Latest / Expert Insights Podcast / #60 – The Next Evolution of AI in the SOC (Nicole Carignan, Darktrace)
Transcript
- 0:04SPEAKER_00: Hello and welcome to the Expert Insights podcast.
- 0:13SPEAKER_00: I'm Caitlin Harris,
- 0:14SPEAKER_00: Deputy Head of Content at Expert Insights and your host for this episode,
- 0:17SPEAKER_00: which we're recording live at the RSAC 2025 conference.
- 0:21SPEAKER_00: I'm joined today by Nicole Carrigan,
- 0:24SPEAKER_00: SVP for Security and AI Strategy and Field CISO at Darktrace,
- 0:28to discuss how AI is evolving from a support tool to a key
- 0:32decision-making partner
- 0:33SPEAKER_00: and what that means for the future role of the analyst.
- 0:36SPEAKER_00: Nicole, thanks for joining us.
- 0:38SPEAKER_00: First off,
- 0:39I want to talk about your cyber AI analyst and the new
- 0:42enhancements that you have
- 0:43SPEAKER_00: been making to that.
- 0:45Can you share what's new and how these updates are going to
- 0:47help improve the way
- 0:48SPEAKER_00: security teams are detecting and responding to threats?
- 0:51SPEAKER_01: Oh, that's great.
- 0:52SPEAKER_01: So our cyber AI analyst is our investigative AI.
- 0:55And what it's able to do is it actually has a layers of machine
- 0:59learning underneath.
- 1:00SPEAKER_01: So think agentic system,
- 1:02but not large language model based or generating synthetic data
- 1:06or focusing only on
- 1:08SPEAKER_01: semantic analysis.
- 1:09SPEAKER_01: We do have semantic analysis portions of it,
- 1:11but we have a lot of different types of machine learning
- 1:13techniques that are finite
- 1:14SPEAKER_01: and focused on optimizing accuracy of that specific use case of analysis.
- 1:19SPEAKER_01: Fiber AI Analyst takes that SOC Level 1 and SOC Level 2 triaging.
- 1:23SPEAKER_01: It takes in the alerts,
- 1:25SPEAKER_01: the events that are high anomalous and high risky behavior,
- 1:29SPEAKER_01: and it starts to set hypotheses based on what it thinks it might be.
- 1:33Like,
- 1:34does that specific external communication look like command and
- 1:37control?
- 1:37SPEAKER_01: Or exfil, data exfiltration?
- 1:39SPEAKER_01: Or just data backup?
- 1:41SPEAKER_01: and transfer.
- 1:41And it analyzes the underlying logs to see if it can confirm,
- 1:45refute,
- 1:45or modify that hypothesis.
- 1:47SPEAKER_01: And it does this iteratively when it comes to a conclusion.
- 1:50SPEAKER_01: If it confirms that it gets escalated to a critical incident,
- 1:54SPEAKER_01: prioritizing what humans should be looking at.
- 1:56SPEAKER_01: So think using advanced behavior analytics.
- 1:59If you think like the definition of an agentic system where
- 2:02you've got a multitude
- 2:04of different machine learning techniques that are working
- 2:06collaboratively to
- 2:07SPEAKER_01: ingest,
- 2:08SPEAKER_01: perceive,
- 2:09SPEAKER_01: analyze,
- 2:10SPEAKER_01: plan,
- 2:11SPEAKER_01: decision making,
- 2:12SPEAKER_01: and then autonomously taking action.
- 2:13But really,
- 2:14from our perspective,
- 2:15we don't want to trust the output of one model.
- 2:17We want corroborating sources that help build our confidence
- 2:20elicitation or of the
- 2:22SPEAKER_01: accuracy of the output of that concept.
- 2:24SPEAKER_01: conclusion.
- 2:25So we look at it where we layer it with a whole bunch of
- 2:28different machine learning
- 2:29SPEAKER_01: techniques that can work collaboratively and critically with each other,
- 2:32SPEAKER_01: ultimately come to the end of the conclusion of the analysis.
- 2:35SPEAKER_01: The next step then is plan to take autonomous action.
- 2:37And in our particular case,
- 2:39we want to take a very surgical response that's very finite and
- 2:43SPEAKER_01: and non-disrupted to a business.
- 2:45SPEAKER_01: So if I were to do that external data communication,
- 2:49SPEAKER_01: but I normally don't go to XYZ legitimate file sharing sites at 11 p.m.
- 2:54SPEAKER_01: at night, it's trying to upload three gigs of data.
- 2:57SPEAKER_01: It'll allow me to perform my normal pattern of life.
- 3:00SPEAKER_01: It will only sever that anomalous or high risky behavior.
- 3:04helping to facilitate the protection against use cases like
- 3:07insider threat or data
- 3:08SPEAKER_01: loss prevention.
- 3:09SPEAKER_01: So that's the premise of Cyber AI Analyst.
- 3:11SPEAKER_01: So the two cool new innovations, which I get very excited about.
- 3:14SPEAKER_01: So Digest,
- 3:15it's a graph neural network that's been trained on graphs of
- 3:19incidents of Cyber AI
- 3:20SPEAKER_01: Analyst detections over the last several years.
- 3:22What this allows us to do is it allows us to look at the early
- 3:26stages of an attack
- 3:27and then apply predictive modeling to see which ones are most
- 3:31likely to explode,
- 3:33SPEAKER_01: to cause a bigger impact,
- 3:35SPEAKER_01: to have a more critical outcome.
- 3:38And it allows us to,
- 3:39again,
- 3:40prioritize where humans should be looking at first.
- 3:43SPEAKER_01: As in my former world, I'm former U.S.
- 3:45SPEAKER_01: government working in cyber operations.
- 3:47SPEAKER_01: We know we needed advanced behavioral analytics.
- 3:50SPEAKER_01: We know we needed it at the network layer or the full domain level.
- 3:53SPEAKER_01: We needed that visibility.
- 3:54We know we needed anomaly detection to facilitate advanced
- 3:57threat detection.
- 3:58SPEAKER_01: But that just gives you something that's anomalous.
- 4:01How do you focus the humans to what they need to prioritize to
- 4:05investigate quickly?
- 4:06SPEAKER_01: So AI should really be that supplemental trickle.
- 4:08triaging of the alerts so that the humans can only focus on the
- 4:12stuff that needs
- 4:13SPEAKER_01: really in-depth investigation.
- 4:15SPEAKER_01: And so Digest does that exact same thing.
- 4:17SPEAKER_01: It's saying based on years of incidents we've seen play out,
- 4:22which of the earlier stages are more likely to be a critical
- 4:25incident to focus
- 4:26SPEAKER_01: humans on that first?
- 4:27SPEAKER_01: Then the other one is DMIST.
- 4:30And this is just really cool because we do need to have
- 4:32semantic analysis.
- 4:34SPEAKER_01: Large language models are great at that.
- 4:35SPEAKER_01: We don't use it in a generative capacity,
- 4:37SPEAKER_01: but we want to have semantic analysis to look at file name structure,
- 4:40SPEAKER_01: host name structure,
- 4:41SPEAKER_01: anomalies or patterns of life in that type of analysis,
- 4:45SPEAKER_01: as well as anomalies to help to facilitate more finite,
- 4:48SPEAKER_01: accurate security detection.
- 4:49SPEAKER_01: But large language models tend to be very computationally extensive.
- 4:53SPEAKER_01: So our DMS2 is a LoRa LLM,
- 4:56SPEAKER_01: so think more of a local language model,
- 4:58SPEAKER_01: trained on a much more finite parameter,
- 5:02SPEAKER_01: specifically to security use cases,
- 5:04so that we can drive accuracy for the security analysis use
- 5:07cases we're looking
- 5:08SPEAKER_01: for,
- 5:08SPEAKER_01: but we only have to run on CPU and not GPU,
- 5:11SPEAKER_01: and we can run locally.
- 5:12So you don't have to suck your data out into a big,
- 5:14large language model classifier.
- 5:15SPEAKER_01: You don't have to worry about data privacy or data residency.
- 5:18SPEAKER_01: It can all be run locally and it can perform the same level of security
- 5:21identification and classification tasks that you would need to
- 5:25have a much more
- 5:26SPEAKER_01: advanced analytics,
- 5:27but without all the resources that are needed in order to
- 5:29accomplish that.
- 5:30SPEAKER_00: Absolutely.
- 5:30It sounds like you've got some really exciting things happening
- 5:33at the moment.
- 5:33SPEAKER_00: I'm very excited to follow as you continue to innovate as well.
- 5:36SPEAKER_00: We've talked a little bit about the evolution of your analyst tool,
- 5:40but this sort of evolution from agentic AI as being a support
- 5:44tool to a real
- 5:45decision-making partner is something that I think we're hearing
- 5:49quite a lot about
- 5:49SPEAKER_00: at the moment.
- 5:50Could you talk to us a little bit more about that evolution
- 5:53across the industry as
- 5:54SPEAKER_00: a whole and what that might mean for the future of the SOC analyst role?
- 5:58SPEAKER_00: I love that.
- 5:59SPEAKER_00: I'm going to start with the why, the need.
- 6:01SPEAKER_01: So over the last year,
- 6:03we've seen edge security devices getting hammered with
- 6:07vulnerability disclosures.
- 6:08SPEAKER_01: And over the last year,
- 6:09Cyber AI Analyst and our threat research team has also started
- 6:12to see a trend where
- 6:14SPEAKER_01: we are seeing anomalous,
- 6:16high-risky behavior attempts to exploit a vulnerability that
- 6:19has not been disclosed
- 6:21SPEAKER_01: yet.
- 6:21SPEAKER_01: So definition of a zero day, critical vulnerability is released.
- 6:25SPEAKER_01: Defenders race to patch it.
- 6:26SPEAKER_01: Threat actors race to exploit it.
- 6:28SPEAKER_01: Day zero starts the race.
- 6:30We're actually saying that it starts T minus 11 to 17 days
- 6:33prior to the disclosure of that CDE.
- 6:36And we saw this over multiple different product lines over the
- 6:40course of the last year.
- 6:41SPEAKER_01: We did finally have a customer who had autonomous response in place,
- 6:45SPEAKER_01: which meant to contain the incident,
- 6:47to mitigate the risk of the exploitation of a vulnerability
- 6:50that they didn't even
- 6:51SPEAKER_01: know they had.
- 6:52And so in this particular case,
- 6:53the vulnerability was disclosed in mid-January.
- 6:56SPEAKER_01: Our customer,
- 6:56SPEAKER_01: December 30th of 2024,
- 6:57SPEAKER_01: started to see anomalous activity on their edge security device.
- 7:01SPEAKER_01: I won't name the product line.
- 7:04SPEAKER_01: And,
- 7:04SPEAKER_01: you know,
- 7:05SPEAKER_01: connections to a rare IP endpoint,
- 7:07SPEAKER_01: no host name associated with it,
- 7:09SPEAKER_01: trying to download a script of some sort,
- 7:11SPEAKER_01: trying to exploit something that we don't know is actual vulnerability.
- 7:14It started acting so anomalously and risky behavior that we
- 7:19took the simple
- 7:21SPEAKER_01: surgical response of sever the connection to the rare endpoint.
- 7:24SPEAKER_01: But it kept trying to do things.
- 7:26And so we actually escalated our action to enforce the pattern
- 7:31of life of that gateway.
- 7:33SPEAKER_01: Don't allow it to do anything it hasn't done before.
- 7:35Allow Caitlin or Abigail to connect in and access the resources
- 7:40that they normally do.
- 7:41SPEAKER_01: But don't go to any rare endpoints that you've never connected to before.
- 7:45And so it started enforcing the pattern of life of that edge
- 7:48security gateway,
- 7:48SPEAKER_01: which is actually the first line of defense for most organizations.
- 7:51SPEAKER_01: Fast forward 11 days later, a vulnerability was disclosed.
- 7:54SPEAKER_01: And that was what this threat actor was trying to exploit.
- 7:56And so we're starting to see the need for autonomous action
- 8:00more and more critical
- 8:01SPEAKER_01: because that isn't a matter of human speed.
- 8:03SPEAKER_01: Humans didn't even know they had a problem on that device.
- 8:06They wouldn't know that there was a problem on that device for
- 8:09another 11 days.
- 8:10SPEAKER_01: So it's not even just a machine speed scale issue.
- 8:13SPEAKER_01: It's not a human versus AI issue.
- 8:15SPEAKER_01: It is...
- 8:17the need for AI to be able to perform such advanced behavior
- 8:21analytics and an
- 8:22understanding of the environment and the assets in that
- 8:24environment to be able to
- 8:25SPEAKER_01: defend against things that we don't even know exist.
- 8:27SPEAKER_01: That's the need.
- 8:28SPEAKER_01: How do we get to autonomous action?
- 8:30SPEAKER_01: And so from the AI space, very similar to the cybersecurity space,
- 8:33SPEAKER_01: We like to throw away around terms.
- 8:35SPEAKER_01: We like to change terminology at any given time.
- 8:38SPEAKER_01: Before there was agentic AI, it was just how do you achieve autonomy?
- 8:42SPEAKER_01: Agentic AI is a vehicle to do that.
- 8:44SPEAKER_01: It's,
- 8:44SPEAKER_01: you know,
- 8:45using a collaboration of multiple different models and
- 8:47techniques that work
- 8:48SPEAKER_01: collaboratively and critically in order to plan,
- 8:50SPEAKER_01: perceive,
- 8:51SPEAKER_01: make decisions,
- 8:52SPEAKER_01: and hopefully autonomously take action.
- 8:53SPEAKER_01: To be able to autonomously take action,
- 8:55SPEAKER_01: specifically in security,
- 8:57SPEAKER_01: you have to get the practitioners on board.
- 8:59SPEAKER_01: There has to be trust.
- 9:00SPEAKER_01: That comes with data,
- 9:01SPEAKER_01: transparency,
- 9:02SPEAKER_01: and explainability,
- 9:03helping them to understand how the AI has come to its analysis
- 9:07and decision-making
- 9:08SPEAKER_01: process,
- 9:09SPEAKER_01: educating the stock users on how to operationalize.
- 9:13SPEAKER_01: It's the use of its intelligence.
- 9:14It becomes a much more collaborative relationship,
- 9:16but it has to be built on trust.
- 9:18SPEAKER_01: And so then you talk about data residency, data privacy, apps.
- 9:21SPEAKER_01: Accuracy,
- 9:22all the core tenets of data science really kind of come into
- 9:25play to be able to
- 9:25SPEAKER_01: build out that trust to achieve autonomy.
- 9:28SPEAKER_01: And then you get to the autonomous action.
- 9:30And what we have found is you have to take the most surgical
- 9:33approach as possible
- 9:34SPEAKER_01: because AI is not going to get it 100% right every time.
- 9:37SPEAKER_01: But if it's not a really disruptive event...
- 9:40SPEAKER_01: and it's really protecting and potentially containing incidents,
- 9:43SPEAKER_01: then that autonomous action is accepted and trusted.
- 9:46SPEAKER_01: Same with a human.
- 9:47SPEAKER_01: Analyst isn't going to get it right 100% of the time.
- 9:49SPEAKER_01: And honestly,
- 9:50SPEAKER_01: most of the time,
- 9:50I could actually rely a little bit more on the output of
- 9:53mathematical models than I
- 9:54SPEAKER_01: can in my own brain.
- 9:55So it's really understanding that it's not going to be right
- 9:58100% of the time.
- 9:59SPEAKER_01: But if you're taking more finite actions that will be disruptive,
- 10:03that actually maybe potentially contain a much more serious
- 10:07incident and reduce the
- 10:08SPEAKER_01: risk,
- 10:08SPEAKER_01: the potential damage of an incident,
- 10:10SPEAKER_00: it's worth that.
- 10:11SPEAKER_00: Absolutely.
- 10:12And you've spoken a little bit there around the sort of like
- 10:14risks and rewards of
- 10:15SPEAKER_00: deploying AI in the SOC.
- 10:17How do you recommend that organizations approach it responsibly
- 10:21so they're kind of
- 10:22SPEAKER_00: balancing both sides of that?
- 10:24SPEAKER_01: That's a great, I mean, great question.
- 10:26SPEAKER_01: And I think that's really what we should lead with SOX to start with.
- 10:29SPEAKER_01: Start with questions.
- 10:30SPEAKER_01: What is the AI that powers the system that they're adopting?
- 10:33SPEAKER_01: What are the machine learning techniques?
- 10:35SPEAKER_01: How do you effectively evaluate accuracy?
- 10:39How do you effectively train it,
- 10:42learn reinforcement learning or fine tune it or control it?
- 10:46That's a big component of security is there has to be a great
- 10:49vehicle for control
- 10:50SPEAKER_01: for humans to ensure that it's being operationalized well.
- 10:54And there's a lot of mitigations that comes with different
- 10:56types of machine learning techniques.
- 10:58SPEAKER_01: And we can pick a little bit on generative AI,
- 11:01specifically large language model based agents in a generative
- 11:04capacity a little
- 11:05SPEAKER_01: bit.
- 11:05SPEAKER_01: There is a data concern.
- 11:08SPEAKER_01: You have to talk about data security,
- 11:09SPEAKER_01: data governance,
- 11:10SPEAKER_01: data classification,
- 11:11ensuring that the large language models have access to the data
- 11:14sets that it should
- 11:15SPEAKER_01: have access to,
- 11:16SPEAKER_01: controlling access permissions,
- 11:18SPEAKER_01: making sure it's not too permissive,
- 11:20looking at clear communication guardrails because it can become
- 11:23self-discoverable
- 11:24SPEAKER_01: and start communicating with agents that might not have,
- 11:26SPEAKER_01: should have access to the type of data that it's working with,
- 11:29SPEAKER_01: etc.
- 11:30SPEAKER_01: So putting in those hard guardrails.
- 11:38So knowing the types of machine learning models or the
- 11:41interfaces or the API
- 11:42connections allows you to really mitigate a lot of the risk of
- 11:46actually deploying a
- 11:47SPEAKER_01: solution,
- 11:48but also understanding that one size doesn't fit all and won't
- 11:52get you to the
- 11:52SPEAKER_01: accuracy you're hoping.
- 11:54SPEAKER_01: And I loved our success.
- 11:55SPEAKER_01: CEO of Dark Trace Federal told me right before this,
- 11:58SPEAKER_01: he was like,
- 11:58SPEAKER_01: we can talk about compliance all day long,
- 12:00SPEAKER_01: but if I don't convince the practitioner that it trusts its output,
- 12:04SPEAKER_01: compliance doesn't even matter.
- 12:05Like really the first gate there is how do you build trust with
- 12:09the people who are
- 12:09SPEAKER_01: actually using the output of these models?
- 12:12SPEAKER_01: And it has to come with, they have to start to understand.
- 12:14SPEAKER_01: So that cross-skilling into machine learning and AI,
- 12:17SPEAKER_01: they have to understand how it's getting there.
- 12:19They have to be able to have the transparency and the
- 12:23explainability to understand what
- 12:26SPEAKER_01: how it came to its conclusion so it can validate or mitigate,
- 12:30SPEAKER_01: essentially,
- 12:31SPEAKER_01: and control the workflow process of how do they use that intelligence.
- 12:36SPEAKER_00: Absolutely.
- 12:36SPEAKER_00: And in a similar vein,
- 12:38SPEAKER_00: I suppose,
- 12:38how can security leaders set realistic expectations around what
- 12:43Agenda KI can and
- 12:44SPEAKER_00: can't do today?
- 12:45SPEAKER_00: I mean,
- 12:45SPEAKER_00: we're hearing a lot of really interesting real world use cases,
- 12:49SPEAKER_00: but also theoretical applications for it on the show floor today.
- 12:51SPEAKER_00: And it's quite exciting and in some cases a little overwhelming.
- 12:54SPEAKER_00: So how
- 12:55SPEAKER_00: How can we set those expectations?
- 12:57SPEAKER_01: I think that's great.
- 12:58SPEAKER_01: I mean, is this a use case they're trying to solve?
- 13:01SPEAKER_01: And let's talk about large language model agent-based agentic system.
- 13:04SPEAKER_01: Would it thrive or be accurate in a semantic analysis capacity?
- 13:07SPEAKER_01: So I went to an AI conference last July,
- 13:10and one of the really interesting use cases I saw where it was
- 13:14semantic analysis.
- 13:15And they were looking at equitable data distribution to
- 13:19disenfranchised communities
- 13:21SPEAKER_01: for healthcare and wealth management purposes.
- 13:24SPEAKER_01: It's like, what a noble cause.
- 13:26But then you have to get down to securing the AI is actually
- 13:29really securing the
- 13:31SPEAKER_01: data that you're talking about in data distribution.
- 13:33SPEAKER_01: And a lot of that then comes down to data security and permissions and
- 13:38SPEAKER_01: communication guardrails and API security.
- 13:41So a lot of the things that are around the model,
- 13:43but not necessarily just the model themselves.
- 13:45I think a lot of security leaders kind of get focused on how do
- 13:48I secure AI?
- 13:48SPEAKER_01: I'm like, you've got to start with data.
- 13:50And then you have to start applying the same foundational
- 13:53cybersecurity principles
- 13:54SPEAKER_01: that you would to any other asset in your environment.
- 13:56SPEAKER_01: And I think that's kind of a first step to start with.
- 13:59SPEAKER_01: But in many cases,
- 14:00SPEAKER_01: and this is the data scientist in me,
- 14:01semantic analysis may not be your best case to get to that
- 14:06accuracy that you're
- 14:07SPEAKER_01: hoping for.
- 14:08Creating or generating new synthetic data that's based on a
- 14:12data corpus that you
- 14:13had nothing to do with it might also not fulfill the use case
- 14:17that you're trying to
- 14:17SPEAKER_01: achieve very accurately.
- 14:19SPEAKER_01: And that's just going to erode trust with the practitioner.
- 14:22SPEAKER_01: You give me junk, I'm not going to trust you.
- 14:25SPEAKER_01: I'm not going to use you.
- 14:26SPEAKER_01: I'm not going to operationalize you.
- 14:28And now you've just introduced potentially risk into your
- 14:31environment with no gain
- 14:32SPEAKER_01: of productivity.
- 14:32And so it's thinking through that that I think security leaders
- 14:37have to get in the mindset of.
- 14:38SPEAKER_01: And I also think similar to the journey with cloud, right?
- 14:41SPEAKER_01: We took off into cloud well before they brought security along with it.
- 14:44SPEAKER_01: We're taking off into AI, right?
- 14:46SPEAKER_01: Security and AI should be one.
- 14:48SPEAKER_01: Developers that are building the models,
- 14:50SPEAKER_01: that are building the Agentex systems should be embedded with security,
- 14:54SPEAKER_01: security embedded with them.
- 14:55SPEAKER_01: It should be secure by design.
- 14:57I love the guidance that's coming out of NCSC and CISA and also
- 15:00NIST AI risk
- 15:01SPEAKER_01: management framework.
- 15:02I mean,
- 15:03also now we're almost completing our ISO 42001 certification AI
- 15:08management system.
- 15:09So really thinking through all of the mitigations that you
- 15:12would think from a
- 15:12SPEAKER_01: security at heart system
- 15:14SPEAKER_01: system,
- 15:15but also being able to innovate and operationalize something
- 15:18that's really going to
- 15:18SPEAKER_01: be a game changer to achieving autonomy.
- 15:22SPEAKER_00: Absolutely.
- 15:22SPEAKER_00: Just on that note,
- 15:23SPEAKER_00: I mean,
- 15:24you mentioned a few of the different guidelines that are out
- 15:26there that are really
- 15:27SPEAKER_00: helpful and sort of innovating,
- 15:29SPEAKER_00: but securely.
- 15:30SPEAKER_00: Do you think that we also need any kind of regulation around that?
- 15:34Or do you think that this is something that the cybersecurity
- 15:36industry needs to
- 15:37SPEAKER_00: kind of realize for itself without that external sort of government in
- 15:41SPEAKER_01: I mean, obviously, we're a global company.
- 15:44SPEAKER_01: We started in the heart of the U.K.
- 15:46SPEAKER_01: and Cambridge.
- 15:47So I'm a firm believer of ensuring that we meet the regulatory
- 15:50requirements for any
- 15:52SPEAKER_01: organization or any country or environment that we work in.
- 15:54SPEAKER_01: EU, U.K., German Workers' Council, as well as the U.S.
- 15:59SPEAKER_01: So I think that there is an element where all of that is,
- 16:02SPEAKER_01: I think,
- 16:03SPEAKER_01: very valid advising guidance that companies need to be followed.
- 16:07SPEAKER_01: So I'm going to take a step back.
- 16:08SPEAKER_01: I've got two answers, and I apologize.
- 16:10SPEAKER_01: Thank you.
- 16:10SPEAKER_01: No, I feel not consistent.
- 16:13SPEAKER_01: And the data scientist in me says,
- 16:16SPEAKER_01: yes,
- 16:16SPEAKER_01: there needs to be firm guidance to guide people to do this responsibly.
- 16:21SPEAKER_01: Because I am not at all scared about AI.
- 16:24SPEAKER_01: I mean,
- 16:24SPEAKER_01: literally getting my computer science degree over 25 years ago,
- 16:27SPEAKER_01: I went down the AI and hardware language section of the computer science
- 16:31SPEAKER_01: engineering course.
- 16:32SPEAKER_01: So it's not new.
- 16:33SPEAKER_01: We've been using it for a long time and I'm not at all scared about it.
- 16:36But I've also been ingrained with good foundational principles
- 16:39of data science.
- 16:40Like how do I do this responsibly,
- 16:42ethically,
- 16:43can do this accurately,
- 16:44effectively.
- 16:45SPEAKER_01: Like really,
- 16:46I look at it as automating a lot of the things so I don't have
- 16:48to do the same task
- 16:49SPEAKER_01: more than five times.
- 16:50I can get a machine learning model to replicate what I do so I
- 16:54can start thinking
- 16:54SPEAKER_01: more strategically and move on to like more harder and complex problems.
- 16:58SPEAKER_01: And I feel like that's like the game changer of AI.
- 17:01SPEAKER_01: But then as Arnold Schwarzenegger was asked on the Kelsey podcast,
- 17:05SPEAKER_01: I think he said this also at a Gartner conference later,
- 17:08SPEAKER_01: but he said it on the Kelsey podcast first.
- 17:10SPEAKER_01: They asked him if he was scared of AI.
- 17:11SPEAKER_01: And he was like, I'm not scared of AI.
- 17:12SPEAKER_01: I'm scared of stupid people.
- 17:13SPEAKER_01: As a data scientist, I am scared of stupid people using AI.
- 17:18SPEAKER_01: I am scared of stupid people innovating with AI,
- 17:20SPEAKER_01: not thinking through the ethical and security implications.
- 17:23SPEAKER_01: And that is one of my biggest fears.
- 17:25SPEAKER_01: Now I'm going to jump onto the other side that I said, totally opposite.
- 17:28SPEAKER_01: As an innovator, you have to be able to run fast.
- 17:31SPEAKER_01: And I know that that as the U.S.
- 17:33SPEAKER_01: has a very different perspective from a lot of other countries globally.
- 17:37SPEAKER_01: They don't want to stifle innovation.
- 17:39SPEAKER_01: And I totally understand that because I think with this great innovation,
- 17:43SPEAKER_01: we can achieve some really cool,
- 17:45SPEAKER_01: almost miraculous things.
- 17:46I think about implications of cures to disease that could come
- 17:50out of this.
- 17:50SPEAKER_01: Yes.
- 17:51SPEAKER_01: Innovation.
- 17:51SPEAKER_01: Yes.
- 17:52SPEAKER_01: I would love to see those two different hats kind of come together.
- 17:55How can we innovate quickly with good data science principles
- 17:59to do it safely,
- 18:00SPEAKER_01: responsibly,
- 18:01SPEAKER_01: ethically,
- 18:01SPEAKER_01: securely?
- 18:02SPEAKER_01: And I think we can.
- 18:03SPEAKER_01: I think most of us,
- 18:03SPEAKER_01: at least in the who've been in the industry for a while,
- 18:06SPEAKER_01: I don't check security out the window when I'm thinking through this.
- 18:10SPEAKER_01: Obviously, most of my career has been security.
- 18:12SPEAKER_01: So it's a little bit too ingrained in me.
- 18:14I think there is an element of innovate,
- 18:16but innovate also very responsibly.
- 18:19SPEAKER_00: Absolutely.
- 18:20SPEAKER_00: And with all of that in mind,
- 18:22SPEAKER_00: I suppose,
- 18:22do you think that we'll ever see a fully autonomous SOC analyst
- 18:26that is capable of
- 18:27SPEAKER_00: replacing humans,
- 18:28SPEAKER_00: whether or not it should?
- 18:30SPEAKER_00: Do you think that we'll ever see that?
- 18:31SPEAKER_00: And what would it take us to get there if so?
- 18:34SPEAKER_00: Yeah.
- 18:34SPEAKER_01: Yes, I do.
- 18:35SPEAKER_01: And I think we're working our way already there.
- 18:38SPEAKER_01: But I think humans are always still going to be a part of it.
- 18:41SPEAKER_01: And so if you think about right now,
- 18:44we're looking with cyber AI analysts to offload a lot of the
- 18:48SOC level one and SOC
- 18:49SPEAKER_01: level two triaging and also with autonomous analysis,
- 18:53detection and autonomous response in order to buy humans more
- 18:57time to think more
- 18:58SPEAKER_01: strategically for remediation or treatment.
- 19:01stock level three threat hunting or proactive and cyber
- 19:04resiliency tasks that can
- 19:06SPEAKER_01: actually have a bigger impact to the organization and risk reduction.
- 19:10I think humans are a part of that whole process that their
- 19:13roles are going to
- 19:14SPEAKER_01: drastically change.
- 19:15SPEAKER_01: And I think that's really exciting.
- 19:17SPEAKER_01: I know a lot of people have a lot of concerns about that.
- 19:20But I think if you have an infectious attitude of never want to
- 19:24stop learning,
- 19:25you get the opportunity to continue to work with more and more
- 19:28innovative tech that
- 19:29allows you to free up and think of more creative ways to defend
- 19:33and to make your
- 19:34SPEAKER_01: organization more secure,
- 19:36especially as threat actors are innovating with AI and they're
- 19:40not doing it safely
- 19:41SPEAKER_01: responsibly.
- 19:43SPEAKER_01: Ethically or securely.
- 19:45So you're going to have to be able to scale the humans to be
- 19:51able to continue to
- 19:52SPEAKER_01: defend against that type of onslaught of threat.
- 19:55SPEAKER_01: So I do think we're already kind of almost there.
- 19:58I don't think it will ever get to a point where humans are out
- 20:00of the picture whatsoever.
- 20:02SPEAKER_01: I think humans will be key and critical into the operationalism.
- 20:05SPEAKER_01: of AI,
- 20:06SPEAKER_01: key and critical into detection engineering,
- 20:09key and critical to setting up security policies for an
- 20:12organization and using the
- 20:13SPEAKER_01: technology they have in order to enforce those policies.
- 20:16Think data loss prevention and insider risk different to every
- 20:19organization.
- 20:20You can't give just a one size fits all and humans are going to
- 20:23be leading that.
- 20:24SPEAKER_01: And then to take a step back even further,
- 20:26SPEAKER_01: how do I now have AI that can learn my entire environment?
- 20:31and has greater visibility and bandwidth for analytics than I
- 20:36as a single human will ever have.
- 20:38SPEAKER_01: I can use that to think,
- 20:39SPEAKER_01: okay,
- 20:40SPEAKER_01: how could an adversary get from point A to point B in my environment?
- 20:44What are hardening mitigations that I can take that would
- 20:46actually reduce that risk?
- 20:48What are the hardening mitigations that would actually have the
- 20:50most risk reduction
- 20:51SPEAKER_01: within my environment?
- 20:52What are the potential CVEs that could be more impactful in my
- 20:56organization?
- 20:56SPEAKER_01: It doesn't necessarily have to be the ones that have the highest score.
- 21:00It could be how is the rest of my security stack interacting
- 21:03with each other to see
- 21:04what is actually the art of the possible and then how do I
- 21:07mitigate against that?
- 21:08So I think it will drastically be changing what stock analysts
- 21:12used to do.
- 21:13SPEAKER_01: And I find that really exciting.
- 21:15But I know we've got to ensure that stock analysts are getting
- 21:19that cross-skilled
- 21:20and that education and how do they operationalize the use of AI
- 21:24and how do they
- 21:26SPEAKER_01: actually start to think strategically to have a bigger impact,
- 21:29especially when they're hearing things or potentially their
- 21:31leadership on X percent
- 21:33SPEAKER_01: workforce reduction.
- 21:34which is causing a totally different exacerbated insider threat
- 21:40concern with AI
- 21:41SPEAKER_01: adoption,
- 21:42SPEAKER_01: economic instability,
- 21:43SPEAKER_00: job insecurity.
- 21:45SPEAKER_00: Absolutely.
- 21:46SPEAKER_00: So I just have one last question.
- 21:47SPEAKER_00: It's off the back of that.
- 21:48What is one practical step that organizations can take today to
- 21:52prepare for that
- 21:54SPEAKER_00: evolution and for that future of the SOC team?
- 21:56SPEAKER_01: Absolutely.
- 21:57SPEAKER_01: I'm going to take a step back before I get to organizations.
- 22:00SPEAKER_01: I want to say call out to education institutions.
- 22:03SPEAKER_01: Like,
- 22:03AI needs to be embedded in cybersecurity degrees as well as
- 22:08computer science
- 22:09SPEAKER_01: engineering degrees.
- 22:10And data scientists and machine learning engineers need to be
- 22:13taking courses in security.
- 22:14SPEAKER_01: Like,
- 22:15SPEAKER_01: there needs to be a trifecta of education to prepare for this onslaught,
- 22:19SPEAKER_01: which I'm very excited about.
- 22:20SPEAKER_01: All right.
- 22:20SPEAKER_01: Now that's education.
- 22:21SPEAKER_01: Now let's take a step back to the organization.
- 22:23SPEAKER_01: I think obviously you lead from the top and you lead from your example.
- 22:27Organizations shouldn't be leaning into a message of workforce
- 22:32reduction or job instability.
- 22:35They should be encouraging their security practitioners to
- 22:39cross-skill into AI.
- 22:40SPEAKER_01: And I think they're going to actually see a lot more productivity gains.
- 22:45SPEAKER_01: They're going to be able to achieve so many more goals and objectives,
- 22:49SPEAKER_01: not just in security,
- 22:50but definitely in security and like innovations across the
- 22:53organization in whatever
- 22:54SPEAKER_01: field that they're in without having to increase the workforce.
- 22:58SPEAKER_01: So looking,
- 22:58SPEAKER_01: kind of flipping it and not thinking workforce reduction,
- 23:01SPEAKER_01: think workforce optimization and scalability,
- 23:04SPEAKER_01: and you can achieve so much more.
- 23:06SPEAKER_01: I think you're going to entice and excite
- 23:09Like the cybersecurity personnel within organizations to learn
- 23:13new technologies,
- 23:14learn new domain skills that are outside of what they've
- 23:18traditionally done,
- 23:19because now they'll have the ability to have the freedom to
- 23:23kind of continue to
- 23:24SPEAKER_01: learn,
- 23:24SPEAKER_01: sharpen themselves,
- 23:25SPEAKER_01: but also I like think more strategically.
- 23:28SPEAKER_00: Yeah, absolutely.
- 23:29SPEAKER_00: I love that.
- 23:29Such a positive answer,
- 23:31just having that positive mindset and outlook on it.
- 23:33And I think that so many cybersecurity professionals have that
- 23:37innate willingness
- 23:38SPEAKER_00: to learn and that desire to learn and grow.
- 23:41SPEAKER_00: Okay, well, thank you so much for that.
- 23:43SPEAKER_00: I really, really enjoy the conversation.