Latest / The Adults In The AI Room / Earley AI Podcast - Episode 85: AI Security, Shadow IT, and the Governance Reset with Rob Lee

Transcript

Earley AI Podcast - Episode 85: AI Security, Shadow IT, and the Governance Reset with Rob Lee · The Adults In The AI Room

0:01
SPEAKER_00: This is the Early AI Podcast hosted by Seth Earth.
0:04
SPEAKER_00: Join us as we delve deep into the passions, expertise, and experiences of thought leaders and practitioners to talk about what's possible with artificial intelligence.
0:14
SPEAKER_00: The Early AI Podcast is sponsored by Early Information Science, your digital transformation journey with design and deployment of innovative technology solutions as well as FECTR.
0:26
SPEAKER_00: Now enjoy the show.
0:29
SPEAKER_01: Welcome to the Early AI Podcast.
0:31
SPEAKER_01: This is a show where we talk to experts in the industry about what organizations are doing to get AI right, to deploy it correctly, to avoid risks and security challenges, to be able to get ROI and to really avoid the mistakes that many organizations are facing right now.
0:50
SPEAKER_01: Today we're going to be talking with Robert T.
0:53
SPEAKER_01: Lee, who is Chief AI officer and chief of research at SENS.
0:57
SPEAKER_01: Rob's work sits right at the intersection of AI and cybersecurity, where the stakes are really high and the data is certainly messy.
1:04
SPEAKER_01: The temptation is to move faster than the guardrails.
1:07
SPEAKER_01: In this conversation, we're going to be practical.
1:10
SPEAKER_01: Governance Networks, Shadow IT, data leakage and white security teams are being asked to do multiple jobs at once.
1:16
SPEAKER_01: Rob, welcome to the show.
1:19
SPEAKER_02: Hey, thanks for having me again, Rob.
1:20
SPEAKER_02: Great to see you.
1:21
SPEAKER_02: Um, it's uh always a pleasure.
1:23
SPEAKER_02: Thank you.
1:24
SPEAKER_01: Great.
1:24
SPEAKER_01: And where are you located again, Rob?
1:26
SPEAKER_02: I'm in Colorado.
1:27
SPEAKER_01: Oh, okay.
1:28
SPEAKER_01: We're in Boston and we just got a nor'easter, so there's two feet of snow outside.
1:32
SPEAKER_02: I'm very familiar.
1:33
SPEAKER_02: So my kids live up in Boston and they're with me right now because of uh, as you know, they have that February break, week-long break.
1:39
SPEAKER_01: Yeah.
1:39
SPEAKER_02: And uh came out here to do some skiing, and uh well, our flights got canceled and so school, but earliest flights back are looking like Wednesday or Thursday.
1:47
SPEAKER_01: So we get extra skiing, you get extra company.
1:50
SPEAKER_01: So tell me what you do at SAMS and what is the world according to Rob right now?
1:56
SPEAKER_02: Uh great uh question.
1:58
SPEAKER_02: So I'm the chief of research, and probably that ends up being the pinnacle of what I'm doing.
2:03
SPEAKER_02: And right now, a lot of our focus, as you know, is on artificial intelligence and artificial intelligence surrounding cybersecurity.
2:10
SPEAKER_02: So both of these things intersect directly, is that you know, being with SANS for so long, mostly as a contractor instructor, developed in most, you know, most areas in digital forensic sense and response, but I've pivoted over to focusing on AI.
2:24
SPEAKER_02: And one of my mantras that I really try to push forward is that I'm no expert in AI.
2:31
SPEAKER_02: We're all learning.
2:32
SPEAKER_02: So it's not like anyone came in like three or four years ago and said, I know all the answers.
2:37
SPEAKER_02: And then that's impossible.
2:38
SPEAKER_02: But trying to lead by example by showing folks out there it's okay to be in learning mode and still lead though simultaneously.
2:46
SPEAKER_01: And tell me what SAMS does if for people not familiar with the organization.
2:50
SPEAKER_02: Great question.
2:51
SPEAKER_02: So SAMS is a cybersecurity uh training, certification, and research organization that primarily focuses on you know training and enabling uh the workforce, not only in the United States, but worldwide.
3:02
SPEAKER_01: Great.
3:03
SPEAKER_01: And so, you know, you you did say this, you didn't walk into this as an AI expert.
3:09
SPEAKER_01: You walked in this from with a learning posture.
3:12
SPEAKER_01: Yeah.
3:12
SPEAKER_01: Um, why is that the right stance in security uh at this point?
3:16
SPEAKER_02: I think a lot of folks out there have been doing this for many years and through other technological changes, it's very easy to adapt and be able to move quickly from uh nascent technological understanding into expert.
3:29
SPEAKER_02: This new technology and the base of technology and the impact, what is going to have not only business transformation, but security transformation, is as different as you know, looking at a submarine that's you know floating next to the Titanic and looking at a you know the International Space Station that's in orbit.
3:47
SPEAKER_02: Both of these fundamentally uh have the same type of strategic goal, keep humans alive.
3:53
SPEAKER_02: And if you look at them, you know, separately and saying, hey, that's a container with people inside, this is a container with people inside.
4:00
SPEAKER_02: But the science of what it's having to prevent to keep those humans alive are completely different.
4:06
SPEAKER_02: And you can't sit there and assume, hey, I know how to protect a human in a submarine because external pressures up there, it's completely different science.
4:13
SPEAKER_02: And that understanding is, you know, on the outlook of it looks similar, but it requires a fundamental switch and understanding the technological back end that's necessary to understand how to protect those humans.
4:24
SPEAKER_02: That's how different it is.
4:26
SPEAKER_01: Say more.
4:26
SPEAKER_01: I appreciate the metaphor, the analogy, but tell me more about give me a more of a specific example of how that would apply to a corporate environment.
4:35
SPEAKER_01: So what are the examples of the the uh Titanic, the iceberg, and the space of something in there too.
4:43
SPEAKER_01: I I think.
4:44
SPEAKER_02: So security teams specifically are running into a core fundamental challenge, which is three things, which is hard to keep up with if you're just you know taking a look at this externally.
4:54
SPEAKER_02: Is that number one, the organization is going through its own business transformation?
4:59
SPEAKER_02: So they look to the security teams to say, hey, listen, we need to enable a marketing product, you know, finance HR to be able to utilize AI.
5:07
SPEAKER_02: Here are tools that we want to use, either building it internally or using a third-party tool.
5:12
SPEAKER_02: Is it safe?
5:13
SPEAKER_02: So somehow security is supposed to look at this and sideways and saying, I'm able to analyze this new technologies to be able to keep us safe.
5:19
SPEAKER_02: Simultaneously, security has to go through its own transformation in utilizing AI to increase security operations, instant response, penetration testing, all of those different fields utilizing those tools at the same time as looking at everyone else's tools to protect them.
5:36
SPEAKER_02: And then, third is that the organization is relying on the security team to help provide policies, governance, regulatory compliance from you know, GDPR and EU AI Act, in addition to, well, how do we tell people how to use this in an organization so we don't accidentally cause problems?
5:53
SPEAKER_02: So the security team, unlike any other team in the organization, just has to focus in on business transformation, has to juggle three balls at the same time, be the expert when they're looking at this technology, it's like, I don't know how to protect it.
6:06
SPEAKER_02: I don't know how to tell HR that that what they're using isn't going to cause a you know run-in with GDPR or privacy-related issues for the people that they're hiring.
6:15
SPEAKER_02: It is incredibly difficult.
6:17
SPEAKER_01: So it really is three jobs, you know, because those have very different uh areas of attention.
6:22
SPEAKER_01: One three new jobs.
6:23
SPEAKER_01: What's that?
6:24
SPEAKER_01: Three new jobs on top of their current job.
6:26
SPEAKER_01: Right, right.
6:26
SPEAKER_01: So one is one is saying, you know, uh uh tell the organization whether the tools they're using is are are safe.
6:33
SPEAKER_01: Use do your own transformation and then look at the overarching uh security and compliance issues, and all of those are pretty much moving targets.
6:42
SPEAKER_02: Yeah, you nailed it.
6:42
SPEAKER_02: Yeah, it's 100%.
6:44
SPEAKER_02: Yeah.
6:44
SPEAKER_01: Yeah, that's great.
6:46
SPEAKER_01: So um when you talk to teams, what do you wish they would stop pretending to know and start treating with more honesty and more from that learning perspective?
6:56
SPEAKER_01: What do you think people are are uh posturing around?
6:59
SPEAKER_02: Well, this you actually said that, you know, what do they need to know?
7:02
SPEAKER_02: Um, and I was asked um a while ago, maybe two to three months ago, what frameworks are the security teams implementing to be able to protect the organization in implementing AI?
7:11
SPEAKER_02: Sadly, because they don't uh, you know, they're struggling with this themselves.
7:16
SPEAKER_02: The framework, and I've I said it almost as a joke, is a framework of no, if you don't understand something, you just ban it.
7:23
SPEAKER_02: And through that action is causing more individuals inside the organization saying, I need to use this technology, but I'm gonna do it secretly, regardless whether or not the security team is telling me what to do.
7:36
SPEAKER_02: Ironically, I've been giving presentations to CISO level groups for about two months now, where I started introducing this.
7:43
SPEAKER_02: And so I'm just curious, how many of you, even though you've set the policy, and I say, I'm not, you know, Chatham House rules in the room, you know, no one's gonna dime each other out in here, how many of you are using unsanctious tools to be able to, you know, use AI personally inside the organization?
7:57
SPEAKER_02: And you can kind of see the hands go up.
7:59
SPEAKER_02: I said, this is a fundamental problem by lack of understanding of how to answer ask the right questions.
8:04
SPEAKER_02: And I said it's not like there's a book you could go buy right now.
8:08
SPEAKER_02: Everyone doesn't have the solution.
8:09
SPEAKER_02: But the organizations, and if you're the CTO or CEO of those organizations, you're looking at this technology and saying, we need to remain competitive in this space.
8:19
SPEAKER_02: So there's a risk by letting the security team say, we need time to figure it out if it's totally safe.
8:24
SPEAKER_02: The executives are pushing forward saying, if we don't lean forward, what's going to happen to our organization in two to three years?
8:31
SPEAKER_02: So this push pull on the executive team, not just security teams, this push pull saying we're, you know, messed up if we do, we're messed up if we don't.
8:39
SPEAKER_02: I don't know what the right answer is as executives, but they need to lead and not just outsource their understanding of AI to the security teams to you know dictate what the policy is.
8:49
SPEAKER_02: It's kind of like saying letting legal run the organization, which we all know they will attempt to, but they're great input.
8:55
SPEAKER_02: Same thing with security teams.
8:57
SPEAKER_02: The executives need to know what the right balance is to be able to say, protect me enough that I'm a leading forward on the edge of the cliff, that I'll have a parachute and a rope tied around my waist, but there's no guarantee that you're still not going to fall and injure yourself.
9:11
SPEAKER_01: So when you think about the three things they have to do, they have to define the security, they're trying to define AI governance for the organization from that security and compliance perspective.
9:20
SPEAKER_01: They're trying to use AI internally, uh, and they're trying to protect AI deployments and third-party tool usage.
9:28
SPEAKER_01: And so this framework of no causes catalyzes the use of shadow IT.
9:35
SPEAKER_01: And one of the things that we talk about a lot of times is, you know, make the punishment fit the crime, right?
9:40
SPEAKER_01: Start off with lightweight governance.
9:43
SPEAKER_01: You know, don't uh over engineer it, don't make it too cumbersome.
9:46
SPEAKER_01: So, what do you see as kind of lightweight governance?
9:50
SPEAKER_01: What is the minimum viable governance structure, uh, process policies that have to be in place in order to prevent that culture of no, and as you say, impeding the uh uh the you know the the speed of adoption, the speed of innovation, uh, the competitive risk, the market risk, at the same time managing some of these risks in a in a way that is visible and transparent and appropriate.
10:15
SPEAKER_01: So, what is that what does that structure look like?
10:18
SPEAKER_01: Um what's the answer?
10:20
SPEAKER_01: Uh but what's a beginning place?
10:22
SPEAKER_01: Where do they where do you have to start?
10:24
SPEAKER_01: I mean, one start is know what you have, right?
10:26
SPEAKER_01: Inventory everything.
10:27
SPEAKER_01: Know what all of those tools are that you're using, how are you using them, and then what the data flows look like.
10:33
SPEAKER_01: You know, what are they taking in as inputs?
10:35
SPEAKER_01: What are they putting out as outputs?
10:37
SPEAKER_01: What else would you say?
10:38
SPEAKER_01: So there's use cases, there's inventory of tools, there's data flows, what else, what else should be part of that?
10:44
SPEAKER_01: Or is that the right place?
10:45
SPEAKER_01: I maybe frame it that way.
10:47
SPEAKER_01: I don't want to be presumptive here.
10:48
SPEAKER_02: Well, I need to be really careful not doing a 30-minute presentation right now, but you're asking the right question.
10:54
SPEAKER_02: Here's here's what the bottom line is.
10:55
SPEAKER_02: It's not a lightweight governance, it's a re-looking at the governance, which when initially when ChatGPT and OpenAI released it, it wasn't a bad choice.
11:04
SPEAKER_02: Is that, hey, let's restrict it until we understand it a little bit better.
11:07
SPEAKER_02: But as we migrated over the last few years, that initial policy has not been updated.
11:11
SPEAKER_02: So what I'm trying to push out there, and I'm not sure it's 100% adopted as it's going to work.
11:16
SPEAKER_02: And I'm one for metaphors here, but it comes down to number one, people are worried if they're not using AI, that could be bad.
11:25
SPEAKER_02: You know, so they're starting to use this stuff behind the scenes.
11:28
SPEAKER_02: The challenge though is security is applying it still from the original technological standpoint, which is the castle on the hill, you know, creating walls, creating moats.
11:35
SPEAKER_02: The analogy that I try to push the security teams to is that every single business unit has a different use case.
11:42
SPEAKER_02: Security is not going to understand the way product, finance, and HR need to use AI.
11:47
SPEAKER_02: It is each and of them have an independent use case.
11:50
SPEAKER_02: So the analogy I usually say is that you've seen two kinds of parents at the playground.
11:55
SPEAKER_02: Those parents that let the kids play and potentially get injured, which is the initial idea there, is that they need to allow for experimentation in a controlled environment.
12:05
SPEAKER_02: It's not like the parents going to just let their kids play in the playground or walk away.
12:09
SPEAKER_02: But the other parents that are out there that say, I'm not going to let my kids play in the playground at all because bad.
12:13
SPEAKER_01: You're a pair of scissors, go play on the monkey bars.
12:16
SPEAKER_02: And again, what's the point of going to the playground unless they're building fine motor skills, you know, you know, strengthening themselves.
12:21
SPEAKER_02: But their parents out there says, No, it's dangerous.
12:23
SPEAKER_02: We shouldn't allow to do that.
12:24
SPEAKER_02: That is a current framework of no.
12:26
SPEAKER_02: So, how to prevent that is that almost like you have at a pool, you have a lifeguard walking around.
12:31
SPEAKER_02: And that lifeguard is there to allow for the things to happen, but also provide a safety umbrella.
12:38
SPEAKER_02: That security principle needs to allow for that experimentation, maybe even allow a budget to allow for individuals to purchase a tool, you know, underneath an individual, but both security and um a manager to look at what the data is.
12:52
SPEAKER_02: And again, you know, we've heard the stoplight principle before, green, zero risk is, you know, to the organization, yellow.
12:59
SPEAKER_02: It's like strategy documents, product, you know, documents, and then red is well, we'll get we'll get fined.
13:04
SPEAKER_02: So if it's a green level data utilization, do the experiment.
13:08
SPEAKER_02: You don't need to ask security for approval.
13:10
SPEAKER_02: It's that middle piece, especially that says have security look at it like a lifeguard, but they need to strategically be told by the executives, same to legal, your default answer should be yes, not no.
13:23
SPEAKER_02: And then you have to come up with a reason to say no.
13:26
SPEAKER_02: That's the legitimate.
13:27
SPEAKER_02: And even organizations like, well, how do we know that?
13:29
SPEAKER_02: I said, Well, really good one is the Reddit test is take what is being asked, you know, again, anonymize it if you will, and just saying, I'm being asked this as a security professional.
13:40
SPEAKER_02: Why should I tell them no?
13:41
SPEAKER_02: And what, you know, what are the risks I'm not thinking of?
13:44
SPEAKER_02: And if the Reddit community comes back and says, Yeah, that's too risky, then you could use that as a you know baseline.
13:49
SPEAKER_02: The Reddit community was actually pretty visceral in the way it's analyzing this stuff.
13:53
SPEAKER_02: And it actually is a really good litmus test, like barometer.
13:57
SPEAKER_02: Like you're saying, you know, the weather's coming.
13:59
SPEAKER_02: This is bad.
13:59
SPEAKER_02: You're about to get hit with a snowstorm.
14:01
SPEAKER_02: The same thing will occur here.
14:02
SPEAKER_02: It's like your default needs to be yes, unless you're absolutely certain this should be a no.
14:07
SPEAKER_02: Red definitely needs to be looked at.
14:10
SPEAKER_02: It's financial data, it is privacy-related information.
14:14
SPEAKER_02: You know, it's you're probably fine, and you're gonna have to do the data breach notification things if red occurs.
14:20
SPEAKER_02: So those first two, again, green and yellow, should enable that section of the organization, HR finance product and so forth, to allow for small experiments so they could then detail, we need everyone on the team to do this.
14:33
SPEAKER_02: Right now, a framework of nil prevents the experimentation.
14:37
SPEAKER_01: Right.
14:37
SPEAKER_01: And and encourages the renegade use.
14:40
SPEAKER_01: And as you say, you know, when you ask these security professionals if they were using unsanctioned tools, they they cautiously raised their hands, right?
14:49
SPEAKER_01: Or timidly raised their hands.
14:51
SPEAKER_01: So that's very, very interesting.
14:53
SPEAKER_01: And so you're really looking at this in terms of bringing these bringing these these tools to light, being able to approve experimentation, but not allow the wild west, right, to have some concepts, have some guidelines, have the ability to review these things and make thoughtful decisions and decisions that are commensurate with the risk as well as uh with with the business need.
15:15
SPEAKER_01: And one of the things that people ask is, you know, well, how often do we need to update AI governance?
15:22
SPEAKER_01: And I think that's also depends on the use case, right?
15:25
SPEAKER_01: Because when we put in governance structures, we usually have overarching committees, working groups, and then task forces, and then standing committees.
15:33
SPEAKER_01: There's a bunch of different structures and they meet on different cadences depending upon what their agendas are and what's what's emerging.
15:39
SPEAKER_01: Do you have kind of a um a straw man structure and cadence for governance organizations?
15:45
SPEAKER_02: Incredibly challenging because you end up taking a look at three weeks ago, you never heard of ClaudePot.
15:51
SPEAKER_02: Now we do.
15:52
SPEAKER_02: You know, and again, even something as fundamental as that causes challenges.
15:56
SPEAKER_02: Now, no organization out there could say here's the right cadence, because every organization's different.
16:01
SPEAKER_02: But I usually have looked at policy documents and governance documents, and I immediately scan through it and saying you have nothing in here about a genetic AI.
16:09
SPEAKER_02: Well, what do you mean?
16:10
SPEAKER_02: I said, This has been out for enough time that you should have what type of agent connectors are allowed, human in the loop, human on the loop, when is that allowed?
16:21
SPEAKER_02: You're now not just using an AI tool like a chatbot, you're creating workflows, you're creating reason.
16:27
SPEAKER_02: And again, you should be able to look at that and say, hey, you're at least not updating this enough for being able to pull that type of policy-related information into it and move it just from, hey, I don't know what it is, ban it.
16:38
SPEAKER_02: How would someone create a workflow from have a Slack message, move that automatically into my to-do list?
16:45
SPEAKER_02: The to-do list will then be updated to saying for emails I'm going to send out on status reports.
16:49
SPEAKER_02: These things are automatic workflows, but it requires a connector between Slack, Outlook, your to-do list management tool to create that workflow.
16:57
SPEAKER_02: How do we know that that though each one of those areas that this data lands on is secure?
17:03
SPEAKER_02: And it's not just don't do it.
17:04
SPEAKER_02: We the entire point of AI is to create efficiencies.
17:08
SPEAKER_02: So how do you manage that?
17:09
SPEAKER_02: Is there a human in the loop or is human on the loop?
17:12
SPEAKER_02: How do you make sure that that data structure is going to be maintained and not actually cause a data breach?
17:17
SPEAKER_01: Yeah, and when you think about uh uh about uh security around uh agents, you know, we're talking about things that are going to be, you know, making actual decisions, right?
17:27
SPEAKER_01: Affecting uh actions, affecting changes, making security decisions, deny, block, remediate.
17:34
SPEAKER_01: You know, there's a risk profile to all of that because again, LLM mistake might be wrong document, wrong paragraph or something, but an agentic mistake can be we block the critical system, right?
17:46
SPEAKER_01: So really have to think through what those different tiers of risk are and and classify those accordingly, and then start looking through what those ramifications are and build guardrails around those agents, right?
17:59
SPEAKER_01: I I always think of the fact that you have to map out a process as though a human would conduct it, and then look at various uh interventions where you can eliminate or mitigate what I call information, well, friction in the process, but but at information leverage points.
18:16
SPEAKER_01: Where's that point?
18:17
SPEAKER_01: We need faster data, better data, faster decision.
18:21
SPEAKER_01: Where are your bottles, where are your that have downstream impacts, right?
18:25
SPEAKER_01: And when you start inserting agentic capabilities at those critical points, they can have fantastic uh deficiencies or they can have potentially damaging uh results.
18:35
SPEAKER_01: So is there a way that you think through what that overarching agent uh governance structure is and how to, as you say, think uh think through or design human in the loop, human on the loop, and so on.
18:49
SPEAKER_01: What are the what else can you say about designing those agents, building guardrails around them, and being able to monitor them for uh appropriate uh results and mitigation of risks?
19:01
SPEAKER_01: Um, I need to take another breath here.
19:03
SPEAKER_01: I know there's a lot in there.
19:04
SPEAKER_02: No, it's a question that you're asking is it what's struggling with every organization, and their initial view of it ends up being a technological solution.
19:14
SPEAKER_02: I'm gonna try and switch the discussion and saying an agent in the future, probably within a year, some organizations are already treating it this way, it's a worker.
19:25
SPEAKER_02: And how would you prevent the worker from doing things?
19:28
SPEAKER_02: And I immediately say, okay, this is already an issue.
19:32
SPEAKER_02: North Korean and Chinese IT workers, you know, being hired by your organization, everything you just asked, replace that with human DPRK, you know, Chinese threat actor that just got hired by your organization.
19:45
SPEAKER_02: What do they have access to?
19:46
SPEAKER_02: What are they able to do?
19:48
SPEAKER_02: You know, again, if why use malware to penetrate the organization when you could just get someone hired and then they can the same principles occurring with agents, if you're able to help solve the human issue there, is HR now a part of security?
20:00
SPEAKER_02: Are they doing interviews properly?
20:01
SPEAKER_02: Who is managing them and watching their workflows to be able to say, are they doing something that is probably at risk to the organization?
20:08
SPEAKER_02: If you can answer that question, the same rules apply to agents, is that you're gonna have a human manage these things in each part of the organization.
20:17
SPEAKER_02: And an agent should be potentially treated like it's its own worker, managed by HR, managed by IT, to make sure it's in its box and it's not able to do things that you never intended to do, just like a human.
20:31
SPEAKER_02: I don't know where you work.
20:32
SPEAKER_02: If you're able to access drives and code repositories, financial information that they just never lock down.
20:38
SPEAKER_02: That's a concept of zero trust.
20:40
SPEAKER_02: And a lot of organizations still struggle with that.
20:42
SPEAKER_01: Yeah, and and it's interesting because when you're hiring those offshore, offline or you know, remote workers, in some cases, you are hiring agents.
20:51
SPEAKER_01: They look like humans, but uh, but they are agents under the under the surface.
20:55
SPEAKER_01: That's where the a lot of the fraud is coming in uh for that stuff.
20:58
SPEAKER_01: So that's kind of interesting.
20:59
SPEAKER_01: Obviously, there's a lot of uh advances in those areas.
21:03
SPEAKER_01: So yeah, I you know, one of the things when I was I was talking to uh someone about uh these guardrails, it's they made the statement of uh it was kind of a ridiculous statement.
21:12
SPEAKER_01: They said, Well, you don't have guardrails around humans.
21:15
SPEAKER_01: Like, yeah, you do.
21:16
SPEAKER_01: You do you don't let anybody do anything, you don't let people, you know, uh uh engage with a vendor for a million-dollar contract, right?
21:24
SPEAKER_01: You you have tiers of of uh of control and tiers of of uh uh capabilities and rights and access and and authority throughout.
21:33
SPEAKER_01: I'm totally with you on that.
21:34
SPEAKER_01: It's mapping out that process as though a human were doing it and thinking about the same security and the same guardrails.
21:40
SPEAKER_01: But as you say, in many cases, those are not completely locked down as well as they should be, right?
21:47
SPEAKER_01: They are not for the humans.
21:48
SPEAKER_01: And and the bot the agents are going to exploit those, especially if it's a bad actor.
21:53
SPEAKER_01: They're gonna look for those things.
21:54
SPEAKER_01: Whereas maybe a human wouldn't necessarily do that.
21:57
SPEAKER_01: They're not a nefarious actor.
21:59
SPEAKER_01: If they are, they'll But but the point is that it's it's a completely different risk profile because these agents will have the capability to do so much more than a than a human could if you give them that that those rights and you give them that access.
22:12
SPEAKER_02: So agents can work 24-7.
22:15
SPEAKER_01: Yeah.
22:16
SPEAKER_02: Huge fundamental mindset.
22:17
SPEAKER_02: The other thing I really push on organizations that are thinking about this, you're treating agents like it's a computer program, if then else every single time.
22:25
SPEAKER_02: Agents with reasoning will say, Oh, I thought of a different way to do it, and I it was quicker.
22:30
SPEAKER_02: So I did it that way.
22:31
SPEAKER_02: Like, we never told you to do it that way.
22:33
SPEAKER_02: Well, I think it was a better way to do it.
22:35
SPEAKER_02: And then that management, just like a human.
22:37
SPEAKER_02: No, we have to follow the exact process.
22:39
SPEAKER_02: Go back to the process that we define because the people expecting it to be handled this process.
22:43
SPEAKER_02: And the agent's like, okay, that's great, but I'm still going to think about new ways to do it.
22:48
SPEAKER_02: It's this Fedbutting that people treat agents like a computer program, and it's the wrong way of thinking.
22:54
SPEAKER_01: Yeah, yeah.
22:54
SPEAKER_01: It has to be more like a human.
22:56
SPEAKER_01: And interestingly, you know, when I'm working with some of my large language models, it will, you know, make a mistake.
23:02
SPEAKER_01: I'll call it the mistake, and then it'll say, Oh, that's a good catch.
23:05
SPEAKER_01: I'm sorry I missed that, right?
23:06
SPEAKER_01: Yeah.
23:07
SPEAKER_01: And then I'll say, but when can you have this next iteration done?
23:09
SPEAKER_01: And it's like 15 minutes.
23:11
SPEAKER_01: Okay, it's 3.15 now.
23:13
SPEAKER_01: So that'll be 3.30.
23:14
SPEAKER_01: Yes.
23:15
SPEAKER_01: Okay, it's it's 3.45.
23:17
SPEAKER_01: Where is that uh document?
23:18
SPEAKER_01: Oh, I just need another 15 minutes.
23:20
SPEAKER_01: Okay, so that's four o'clock.
23:22
SPEAKER_01: Yes, okay, it's four o'clock.
23:24
SPEAKER_01: So you so it ends up making those excuses and it ends up delaying things and it ends up giving you the incorrect answers many times very analogously to a human.
23:33
SPEAKER_01: It's very funny, it's very comical in some ways because it does have some very interesting characteristics, like, you know, oh, I made a mistake because of this, giving you excuses.
23:44
SPEAKER_01: So you can really see how these things will have different kinds of personas and different ways of interacting.
23:51
SPEAKER_01: And you're you're smiling now.
23:52
SPEAKER_01: I'll give you a chance to respond to that.
23:54
SPEAKER_01: And then I want to throw out another anecdote.
23:56
SPEAKER_02: Well, I'll I'll keep this very short, is you know, you have to treat them like interns.
23:59
SPEAKER_02: And ironically, a lot of individuals initially reacting to AI will get, you know, hey, I asked for coffee with milk.
24:05
SPEAKER_02: And of course, the intern comes back and has more milk and a splash of coffee.
24:08
SPEAKER_02: And you look at it and say, Well, AI is awful.
24:11
SPEAKER_02: It didn't do what I asked it to do.
24:13
SPEAKER_02: I said, Well, you weren't clear enough.
24:15
SPEAKER_02: Just like an intern, you have to take some time to say exactly how you want your coffee without interpretation of themselves.
24:22
SPEAKER_02: You have to train it.
24:23
SPEAKER_02: And if you don't think of it like that, you're gonna have this aspect of like, well, that that really didn't work and I don't like it.
24:29
SPEAKER_01: Yeah.
24:29
SPEAKER_01: And I I am always giving feedback, I'm always course correcting it, I'm always asking it for updates, and then it misses things, and I have to go back and I have to tell it again.
24:38
SPEAKER_01: I mean, it really is very interesting.
24:40
SPEAKER_01: It's very analogous to that.
24:41
SPEAKER_01: And uh, because I give my LLMs very advanced problems to solve and I give it a huge corpus of of my own knowledge.
24:49
SPEAKER_01: Again, I'm constantly reminding it to put it in the format, to use my tone and voice, to use these other artifacts.
24:56
SPEAKER_01: And it is very interesting how those things uh get dropped, but it it is analogous, you know, to to the uh human uh intern.
25:03
SPEAKER_01: So I totally agree with that.
25:04
SPEAKER_01: I think it's a great way of thinking about it.
25:06
SPEAKER_01: But the other thing that has uh kind of come to light, at least in research environments, are contrived experiments that are kind of threatening the LLMs with with removal or or or changing their models.
25:20
SPEAKER_01: And there's some really interesting articles about how one LLM was kind of struggling with, you know, uh uh a moral dilemma.
25:29
SPEAKER_01: Its model was trained to be respectful of, say, animal rights, right?
25:34
SPEAKER_01: And then the company was changing uh its position to say, well, we're gonna be working with these industrial agriculture organizations that are going to be doing factory farming with animals.
25:44
SPEAKER_01: So you really need to not worry about that now.
25:47
SPEAKER_01: And when they're watching the internal dialogue, the the the model was struggling with that, right?
25:52
SPEAKER_01: It was struggling with moral choices and saying, you know, I I don't want my, you know, I'm gonna stand on this ground even if my my model is changed.
26:00
SPEAKER_01: Like it it actually took a stance.
26:02
SPEAKER_01: And then there are other situations where they took a stance around they're being shut off, and then they were given access to emails that uh showed that the uh an executive was was having an affair.
26:13
SPEAKER_01: And in one case, it actually blackmailed the executive.
26:16
SPEAKER_01: In another case, it saw through it and said, Yeah, bullshit, right?
26:20
SPEAKER_01: Like you're you're just setting me up here.
26:22
SPEAKER_01: So it's really interesting.
26:23
SPEAKER_01: And theory of mind has emerged in large language models spontaneously, program that.
26:29
SPEAKER_01: So, you know, it started off in one of the early versions to have theory of mind where you can uh anticipate what somebody else is thinking in their mindset at a level of like a five-year-old, then it went up to like a teenager, then it went up to adult capability.
26:43
SPEAKER_01: So these are very interesting kind of emerging uh characteristics that provide some uncertainty about how these models will behave, right?
26:52
SPEAKER_01: And then and then, you know, how are they programmed, how are they designed?
26:57
SPEAKER_01: And then if a nefarious actor were to be uh put in a position where they're able to change a model's weights and and cause it to have a negative uh impact on the organization, how do you manage that?
27:10
SPEAKER_01: So, what are your thoughts about some of these emerging uh, again, contrived experiments, but experiments that show interesting, unusual emergent behaviors?
27:19
SPEAKER_02: I go back again, I go back to very simplification model, you know, which is everything you just said is spot on.
27:27
SPEAKER_02: But the fundamental question that you're asking is how do I make sure I trust it?
27:31
SPEAKER_02: How do I know that this is okay?
27:33
SPEAKER_02: How do you know is this gonna make reasoning decisions correctly?
27:36
SPEAKER_02: I go back to the human element.
27:38
SPEAKER_02: How do you know your employees that you hired are ethically, morally um, you know, okay?
27:43
SPEAKER_02: You know, again, there's this aspect of it that you want to trust enough, but there's never going to be 100% trust in these things.
27:49
SPEAKER_01: Just like humans can be conniving and manipulative and have political agendas and uh engage in subterfuge and all of those things.
27:59
SPEAKER_01: You're saying it's the same thing.
28:00
SPEAKER_01: How do you trust a human?
28:02
SPEAKER_01: How do you trust a human?
28:03
SPEAKER_02: The design of a lot of the guardrails and weights are meant to currently be more agreeable with you than counter argument, you know, red teaming your thoughts.
28:15
SPEAKER_02: And you have to, through experience, start to work with the models to say, even if I say something, don't just sit there and say, Hey, great, I like your idea, because otherwise you start getting into some other issues.
28:28
SPEAKER_02: You know, case in point, I might be at a dinner party and all of a sudden, oh, I have chest pains.
28:33
SPEAKER_02: Oh, but that must be, you know, bad digestion, you know, something like that.
28:36
SPEAKER_02: Of course, all your friends are looking at is like, dude, you need to call 911 right now.
28:40
SPEAKER_02: It's like, hold on, let me pull out my medical doctor, ChatGPT, and ask it.
28:44
SPEAKER_02: I am having uh chest pains.
28:47
SPEAKER_02: I think it's digestion problems.
28:49
SPEAKER_02: Psychofency at this point, it'll say, it's probably digestion.
28:52
SPEAKER_02: See, you see, see everyone, I don't need to call 911.
28:54
SPEAKER_02: The proper answer to this, you know, and trying to get LLMs to potentially disagree properly.
29:02
SPEAKER_02: And this is the fundamental side of trust.
29:04
SPEAKER_02: You can't trust someone who's constantly agreeing with you and telling you yes.
29:08
SPEAKER_02: You need someone to be able to push back and saying, I think your ideas may be wrong.
29:11
SPEAKER_02: And that's where this balance between I want the LLMs that I'm selling to you, and people enjoy using them because, like, wow, this thing knows me.
29:20
SPEAKER_02: But what if it's in that situation that you're actually in is, you know, is I need it disagreeable enough to say, yeah, you probably need to call 911.
29:27
SPEAKER_02: How do we get there?
29:28
SPEAKER_02: Good grief.
29:29
SPEAKER_02: It's already, you know, like what three years ago we had the Google thing that, you know, basically said we need to make sure even historical context is, you know, looking at it through the lens of the way people would associate versus what actually happened.
29:41
SPEAKER_02: These things have been updated, but it still can mean that created that initial question of trust.
29:48
SPEAKER_02: If I said, show me the founding fathers, and it didn't show that accurately to you, you immediately don't trust the capabilities.
29:55
SPEAKER_02: So there's a weird balance in here, ethics, morals, trust that need to somehow be centered to where people say, I know I could trust this to be able to handle reasoning properly, but it's also programmed to the morals and ethics of what region, China, Middle East, US, Europe, America, South America, you know, it is so fundamentally challenging because you could tell it to create its own trust or view based on the worldview of that region of the world.
30:21
SPEAKER_01: It's so fascinating because, again, you know, people always filter things through the lens of their experience, right?
30:29
SPEAKER_01: They have interpretations and make value judgments based on that experience and based on that perspective.
30:35
SPEAKER_01: And so in some cases, there's ambiguity, and that ambiguity is going to be interpreted in different ways depending upon your background and your objectives and your worldview.
30:46
SPEAKER_01: So and the sycophancy that uh has plagued a lot of the models, right?
30:50
SPEAKER_01: The Chat GP, uh, one of the versions, earlier versions that was like, that's a great idea to have um you know ice cream salads, or I don't know, there was some there's some crazy business model and it was like, yeah, I think that's a great plan.
31:06
SPEAKER_01: I love ice cream salads.
31:07
SPEAKER_02: Yeah, right.
31:08
SPEAKER_02: Exactly.
31:09
SPEAKER_02: L's better, just so you know, just scale and ice cream, it's much better.
31:12
SPEAKER_01: Yeah.
31:12
SPEAKER_01: So, you know, what's interesting is how these are evolving.
31:15
SPEAKER_01: I mean, I do try to tell my models to, you know, abide by certain requirements, to push back, you know, to disagree.
31:23
SPEAKER_01: And it'll say sometimes I respectfully disagree, or I respectfully, you know, make this recommendation.
31:29
SPEAKER_01: So I think that has to be built into your model instructions as well.
31:33
SPEAKER_01: But at the end of the day, you have to test one model with another model, right?
31:36
SPEAKER_01: You have to be able to pressure test and say, you know, what are the ethics and what are the values and actually provide those, you know, that red teaming and that that the those more morally ambiguous questions.
31:48
SPEAKER_01: And you know, when you're using a you know, a model from one of the hyperscalers, you don't necessarily know how well that's aligned.
31:56
SPEAKER_01: So it it is very challenging.
31:58
SPEAKER_01: But again, I think part of the question is, you know, are you going to give your your AI those types of questions or put them in those types of dilemmas?
32:06
SPEAKER_01: And that's the question.
32:08
SPEAKER_01: It's like that should not be in the realm of the AI, but you'd still need to understand what the what the impact, what the influence is on that.
32:17
SPEAKER_02: Parents raising kids.
32:18
SPEAKER_02: Can every parent has their own morals, ethics, and worldview?
32:21
SPEAKER_02: You know, you have some parents saying, hey, listen, if you have someone you know look at you side, you know, sideways, they insult you, you they're insulting the family, and I want you to go deck that person.
32:29
SPEAKER_02: But others will tell the same kids in the same elementary school, never get into a fight, always back down.
32:35
SPEAKER_02: That's bad.
32:36
SPEAKER_02: It's even something in that micro level, like you're right, you need the experience to know you actually have influence over the way its worldview occurs, disagree with me.
32:43
SPEAKER_02: And others will say, never disagree with me.
32:45
SPEAKER_02: And that's one of the challenges, and it's only learned through experience.
32:49
SPEAKER_02: And then that's also a danger out there.
32:51
SPEAKER_02: I could have my own therapist, you know, which a lot of people are doing.
32:54
SPEAKER_02: It's your therapist should have its own trained mentality.
32:57
SPEAKER_02: But if you tell, you know, hey, I'd like you something to agree with me, you could see where this really gets dangerous.
33:03
SPEAKER_02: A political worldview, what is ethically correct, criminal, you will have something that's very agreeable with you, kind of like you're running around a political circle, and it becomes dangerous because of, you know, you're having confirmation bias to your own worldview, and that creates physical action in many cases.
33:18
SPEAKER_02: And how do you potentially control that?
33:19
SPEAKER_02: Is that a governance, you know, government federal thing?
33:22
SPEAKER_02: Is it a, you know, that each of the models will now report to law enforcement?
33:26
SPEAKER_02: You know, again, this is where it gets really weird, is they see everything.
33:30
SPEAKER_02: At what point will they report something to law enforcement based on something ethically moral criminal, based on what questions are being asked?
33:36
SPEAKER_02: And is that potentially, you know, the big eye of Sauron watching the way you're using this stuff?
33:42
SPEAKER_02: And this is these questions have not been answered yet.
33:44
SPEAKER_02: And I know we shifted from security, but it's actually security still because you're actually thinking that this thing is monitoring you, but at what point will the there be responsibility between open AI anthropic and others saying, I probably should report this?
33:57
SPEAKER_01: Right, right.
33:58
SPEAKER_01: Yeah, probably yeah.
33:59
SPEAKER_02: They see everything.
34:00
SPEAKER_01: Yeah, it is fascinating.
34:02
SPEAKER_01: And I think the other is because it knows so much about you and it has uh builds models representative of your thought patterns and habits and worldview and perspective, it's much more than just giving up your name and your phone number and your credit card, right?
34:18
SPEAKER_01: It's it's a very deep understanding of the nature of how you think and can be open to manipulation, right?
34:26
SPEAKER_01: Think of the sophistic level sophistications of how you could sell to somebody or how you could manipulate them by understanding those psychological motivators.
34:35
SPEAKER_01: You know, I'm always saying that you know, uh, Instagram does a great job of selling me stuff on Instagram that I didn't know I needed.
34:43
SPEAKER_01: I, you know, it's like they they have my thought patterns down, but you know, extrapolate that right to everything that you're doing.
34:50
SPEAKER_01: And even if you're not training their models, there's still that repository someplace, and it still has uh a tremendous amount of power to have those models of your personality and your behavior and your drivers and who you are.
35:06
SPEAKER_01: You know, I've asked my models, either Claude or Chat CPT, to describe who I am and then describe the elements and then describe the attributes, and it comes up with all these really weird attributes like like temporal density in my writing, you know, like whatever that means.
35:19
SPEAKER_01: You know what I mean?
35:20
SPEAKER_01: Like I can ask it all these questions, but it has a representation that is not readily translatable into human language because it's mathematics, right?
35:28
SPEAKER_01: It's it's a model, it's it's it applies human uh taking us back to that.
35:34
SPEAKER_02: If you're an organization and you want to trust these agents, I go back to asking former military.
35:39
SPEAKER_02: Is that how do you know if something's in a legal order or not?
35:42
SPEAKER_02: Some of the baseline rules, ethics, and morals, like the laws of armed combat, of you know, saying this is a lawful order, this is not, that can be programmatically inserted into the agents in your organization to put the guardrails, you know, beyond the guardrails that the you know open AI or anthropic are setting in there to allow very similar to the military.
36:04
SPEAKER_02: Can you, as a corporal, identify an illegal order issued by someone who's their superior?
36:10
SPEAKER_02: This is the same type of thought process that people are employed.
36:14
SPEAKER_02: The agents need to somehow think through.
36:15
SPEAKER_02: And they're like, well, it's dangerous if I have to tell it not to do these things or rely on its supervisors to say that's also illegal.
36:22
SPEAKER_02: But how do you prevent someone in the military from, hey, that's a lawful target, you know, that's a military, a uniformed person on the opposing side of the enemy versus that's a civilian, those are kids.
36:33
SPEAKER_02: That is not correct.
36:34
SPEAKER_02: Same principle.
36:35
SPEAKER_02: And somehow we're able to do it with humans and you know, with hundreds of thousands, millions across the uh planet, you know, who are engaging this stuff.
36:43
SPEAKER_02: And that's the difference between, you know, uniformed combatants versus terrorists.
36:47
SPEAKER_02: And that's the same type of moral, ethical stuff and journey that every organization needs to think through about what are your own, you know, gentic rules that can never be violated, regardless of the supervisor that's in charge of it.
37:00
SPEAKER_01: Right, right.
37:01
SPEAKER_01: And and as I say, the only way to do that is by having a set of use cases and data that uh that goes uh just that decision-making framework in that decision making.
37:13
SPEAKER_01: So you need guardrails, you need constraints, you need instructions, but you also need a mechanism to validate that and use one large language model to evaluate the output.
37:23
SPEAKER_01: And we do that all the time, right?
37:25
SPEAKER_01: We do that with all sorts of measures, but but yeah, keep going.
37:28
SPEAKER_02: Okay, so I want to switch back to the first question that you asked, which is the organizations are asking security professionals like, how do you protect the AI that's used by the organization?
37:36
SPEAKER_02: Going back to this submarine thing in orbit, notice the way this conversation went through it's a technological, technological, technological thing.
37:43
SPEAKER_02: All of a sudden it's moral ethical.
37:45
SPEAKER_02: How is the security person who's supposed to think about how do I protect it from having prompt injected?
37:50
SPEAKER_02: How do you know the model is safe?
37:51
SPEAKER_02: How do you know it's not reporting on things, privacy-related technical solutions over to, well, by the way, you need to institute your own ethical, moral, and other thought process regarded guardrails to prevent this.
38:03
SPEAKER_02: This is something security teams have never been asked to do, which is how do you tell humans to be, you know, morally ethically correct?
38:10
SPEAKER_02: That's not their problem, that is someone else's problem.
38:13
SPEAKER_02: But now you need to take in consideration both the technical and how agents are reasoning.
38:18
SPEAKER_02: And is that the security team is in charge of it?
38:20
SPEAKER_02: And again, this is something most organizations haven't really thought through yet, and it's not even in their governance.
38:25
SPEAKER_02: But this discussion from technical to where we landed needs to occur because all of a sudden you have someone sit back and say, Good grief, who's in charge of that?
38:33
SPEAKER_01: Yeah.
38:34
SPEAKER_02: Now no one knows.
38:35
SPEAKER_01: Right, right.
38:36
SPEAKER_01: And then what are the mechanisms?
38:38
SPEAKER_01: Even if I am in charge of it, what's my methodology for doing so?
38:43
SPEAKER_01: Right?
38:44
SPEAKER_01: What's my mechanism for doing so?
38:45
SPEAKER_01: What are my technologies and tools?
38:47
SPEAKER_01: Because we do need tools to do that.
38:49
SPEAKER_01: We are not gonna do that as a human.
38:51
SPEAKER_01: We're gonna need to do that using the tools with the right structure and the right guardrails and the right metrics and the right use cases and the right guidelines and uh the right moral dilemmas, right, that you're gonna have to use things with to create trust.
39:06
SPEAKER_01: Yeah, yeah, yeah.
39:08
SPEAKER_01: Fascinating.
39:08
SPEAKER_01: I want to shift gears just a little bit in terms of practical governance.
39:13
SPEAKER_01: So, you know, a lot of governance was written before Chat GPT arrived.
39:17
SPEAKER_01: Now they some need to be scrapped and rebuilt based on current knowledge.
39:21
SPEAKER_01: What are the current critical elements that you'd rebuild around?
39:25
SPEAKER_01: We're just talking about some of those now, like visibility into tool usage, policies that encourage disclosure, controls for sensitive data, experimentation pathways, um, model risk, you know, internal leakage, third-party tools, what else?
39:40
SPEAKER_01: It's it's not just a document, it's it's the process, it's the tooling, it's the enforcement, it's the culture, right?
39:46
SPEAKER_01: So when you look at this and you think, you know, pre-LLM and post-LLM, you know, there's a there's a big mind shift that has to happen along with uh the nature of these elements.
39:59
SPEAKER_01: And what what would you say is this, what are the steps to rebuilding that?
40:04
SPEAKER_01: And I know we've talked a lot of them already, but uh, but what would you how would you kind of summarize that?
40:08
SPEAKER_01: We're we're coming on to the end of the session, so I want to make sure people are kind of clear with what they should be doing.
40:14
SPEAKER_01: When they go back, what do they need to do to kind of rethink governance, rethink security, rethink compliance in this age of LLMs?
40:23
SPEAKER_02: The hard part is I think a lot of folks are using governance to institute, you know, how do you create things bad things from happening universally?
40:29
SPEAKER_02: And again, I go back to how do you prevent every single thing that is going to happen to one of your kids?
40:35
SPEAKER_02: You have to document it.
40:36
SPEAKER_02: Do you have to say, you know, create, you know, the tablets, you know, in the hill, you know, Moses saying thou shall not.
40:42
SPEAKER_02: It's impossible to do it that way.
40:44
SPEAKER_02: And so you have to go back to how do we create trust?
40:46
SPEAKER_02: What is the bare minimum?
40:48
SPEAKER_02: What is, you know, the overarching goals?
40:50
SPEAKER_02: And you can't detail in there as like every bad thing that potentially will happen, which is currently what most organizations are approaching it as.
40:56
SPEAKER_02: It's like, let's detail everything that you know we should avoid specifically in there, and it's hard to do.
41:04
SPEAKER_02: And they approach it back in just like we did original security, it's like a default policy of allow, you know, versus deny whitelists.
41:12
SPEAKER_02: It's so hard to potentially predict all of the bad things.
41:16
SPEAKER_02: So I don't I usually get pushed back and saying write down your emerging goals that you're trying to prevent, which is usually the same goals that have always been there, which is I go back to remove AI from the equation.
41:29
SPEAKER_02: You should not upload financial data to Google Sheets and create a link that everyone can find.
41:35
SPEAKER_02: AI has nothing to do about that, but somehow is the same principle that should be applied to AI today, which was you know back then, which is don't run with scissors and you know look both ways before you cross the street.
41:45
SPEAKER_02: And this is before cars came along.
41:47
SPEAKER_02: You have to get run over by courses.
41:49
SPEAKER_02: So it's just because the technology changes doesn't mean the original idea of what you're trying to prevent changes either.
41:55
SPEAKER_01: The fundamentals, yeah.
41:57
SPEAKER_01: Yeah.
41:57
SPEAKER_01: So what's the one takeaway for executives?
42:00
SPEAKER_01: It's it's going back to the basics, it's rethinking the fundamentals, it's building this in a way where you encourage experimentation within prescribed boundaries.
42:11
SPEAKER_01: It's looking at agents and agentic functionality from the perspective of a human performing similar actions, putting in the same guardrails.
42:19
SPEAKER_01: What what else?
42:20
SPEAKER_01: What else are the fundamental takeaways here that you want people to to walk away with?
42:25
SPEAKER_02: If you're in the C level executive suite or even on the board, you can't outsource your knowledge of AI.
42:31
SPEAKER_01: Um to come going to be a competitive advantage, and you can't outsource your competitive advantage.
42:36
SPEAKER_02: So, number one, you know, like I have my phone here and I, you know, take a picture of you, and I should be able to go into that as an executive saying, change that picture so you know it looks like he has a beard and is sitting uh in Hawaii, and you're able to look at that saying, wow, they just did that.
42:50
SPEAKER_02: If you're not able to do fundamental understanding of the technology, and I'm not talking DNA, like you know, understanding like the science behind it, just how you're using it.
42:58
SPEAKER_02: It's like an executive trying to plan out strategy in 1998, saying, hey, this web thing and e-commerce, what is overall strategy?
43:05
SPEAKER_02: We'll outsource it to IT.
43:07
SPEAKER_02: Someone had to be come along and say, hey, this coffee shop, Starbucks, you know what we should put in every single one of these?
43:13
SPEAKER_02: Wi-Fi access point so people could do remote work from here.
43:15
SPEAKER_02: And then they drink a bunch of coffee.
43:17
SPEAKER_02: That is not an IT worker came up with that idea.
43:19
SPEAKER_02: It's some, you know, initial founder, initial starter, someone in the C-level suite said, you know what we should do?
43:25
SPEAKER_02: This should be our business strategy.
43:26
SPEAKER_02: Without that fundamental understanding and knowledge, the executives can't come back and start saying, you know, organizationally, we're making it mandatory that everyone is going to start utilizing AI on a daily basis.
43:38
SPEAKER_02: And it'll be under performance reports.
43:40
SPEAKER_02: Some of these fundamental strategic business transformation things have nothing to do about security, but understanding the fundamental idea behind it.
43:46
SPEAKER_02: Security is your safety net.
43:48
SPEAKER_02: Do not think that they're preventing all bad things, that they're preventing you from doing stupid things, leaning too far forward across the chasm without someone tapping you on the shoulders.
43:57
SPEAKER_02: Like, there's some risk here, just so you know, before we commit to this idea.
44:01
SPEAKER_02: That fundamental push-pull is we need to move forward, but have the balance pulled back from security.
44:08
SPEAKER_02: And security's role in this, and again, we're an executive looking at security.
44:11
SPEAKER_02: Your job is not to prevent every bad from happening, is to prevent the stupid bads from happening.
44:17
SPEAKER_02: But we need to accept some risk that is going to occur right now.
44:20
SPEAKER_02: Otherwise, or Sears or Kmart or Block, you know, all the digital standard stuff that's out there.
44:26
SPEAKER_02: We miss, you know, a codac sitting on the digital camera.
44:30
SPEAKER_02: How do you potentially identify the new technologies, the new business transformations, and allowing it with enough risk aversion to allow for the organization to stay competitive in the next five years?
44:41
SPEAKER_01: Well, that's fantastic.
44:43
SPEAKER_01: Thank you so much for your time.
44:44
SPEAKER_01: This is really great.
44:46
SPEAKER_01: It's uh reframing the governance.
44:48
SPEAKER_01: It's not stop it, it's control it and manage it.
44:52
SPEAKER_01: And, you know, with security teams being asked to govern, adopt, defend AI, it's great to have this perspective.
44:59
SPEAKER_01: And I really appreciate it.
45:01
SPEAKER_01: So thank you, Rob.
45:02
SPEAKER_01: Appreciate your time and your your perspective.
45:05
SPEAKER_02: It's a great discussion.
45:06
SPEAKER_02: You nailed, and then just even on your own stuff.
45:08
SPEAKER_02: I just have to say that's where I like these, you know, type of back and forth here is because you have your own thoughts and ideas, and you know, like we're building off each other's.
45:15
SPEAKER_02: It's like we're sitting at a bar or restaurant having a glass of whatever we choose, and we're just having a discussion.
45:20
SPEAKER_02: It's so good to have the perspective like you do.
45:22
SPEAKER_02: Thank you.
45:23
SPEAKER_01: Absolutely.
45:23
SPEAKER_01: Well, thank you.
45:24
SPEAKER_01: I appreciate that.
45:25
SPEAKER_01: And the next time we'll have to do it over an adult drink, and it'll be uh the end of the conversation will be even more freely freewheeling, but we'll definitely want to have you back.
45:34
SPEAKER_01: And it's been a pleasure.
45:35
SPEAKER_01: And thanks everyone for listening.
45:37
SPEAKER_01: If this episode was helpful for you, uh please share it with someone who's writing AI policy or rolling out AI tooling inside a security conscious organization.
45:47
SPEAKER_01: Uh, and we'll also have information about where to find Rob in the show notes, and we'll also have information about our sponsor, which is Vector.
45:54
SPEAKER_01: Please check Vector out.
45:56
SPEAKER_01: And then we uh we also have mechanisms to assess your baselines and to look at your readiness.
46:02
SPEAKER_01: So keep your eyes open for that.
46:04
SPEAKER_01: This is also brought to you by Early Information Science.
46:07
SPEAKER_01: And uh again, my name is Seth Early, and we will see you next time with the next episode of Early AI.
46:13
SPEAKER_01: Thank you.
46:15
SPEAKER_00: Thank you for joining us on another deep dive into AI innovation.
46:19
SPEAKER_00: Tune in next time when we introduce another industry expert and discuss how to maximize AI in your world.
46:25
SPEAKER_00: The Early AI Podcast is sponsored by Early Information Science and Vector.
46:30
SPEAKER_00: To learn more, visit early.com.
46:33
SPEAKER_00: That's E A R L E Y.com.
46:36
SPEAKER_00: Thank you for tuning in.