Latest / Becker’s Healthcare Podcast / The Growing Trend of AI Agents in the Health System & What Leaders Can Do to Keep Operations Secure

Transcript

The Growing Trend of AI Agents in the Health System & What Leaders Can Do to Keep Operations Secure · Becker’s Healthcare Podcast

• 0:00
  - Hello, everyone. This is Erica Spicer Mason with
• 0:02
  - Becker's Healthcare. Thank you so much for tuning
• 0:04
  - in to the Becker's Healthcare podcast series.
• 0:08
  - Today, we're going to talk about the growing
• 0:10
  - trend of AI agents in the health system
• 0:12
  - and what leaders can do to keep operations
• 0:15
  - secure.
• 0:15
  - And joining me for this conversation, we have
• 0:18
  - Jason Elrod, a health care chief information security
• 0:21
  - officer and executive adviser to Elisity.
• 0:25
  - Jason, great to have you on the podcast.
• 0:26
  - Thanks for being with us today.
• 0:28
  - Fantastic to be here, Erica. Thanks for inviting
• 0:30
  - me. Oh, so glad to have you. And
• 0:33
  - I know AI agents is something that's becoming
• 0:36
  - more and more a part of health care
• 0:37
  - conversations, not just in the health IT space,
• 0:39
  - but really across the leadership board. So
• 0:42
  - wanted to make sure that we get ourselves
• 0:44
  - grounded first in what you're seeing happening here.
• 0:48
  - As I mentioned, we're seeing AI agents start
• 0:50
  - to show up inside of health systems today,
• 0:52
  - and I'd like to know from you what
• 0:54
  - that looks like operationally,
• 0:56
  - and how does an agent actually end up
• 0:58
  - with access to a health system environment?
• 1:01
  - Well, I think the important thing to understand
• 1:04
  - is that AI agent aren't necessarily
• 1:06
  - arriving as some large formally announced enterprise
• 1:10
  - deployment.
• 1:11
  - It's not like, hey, you know, the AI
• 1:13
  - cohort is here.
• 1:15
  - They arrive quietly.
• 1:17
  - You know, like a vendor adds an agentic
• 1:18
  - feature to a product we already use, a
• 1:20
  - business unit connects AI capability
• 1:22
  - to Microsoft three sixty five, Someone builds an
• 1:25
  - automation around an API. So
• 1:28
  - it's not quite the way that you, would
• 1:30
  - normally expect that. It's not something you see
• 1:32
  - it and you formally put it through. So
• 1:34
  - you do have that capability,
• 1:36
  - but
• 1:37
  - it's
• 1:38
  - less
• 1:40
  - structured.
• 1:41
  - Right?
• 1:42
  - And so so it's kinda sneaking its way
• 1:44
  - in. And I hate to say it that
• 1:45
  - way because there's
• 1:48
  - very
• 1:48
  - obvious ways that we're using AI and we
• 1:50
  - bring those in. But it's all those other
• 1:52
  - little areas that pop in. Right?
• 1:54
  - And from a security perspective,
• 1:56
  - AI is becoming much more interesting,
• 1:59
  - especially when it's moving from generative language
• 2:02
  - to exercising authority. So the difference between, like,
• 2:05
  - a large language model
• 2:07
  - and agentic AI. And that's a that's a
• 2:09
  - significant and very important
• 2:12
  - security boundary that we're seeing nowadays.
• 2:15
  - Yeah. Jason, I really appreciate the overview, and
• 2:18
  - I think you've mentioned some key points here
• 2:20
  - that I'm sure
• 2:21
  - as a CISO
• 2:23
  - raise some flags in your mind, you know,
• 2:25
  - when you think about a technology or or
• 2:27
  - a feature arriving quietly.
• 2:29
  - And in a sense, you know, feeling like
• 2:31
  - it's almost something that can sneak in under
• 2:33
  - the radar into a health systems operations.
• 2:35
  - So that kind of leads me to want
• 2:37
  - to touch on governance here.
• 2:39
  - I'm curious, you know, when we're thinking about
• 2:42
  - who owns an AI agent, who approves it,
• 2:44
  - who can turn it off,
• 2:46
  - you know, are are those things that
• 2:49
  - you advise organizations to take a look at?
• 2:52
  - Is it enough, or where do you think
• 2:53
  - security leaders should start?
• 2:56
  - I'll say governance
• 2:57
  - absolutely matters.
• 2:59
  - And it can't be governance
• 3:01
  - theater. It has the actual operational governance. What
• 3:03
  - I mean by that
• 3:05
  - is,
• 3:07
  - it can't be like the committee
• 3:08
  - approved the use of AI for some use
• 3:11
  - cases.
• 3:13
  - That is too vague. That it it's dilution
• 3:16
  - of responsibility,
• 3:17
  - dilution of accountability.
• 3:19
  - So real governance here has to be
• 3:22
  - operationally
• 3:23
  - specific.
• 3:24
  - What do I mean by that? Well, you
• 3:25
  - have to know who authorized this, who's accountable
• 3:28
  - for it, what data is it gonna access,
• 3:31
  - how are we actually monitoring it for impact
• 3:34
  - that it's gonna have, who's gonna be responsible
• 3:36
  - for that, And then how do we actually
• 3:39
  - enforce
• 3:41
  - containment if it behaves unexpectedly
• 3:43
  - or how quickly can we restrict or stop
• 3:45
  - it? And
• 3:47
  - from that perspective,
• 3:49
  - every use case or every AI that we
• 3:51
  - have, either we formally
• 3:53
  - introduce the organization or we discover,
• 3:55
  - needs to have an actual
• 3:57
  - name, like a person, not some sort of
• 4:00
  - committee because that dilutes it.
• 4:02
  - So, you gotta answer those five questions I
• 4:04
  - just sort of mentioned there. First, like, you
• 4:06
  - know, who authorized it? Who's accountable for it?
• 4:09
  - What data is it accessing?
• 4:11
  - Who's responsible for when it goes wonky? And
• 4:14
  - that's a technical term, by the way, when
• 4:15
  - it goes wonky.
• 4:17
  - And,
• 4:18
  - what are the tools we're using to support
• 4:20
  - that?
• 4:21
  - So you kinda could there's a couple of
• 4:23
  - components there. So governance is not just saying,
• 4:25
  - yes, you can do this.
• 4:26
  - No. So governance really answers the question, should
• 4:29
  - this agent exist? So do we know it?
• 4:31
  - Who's responsible? What are we gonna use it
• 4:33
  - for? What's it gonna access?
• 4:35
  - Security
• 4:36
  - architecture
• 4:37
  - needs to go along with that
• 4:39
  - and that answers how dangerous can this agent
• 4:42
  - become.
• 4:44
  - So the distinction there, governance absolutely needed. It
• 4:47
  - needs to be not governance theater, it needs
• 4:49
  - to be operational governance.
• 4:51
  - And governance in its essence tells an agent
• 4:54
  - what it should do.
• 4:56
  - You always have to link that with good
• 4:58
  - security architecture because architecture determines what it can
• 5:02
  - do, and that's a very important distinction.
• 5:06
  - Yeah. And I love this term that you're
• 5:09
  - using, Jason, the idea of governance theater.
• 5:12
  - Something I haven't heard before, but I've heard
• 5:14
  - so many leaders speak to the heart of
• 5:16
  - that issue.
• 5:17
  - You know, essentially saying that these governance
• 5:20
  - processes and protocols and account accountability
• 5:23
  - point people, it can't be you know, there
• 5:25
  - has to be backing behind it. It has
• 5:27
  - to be meaningful. So
• 5:28
  - appreciate you outlining those specific areas for listeners.
• 5:32
  - And I wanna take a step back here
• 5:33
  - too and go back to this idea of
• 5:36
  - when the AI agent enters a health system.
• 5:39
  - Health systems are incredibly complex environments in most
• 5:42
  - cases.
• 5:43
  - So how what can an AI agent potentially
• 5:46
  - reach inside of a hospital that might surprise
• 5:49
  - someone who's maybe coming from traditional corporate IT?
• 5:53
  - A hospital is an extraordinary
• 5:56
  - technology environment
• 5:58
  - because it's simultaneously
• 5:59
  - it's an enterprise, it's a clinical environment, but
• 6:02
  - it's also an industrial environment and increasingly
• 6:05
  - an IOT environment. We're gonna call it MIOT,
• 6:07
  - medical internet of things. So most corporate environments
• 6:12
  - don't have a CT scanner, an infusion pump,
• 6:14
  - a nurse call system,
• 6:15
  - and a building management system, and an EHR,
• 6:18
  - and a financial application operating in the same
• 6:22
  - ecosystem.
• 6:23
  - So it's
• 6:24
  - really interesting. So I think a lot of
• 6:26
  - folks don't necessarily take that into consideration. It's
• 6:28
  - a very blended environment.
• 6:30
  - So hospital isn't just an IT environment with
• 6:32
  - patients in it.
• 6:34
  - It's really
• 6:35
  - a clinical operating environment
• 6:37
  - powered by technology.
• 6:40
  - And we end up with a very interesting
• 6:42
  - space there because we can have some of
• 6:44
  - the most fantastic
• 6:46
  - sci fi latest technology in there, like surgical
• 6:49
  - robots as an example.
• 6:51
  - Next is some of the most
• 6:53
  - archaic
• 6:54
  - legacy
• 6:55
  - IT stuff,
• 6:56
  - you've ever seen. You did you know, you
• 6:58
  - saw it at Make Me Be in the
• 6:59
  - way back machine or inside of a museum.
• 7:01
  - You're like, wow. So I've got this really
• 7:03
  - old technology, this legacy in place
• 7:06
  - right next to things that are super sci
• 7:08
  - fi cutting edge. So you end up having
• 7:10
  - a little bit of a paradox
• 7:12
  - in the,
• 7:14
  - IT environment itself as well.
• 7:16
  - It makes a lot of sense as you
• 7:17
  - were walking us through that example, Jason. I'm
• 7:20
  - I'm thinking about the surgical robot, as you
• 7:22
  - mentioned, maybe
• 7:24
  - a room away from a fax machine, which
• 7:26
  - we know a lot of providers still use.
• 7:28
  - Yeah. It is it's crazy. Right? Yeah. Yeah.
• 7:31
  - Such a,
• 7:32
  - contradiction, a dynamic environment for sure.
• 7:36
  - So if there is risk around
• 7:38
  - what an agent can reach and also do
• 7:41
  - within this complex system,
• 7:43
  - how should health care leaders be thinking differently
• 7:45
  - about concepts like identity, permissions,
• 7:49
  - segmentation, and even limiting blast radius. Can you
• 7:52
  - shed some light on those?
• 7:54
  - Well, definitely. We we should begin by treating
• 7:56
  - agents as nonhuman identities
• 7:58
  - with potentially
• 7:59
  - significant privileges.
• 8:02
  - What I mean by that is the same
• 8:04
  - zero trust principles we apply to people and
• 8:06
  - devices
• 8:07
  - needs to extend to agents.
• 8:09
  - We don't just simply trust an agent simply
• 8:11
  - because it's internal, because a vendor built it,
• 8:13
  - or just because somebody approved it.
• 8:16
  - You give it minimum
• 8:17
  - necessary to perform its job just like an
• 8:19
  - individual. So this is why it's essential
• 8:22
  - that we give
• 8:23
  - AI, basically
• 8:25
  - agents, their own identity.
• 8:28
  - That way we can assign least privilege, basically
• 8:30
  - the minimum permissions necessary.
• 8:33
  - Segmentation
• 8:34
  - is intently
• 8:35
  - important here, so only in the environments required
• 8:38
  - for its function.
• 8:40
  - And I mentioned identity. You wanna have that
• 8:42
  - distinct identity with them. So I don't wanna
• 8:45
  - hide autonomous activity behind, like, a broad service
• 8:49
  - account. We can have, like, the AI identity
• 8:52
  - and all I all AIs get that one
• 8:54
  - sort of record or identity. We have to
• 8:56
  - give them individual ones just like people.
• 8:58
  - We have to be able to monitor it.
• 9:00
  - Right? You know, we gotta be able to
• 9:01
  - understand, like, what are they doing to the
• 9:04
  - best extent that we can.
• 9:05
  - And we talk about segmentation. Right? Containment.
• 9:09
  - You don't
• 9:10
  - secure
• 9:12
  - autonomy
• 9:13
  - by assuming
• 9:14
  - perfect behavior.
• 9:16
  - You secure it by limiting the consequences
• 9:19
  - of imperfect behavior.
• 9:22
  - And honestly, complex systems fail.
• 9:25
  - And credentials get compromised, applications malfunctions.
• 9:28
  - Humans make mistakes. And agents will sometimes behave
• 9:31
  - in ways we did not anticipate.
• 9:34
  - Then you're looking at some of the, you
• 9:36
  - know, frontier models from some of the big
• 9:38
  - providers out there where they did not anticipate
• 9:40
  - them breaking out and then going hacking another
• 9:42
  - company out there.
• 9:43
  - So you can see that there is going
• 9:45
  - to be a spot where something doesn't act
• 9:47
  - like you thought it would act.
• 9:50
  - That's where the segmentation comes in, especially identity
• 9:52
  - based,
• 9:53
  - segmentation. You wanna be able to help you.
• 9:55
  - Whether it's a human or nonhuman identity,
• 9:58
  - you wanna say, okay, I've authorized you to
• 10:00
  - do this.
• 10:01
  - And so when I've authorized you to do
• 10:03
  - something and you get into that system to
• 10:05
  - do it, you know, I can't really
• 10:07
  - control
• 10:08
  - so much.
• 10:10
  - The damage can be done through that explicit
• 10:12
  - authorization.
• 10:13
  - But what I can do is I can
• 10:14
  - segment that to just that system. You can't
• 10:17
  - take that authorization and then expand it across
• 10:19
  - systems that are not involved in that idea
• 10:22
  - of minimum necessary.
• 10:24
  - So again, I mean, you don't you don't
• 10:26
  - secure autonomy by assuming perfect behavior because nothing's
• 10:29
  - perfect and it will fail. You secure it
• 10:32
  - by limiting the consequences
• 10:34
  - of when it breaks bad or that imperfect
• 10:36
  - behavior.
• 10:38
  - Mhmm.
• 10:39
  - Yeah. So it's sounds like segmentation is essential.
• 10:43
  - And Credible. I'm wondering, Jason yeah.
• 10:46
  - And I'm wondering,
• 10:48
  - even though this concept,
• 10:50
  - it doesn't sound new, but it it does
• 10:52
  - sound potentially challenging for organizations to scale or
• 10:55
  - to execute it at scale. So I'm curious
• 10:58
  - what your view is on that and how
• 11:00
  - you're seeing organizations
• 11:01
  - make meaningful progress in achieving segmentation.
• 11:05
  - In the past, the technology was the problem.
• 11:08
  - You couldn't really do it at scale. So
• 11:10
  - but but today, the technology usually isn't the
• 11:12
  - hardest part.
• 11:13
  - The organizations
• 11:14
  - that succeed
• 11:16
  - succeed in this space, actually stop treating segmentation
• 11:19
  - as this giant network project and start treating
• 11:21
  - it as a risk reduction program. So you
• 11:23
  - can't say, I need to go in and
• 11:24
  - do it everywhere all at once.
• 11:27
  - You know, trying to trying to redesign an
• 11:29
  - entire health system network at once is enormously
• 11:31
  - difficult and just thinking about that can cause
• 11:34
  - pause or stop it, the work of it.
• 11:37
  - A better approach is to understand the environment,
• 11:39
  - identify those meaningful boundaries like we just talked
• 11:42
  - about a little bit there, and then progressively
• 11:45
  - reduce unnecessary
• 11:47
  - connectivity. And what I mean by that, again,
• 11:49
  - reduce the scope.
• 11:51
  - Segmentation isn't about building walls everywhere.
• 11:54
  - It's about removing
• 11:55
  - pathways that have no reason to exist.
• 11:58
  - That's the idea in the past you did
• 12:00
  - this over provisioning. You had everything go and
• 12:02
  - just
• 12:03
  - I'm thinking back in the way back machine.
• 12:04
  - You just put it on the flat network
• 12:06
  - and you gave it advent prevent credentials and
• 12:08
  - you were just able to do your work.
• 12:10
  - Well, nowadays that that's obviously a recipe for
• 12:12
  - disaster and obviously not where you wanna go.
• 12:15
  - But most
• 12:17
  - health care or health IT
• 12:18
  - ecosystems
• 12:19
  - didn't grow up
• 12:22
  - with a real coherent plan.
• 12:25
  - So they're not homogeneous,
• 12:27
  - heterogeneous. Right? They're they're they're completely diverse. I
• 12:30
  - mentioned before, old technology,
• 12:32
  - cutting edge technology
• 12:33
  - right next to each other. So the first
• 12:35
  - thing you need to do is you need
• 12:36
  - to start visibility. Right? You need to know
• 12:38
  - what you have. Understand what the devices are.
• 12:41
  - What's it communicating with? Why does that communication
• 12:43
  - exist? Because sometimes we had those the communications
• 12:46
  - were there,
• 12:47
  - for a reason in the past, but they
• 12:48
  - don't necessarily need to be there now.
• 12:50
  - And
• 12:51
  - what happens if that communication stops? So getting
• 12:54
  - that higher fidelity about what's really going on
• 12:56
  - in your network, extremely important.
• 12:59
  - Then
• 13:01
  - move deliberately,
• 13:02
  - not
• 13:04
  - holistically.
• 13:05
  - And what I mean there is you go
• 13:07
  - from visibility to understanding to policy to enforcement.
• 13:10
  - That's sort of the path you have to
• 13:12
  - do. Like you know what I got,
• 13:13
  - what what it should be doing,
• 13:15
  - what's a rational policy now based on that,
• 13:17
  - and then how do I enforce it? That's
• 13:20
  - the that's the macro of how you do
• 13:21
  - it. But don't do it everywhere all at
• 13:24
  - once.
• 13:25
  - What you have to do is you, hey,
• 13:26
  - hey, what are the highest impact items? So
• 13:28
  - and then go into that and you you're
• 13:30
  - gonna go ahead and do it incrementally
• 13:32
  - in that fashion.
• 13:34
  - And since we're talking about AI agents,
• 13:37
  - you wanna look at that from, again, the
• 13:38
  - identity. This agent performed a function
• 13:41
  - in this environment
• 13:43
  - and has no reason to communicate outside that
• 13:46
  - boundary. So like, oh, I might use it
• 13:48
  - for some for something else later. Great. That's
• 13:51
  - a future Us problem and a future Us
• 13:53
  - opportunity.
• 13:54
  - Don't
• 13:56
  - build in that
• 13:58
  - wide boundary.
• 14:00
  - Build in the constrained boundary.
• 14:03
  - Because, again, I mentioned it's about it's not
• 14:05
  - about building walls everywhere. It's about
• 14:09
  - removing the things that have no reason to
• 14:11
  - exist.
• 14:13
  - And you don't really know that if you
• 14:14
  - don't have that visibility. You don't understand the
• 14:16
  - identity. You don't understand the communications
• 14:18
  - and then the context in which that needs
• 14:20
  - to operate.
• 14:21
  - Mhmm. That helps? It it helps a lot,
• 14:23
  - Jason. It almost reminds me of the concept
• 14:26
  - of kind of harm reduction,
• 14:28
  - but everything you're saying
• 14:30
  - is so clear.
• 14:31
  - And I'm curious too. You know, you mentioned
• 14:33
  - that you advise leaders oftentimes
• 14:36
  - if they're wondering where to start, you advise
• 14:38
  - the highest impact use cases, you know, not
• 14:40
  - taking everything on all at once to achieve
• 14:43
  - this environment.
• 14:45
  - But say a health system leader who's listening
• 14:47
  - to this right now has a short amount
• 14:48
  - of time and essentially no new budget to
• 14:52
  - get started down this path. What are the
• 14:54
  - first two or three things that you would
• 14:57
  - tell them to do to reduce AI agent
• 14:59
  - risk?
• 15:00
  - Before you buy another AI security product,
• 15:03
  - understand the identities, the permissions, and the controls
• 15:07
  - you already have.
• 15:08
  - That's key.
• 15:10
  - So
• 15:11
  - when you find the agents,
• 15:13
  - whether you again, you intended them to be
• 15:15
  - there or you just kind of were gifted
• 15:17
  - them because of shadow IT or shadow AI.
• 15:20
  - Right? You find them, understand where they are.
• 15:22
  - So inventory them wherever
• 15:24
  - they are and the agentic and wherever agentic
• 15:27
  - features exist, which vendors have added them, which
• 15:29
  - is the sneaky part getting in the organization.
• 15:31
  - You might have had something in your organization
• 15:33
  - for a long time and all of a
• 15:34
  - sudden the next released, boom. Now you have
• 15:37
  - an AI agent inside of it. So understand
• 15:40
  - that. What automations exist?
• 15:42
  - Which APIs they use? How are they talking
• 15:45
  - back and forth?
• 15:46
  - And I'll go back to that initial
• 15:48
  - identity.
• 15:50
  - What identities they operate under? So if it
• 15:52
  - says they can't operate under
• 15:54
  - my identities because I created it. They have
• 15:56
  - to have their own identities. And when you
• 15:58
  - discover agentic activities in your organization
• 16:00
  - that are operating
• 16:02
  - hidden or underneath
• 16:03
  - the umbrella of somebody else's identity, start there.
• 16:06
  - Start separating those. So again, yeah, before you
• 16:08
  - buy another AI security product,
• 16:11
  - understand the boring basics here. Understand the identities,
• 16:14
  - the permissions, and the controls you already have.
• 16:18
  - And I'll I'll one more.
• 16:20
  - Anytime you implement an AI,
• 16:22
  - always ask this critical question or have this
• 16:25
  - on that list.
• 16:27
  - How would I shut it down right now?
• 16:30
  - What's my exit criteria from this?
• 16:34
  - See,
• 16:35
  - that key questionnaires is not how I got
• 16:37
  - in the environment, how I started using it.
• 16:39
  - It's how do I stop it? How do
• 16:41
  - I shut it down if it starts to
• 16:43
  - break bad quickly?
• 16:45
  - Mhmm. Because nobody knows that answer.
• 16:48
  - Honestly, that's something you can fix without a
• 16:51
  - multimillion dollar project.
• 16:52
  - Just go through your system and say, how
• 16:54
  - can I turn this off? How can I
• 16:55
  - shut it down immediately
• 16:56
  - if I needed to?
• 16:59
  - Yeah.
• 17:00
  - It's a really
• 17:02
  - it's such a simple but powerful
• 17:03
  - concept, Jason. How would I shut it down
• 17:05
  - right now?
• 17:06
  - An essential question.
• 17:08
  - So I I really appreciate you raising that
• 17:10
  - for our listeners.
• 17:11
  - And, also, I I thought that your pragmatic,
• 17:15
  - recommendation to just understand the boring basics, quote,
• 17:18
  - unquote, you know, that that's another
• 17:20
  - essential piece of this.
• 17:22
  - Jason, it's been such a great conversation. I
• 17:24
  - know we only have a few more minutes
• 17:25
  - together, but I wanna give our our listeners
• 17:28
  - just
• 17:29
  - kind of a forward looking view of
• 17:31
  - where this approach that we've talked today, whether
• 17:34
  - that's segmentations,
• 17:35
  - creating, you know, constrained or contained environments,
• 17:39
  - Where might that approach stop helping, and what
• 17:42
  - risks would you leave listeners with to consider
• 17:45
  - even after they've achieved the constrained access and
• 17:48
  - segmentation?
• 17:50
  - Segmentation is critical. Right? We we've already mentioned
• 17:52
  - that. Lease privilege, again, also helps dramatically reduce
• 17:55
  - risk. So segmentation on lease privilege dramatically reduces
• 17:58
  - risk. But what they don't do is they
• 18:00
  - don't make an agent trustworthy.
• 18:02
  - They limit the consequences when something goes wrong.
• 18:06
  - There are still entire categories of AI risk
• 18:09
  - that must be addressed
• 18:10
  - separately.
• 18:12
  - Here's the thing. The objective isn't to make
• 18:14
  - AI incapable of failure.
• 18:16
  - It's gonna fail.
• 18:18
  - I mean, because that's that's completely unrealistic.
• 18:20
  - The objective
• 18:21
  - is to make failure observable,
• 18:24
  - containable,
• 18:25
  - and recoverable.
• 18:27
  - So I think that's where the focus needs
• 18:29
  - to be. There's no single control for AI
• 18:31
  - risk. We need governance, identity, architecture, security testing,
• 18:34
  - all that stuff. Human oversight, you know, how
• 18:36
  - do I turn it off?
• 18:37
  - Segmentation, special values, it creates a boundary around
• 18:40
  - the failure. And that I think is the
• 18:41
  - most
• 18:42
  - important
• 18:43
  - because you're still gonna have those risks. If
• 18:45
  - I authorize
• 18:47
  - an agent to get into
• 18:48
  - our EHR as an example to do things,
• 18:51
  - I've already authorized it. I can't that no
• 18:52
  - amount of segmentation is gonna help that because
• 18:54
  - I've segmented it down just you can access
• 18:56
  - the EHR. So what's it doing inside the
• 18:58
  - EHR? Right?
• 19:00
  - These are
• 19:01
  - risks that remain. So your authorization risks, goes
• 19:04
  - back to that least privilege,
• 19:06
  - those risks are still there. What do you
• 19:07
  - authorize it for? How do you identify it?
• 19:09
  - And, again, that could that question I asked
• 19:11
  - there,
• 19:12
  - how do I stop it? How can I
• 19:14
  - stop it or constrain it if it does
• 19:16
  - break bad,
• 19:17
  - and has impact patient care or organizational operations?
• 19:21
  - So
• 19:22
  - you can only solve so much.
• 19:25
  - Yeah. Jason, I I just wanna thank you
• 19:27
  - for
• 19:28
  - everything you've shared in this conversation, a lot
• 19:30
  - of
• 19:31
  - actionable
• 19:32
  - steps and considerations that you've left with listeners
• 19:35
  - today,
• 19:36
  - And just wanna thank you for raising attention
• 19:39
  - and shedding light on a a topic that's
• 19:40
  - very important as we see greater adoption of
• 19:43
  - of AI agents. So thank you again for
• 19:45
  - spending the time with Becker's today.
• 19:47
  - Oh, it's been my pleasure, Erica.
• 19:49
  - And we'd also like to thank our podcast
• 19:51
  - sponsor for today, Elisity.
• 19:53
  - Listeners, be sure to tune in to more
• 19:55
  - podcasts from Becker's Healthcare by visiting our podcast
• 19:58
  - page at beckershospitalreview.com.