwierig zu fix werden, weil die Tool Sets nicht mehr sind für das.
You also mentioned a technology that you use in order to detect malicious activity, right?
So can you describe a little bit how that works, how you use this?
So it basically taps into the APIs on the model providers.
So it can detect malicious patterns.
It has anomaly detections.
They run investigative agents and then combine that with human investigators as well.
So that allows it to be much more fast.
And then the very large data enrichment that they do, they enrich with a lot of data sets and they have been able to add more data sets quickly for us, which is quintessential.
And then the enriched data, Actually can show those patterns that previously we were stuck with just using simple pattern recognitions in singular dimensional logs that came from one source.
Now we can just basically combine them all together and see, oh, these combinations are, that's not a good thing.
Interesting.
Das ist ein sehr guter Baselayer.
Wenn das in place ist, dann ein paar Sachen, die in anderen Bereichen kann man bei diesem Baselayer verletzt werden.
Oder es kann man sich verletzt.
Ja, und die Visibilität ist sehr wert.
Ich nehme die Mini-Scheihalud.
Wir hatten Teams, die die Affecten Repository nutzen.
Within hours, we could actually get a clear picture that none of them had actually pulled the affected code at that point in time, which meant we could verify that we were safe.
And that wouldn't have been the case with, for example, log4shell or something like that just a few years back, right?
And I think that capability of doing that correlation and being that quick is extremely valuable.
And also being able to kind of send in as much data as you absolutely can without there being an insane amount of cost to it.
Because a lot of these seams, they bill you by the data set you pull in and that becomes very expensive very quickly.
But it also loses its efficiency if you try to kind of be saving on something.
For sure, yeah.
So most of the topics that you mentioned so far have been like focused rather on like, no, no, it's a mix.
Not only engineers, right?
So MCPs and integrations are also relevant for other knowledge workers.
Pipelines is more an engineering thing.
Franzi also told us that she was very confident or she knew exactly for what she could use, which AI, and also how to handle PII.
Can you maybe double-click on this topic a little bit?
Well, it's a combination.
You have to do awareness training.
Awareness Training Modules.
This is a pattern that I started off already in my previous company, where we built these highly tailored trainings, because it's not very useful to listen to a small talk about GDPR and personal data in the generics.
Most of the people that at least we employ, they are...
People who are smart, they have been around and they have heard most of the things.
Why should I train them on things that they already know?
So tailoring it into what's specific for us is much more valuable.
Also, our ambition is to only train on things that they don't know.
So where we have basically challenges.
So that's one part of the equation.
But then also building in that.
into the system prompts or the prompts that you give the AI to have inside of those principles that you have to make sure that they are not violating privacy.
They're not using a tool for an unintended purpose.
Okay, so that's what you meant before when you said, okay, I think you said you are distributing these via MDM, right?
Yes.
So that everyone has this on their device and then you have already a pretty decent base layer of security due to the system prompt that is with every person who is using any of these LLMs, right?
Well, they're not bulletproof.
Yes, for sure.
But with a layered approach.
They don't need to be, because it's kind of like too many layers.
There's an inertia to kind of get through, to breach it.
But if we're talking about malicious actors, because this is kind of the problem of the security industry in general, that everyone builds for the edge cases.
So if you always build for the edge cases, then...
you don't have enough time to actually build for the 80% cases.
And that's usually where it fails.
And then foundational security aspects of like, okay, make sure that everything is patched and the latest versions is running on people's computers.
Everyone is using MFA and preferably phishing proof MFA or not phishing proof, but at least phishing resilient MFAs.
Those are things that raise the bar quite significantly.
And as long as you're not the slowest-moving animal in the herd, then the attackers also can probably take someone else before you.
And I don't think that's luck.
That's more of like, okay, try to kind of stay two steps ahead.
Yeah, that's it.
Great statement that I will definitely take away.
Don't try to always focus on the edge cases, but try to get the 80% right.
Because those are the more important.
Well, the engineers in us always want to make perfect solution.
And then we want to make elegant solutions.
And those two contradict.
And then we end up in this stall that, oh, I want to make a beautiful...
graceful solution.
And then I need to address the edge cases.
And then it gets slower to kind of deliver.
Whilst delivering something fast is actually very much more valuable most of the time these days.
Even if it isn't perfect because you learn from it, then you did something.
Absolutely.
Also, in the security area, it's better to have something quickly that gives certain levels of protection as opposed to having something close to perfect.
in the long term.
So, definitely.
Alright, thanks a lot.
I think we...
Oh, André, do you have a final question maybe in this segment?
Yeah, I have two things.
First of all, my note, my reminder on patterns compared to past larger changes.
And then second, I'm wondering, so I absolutely get your both view here.
Just wanted to challenge that.
In a world where agents, I think you mentioned that Marc, are tireless, so they can just work, I'm wondering if 80% is good enough.
So even though I agree absolutely that it's better to have something fast than aiming for the 100%, I'm just wondering if the game has changed in a way that...
You need to really invest, invest, invest to get to the 100% because on the other side, there's someone who's like trying, trying, trying tirelessly.
And I think that has changed in actually in the last, or this has changed with the capabilities we now have.
Yes, I understand what you're saying.
I'm definitely not saying that you shouldn't have the ambitions.
I'm a 110% ambition person.
By role, by the way.
Sorry?
By role.
It's your role.
It needs to be like that.
I'm always aiming very, very high.
But I'm also going to be happy when I hit 80%.
Because if you don't aim too high, then you're not going to...
You're not going to come, if you just aim for 80%, you're not going to get to 60%.
But I think most people, and now I'm very much generalizing, basically suffer from the fear of not making it good enough so that they take time to actually roll it out.
And if you're considering that you're working Security people are there to kind of catch.
They are thinking in all of the extreme edge cases.
They are thinking about all of the risks.
We are used to picking things apart and seeing the problem with them.
Hence, it's very difficult for us to kind of deliver something, not calling it half-ass, but at least calling it something which is a first step.
und dann auf dem Aufbau von dem, weil wir Angst haben, dass es nicht gut genug ist.
Engineers usually sagen, aber wir werden nicht mehr able zu re-visit das und aufbauern.
Ich denke, das ist die erste Falle, dass wir uns quenchen.
Das ist, dass du, dass du das erste Ding hast, und dann hast du aufbauern.
Denn wenn du aufbauernst, wenn du aufbauernst, dann hast du aufbauernst.
Denn wenn du aufbauernst, wenn du aufbauernst, dann hast du aufbauernst.
You will never come back and revisit it because you always do this.
You always do this pattern.
But if you always add one more package to it, one more thing, one more step, then eventually it creates that virtuous circle where it feeds back on itself and it improves continuously.
It's kind of like the agile manifesto in a sense.
It's just translated into the security space.
Oder zu sagen, es ist die Anwesendung der guten, richtig?
Aber du solltest immer für die Anwesendung für die Anwesendung.
Die Anwesendung der meisten ist immer die Anwesendung der Arten, die Menschen, die eine riesige Anwesendung der Anwesendung der Anwesendung der Anwesendung sehr gut.
Und ich sehe meine Arbeit als das, als ich auch.
Wenn ich etwas, wie ein Dokument oder ein Dokument, I enjoy making it well, but I can't make it to the price that I don't deliver it on time.
And when I'm talking about security, time is your enemy because the longer you are waiting, the larger is the risk or that something will occur.
So you have to continue to deliver.
And hence this Don't try to be perfect from the first rollout.
Now let's try to spend a couple of minutes on the patterns, right André?
Yeah, you seem to be very relaxed, Marc, while other people are not screaming, right?
They're on fire, right?
Let's phrase it positively.
They're on fire with the new capability.
Oh, I meant hair on fire, but hair on fire is much better, yeah.
Ja, es kann sich beide.
Ich denke, dass alle sehr viele Erinnerungen sind.
Sie haben viele verschiedene Technologien gesehen.
Was ist Ihre Meinung hier?
Are wir in etwas sehr, sehr, sehr?
Oder Sie sehen patterns, die Sie vielleicht auch share mit uns, die Sie sehen, so dass Sie ein bisschen Sicherheit haben, wie klar?
Ja, ich denke, dass eine der meisten interessantee Patternsen, dass wir nicht viel über sprechen, ist die demokratische Entwicklung.
Remember, in der Zeit, LAMP kam, wie man mit sehr limiteden skill sets konnte, man musste eine engineering degree in der Art of zu code.
So they started building code, which was, well, less than perfect.
By the way, we need to introduce LAMP.
I think not like, tell me you are old without telling me you are old.
I'm not sure if everyone is aware of LAMP.
So you want to introduce that shortly?
So basically the Linux stack with then...
Having MySQL, PHP on top of it, which was basically what built Facebook.
And many of these companies in the early IT or early .com stages were built on these stacks.
And the Apache Web Server.
Apache Web Server, yeah, that was the A there.
Linux, Apache, MySQL, PHP.
Those things put together, that was really a bad combination, especially in the hands of someone who wasn't proficient.
Now you're building code with AI, and you haven't even seen the code.
You don't even know that it's there.
I built a product.
I made a website.
It's amazing.
And you're so enthusiastic and you're happy.
It's the same thing.
We're repeating the same patterns.
It took us almost 20 years to perfect the CI-CD pipeline post-LAMP platform.
How do you build and create that continuous delivery?
That took a lot of time.
Now most of that paradigm is being upset.
und schicken über, weil jetzt plötzlich haben wir AI-Coding.
Natürlich, wir sind auf dem Stack, das wir bereits haben, aber wir sind jetzt noch mehr Erlebnisse.
In der Zeit war es, wir sollten sich die Sicherheit für die Engenheuerung machen.
Wie viele von den Mandatoren-Security-Trainings haben Sie in Ihrer Karriere gesehen?
Wie viele hat Ihre AI durchgeführt?
Ja.
Ich bin nicht sicher.
Und da kommt der Challenge.
Du musst jetzt noch neue Toolkits zu diesem.
Und du musst einfach auf das.
Okay, wir müssen kind of keep auf.
Wir müssen eine PACE hier.
Wir müssen schnell.
Aber es ist ein Problem, dass es ein Pattern hat und, well, knock on wood.
The world didn't go under.
But definitely there's good reason for people to be nervous, but having your hair on fire doesn't mean you don't act.
Part of my job used to be helping companies who had data breaches.
I was part of a company which was a managed security service provider.
We started in 2001 in the backlash of the IT.
Und selling security managed security services wasn't really a thing.
And here came the LAMP stack and then came a lot of companies who built technologies and they didn't even have properly configured firewalls or detection capabilities and many of these other things that we today call are taking for granted.
So, of course, they had breaches and issues.
Going around to companies and helping them through that crisis makes you kind of have a very calm sense about this and thinking around, okay, doesn't mean I don't take it more or less seriously.
I just know that there's no reason for me to panic and shout wolf because that never helped.
And poor Seesows who've gone off and shouted wolf to the...
to the board of directors or the management team, they all know where that ended up.
They might have gotten money once, but then second time.
And in today, if they go off to the board and say, AI is very dangerous, there's a lot of vulnerabilities with it, or there's much security issues with it, then what will you win?
Because at the same time, the board is seeing, oh, I need to make my company competitive.
I need to innovate.
And here comes a security person telling me it's all dangerous.
What does that help me?
Nothing.
So you need to come with solutions to that problem.
You need to keep a calm head and then break down the problem and come with solutions.
And then they will listen.
At least the same ones.
Thanks for that perspective.
All right.
With that, I would say it's about time to go to the final segments of our podcast episode, right?
So number one, the reality check.
So do you experience what the fuck moments or also wow moments recently with AI?
Yes.
The stuff that I didn't anticipate, I can basically every time that it improves.
Es gibt mir die W-de-fuck-moment all die Zeit.
Es gibt es auf dem Gibt es und es gibt es auf dem Improving.
Und dann ist es immer noch, wenn es es going to...
Vielleicht ist es einfach nur ein Blattens-Out.
Und dann ist es einfach nur die Blattens-Out.
Und dann ist es auch die Bubble, natürlich, die AI-Bubble.
Aber wenn es sich jetzt flattens-Out ist?
Und wenn es sich jetzt auch ein Blattens-Out flattens, dann haben wir uns an.
two or three years worth of backlog to actually just adopt to become more AI-driven or more capable.
But I love being surprised.
Connecting up old hardware that I have at home.
I used to do a lot of hardware hacking back in the day.
And I have these seasoned interfaces and stuff like that with serial ports that I can connect in smart cards and stuff.
And the stuff that I can do with AI now connected to that is fascinating because it does a bunch of stuff that I never imagined.
Imagine if I had that gear like 20 years earlier.
I definitely would have had much fun.
Yes, absolutely.
Last but not least, do you maybe have a prediction for our listeners?
I think that the companies that are learning how to collect institutional knowledge, they are probably the biggest winners.
There's a couple of things with institutional knowledge of learning how an organization does something and then building it into a system.
und zu beenden, dass es sehr schwierig zu transferieren.
Du kannst es nicht wirklich aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- und aus- Today, quite easily.
But imagine that when you can accumulate a lot of this knowledge and a lot of these data sets, that will give you an enormous lock-in, an enormous power.
And that's what you're seeing all of these companies also trying to do, that they are really trying to get the user to live in their ecosystem because they are learning from that.
So, I don't think it's a very revolutionary prediction, but I think it's a very important one.
And then, well, not really a prediction, but more of a recent hindsight.
Look at all of the vulnerabilities that have been found very recently, even if half of them are false or not actually real.
There's a lot of Interest in weaponizing stuff.
We need to be very quick about building new things.
So the reactive security systems is going to be the biggest future, really.
And we have to rethink how that works.
That was actually my point earlier when I said maybe 80% is not good enough.
Ich denke, wir können alle sagen, es ist ein guter Startpunkt, aber wir müssen mehr in diesem neuen Welt.
Und ich glaube, dass wir, als Menschen, nicht mehr auf die Pace können.
Wir müssen dort sitzen und wir müssen auf die Agents und AI arbeiten mit den Unterlagen und den Unterlagen und den Unterlagen.
Wir müssen nicht genug sein.
Absolutely.
Es kann uns, es kann auch uns, es kann auch die Agents, auf beide Seiten.
Vielen Dank.
Es war sehr gut zu sprechen.
Vielleicht müssen wir auch noch ein paar Jahre später update und dann sehen wir, ob wir noch mehr als Faktor haben, die wir heute diskutieren.
Vielleicht wird es auch noch ein bisschen zu sprechen.
Aber für heute...
Vielen Dank für die Show.
Und danke für mich.
Bye-bye.