Martin Hinton (00:02.406)
Okay, that's going.

Martin Hinton (00:34.822)
Just move the note around.

erin kenneally (00:53.154)
Got two other dogs joining the brigade here. This is gonna be interesting. Well, I've got one, but the neighbors' dogs just decided to cruise over and now they're all Yeah, I yeah. So who knows what's gonna happen.

Martin Hinton (00:57.39)
How many do you have?

Martin Hinton (01:03.432)
come in to watch the show, right? Yeah. Okay. All right. Well, well let's give them something to let's give them something to listen to. All right, here we go.

erin kenneally (01:10.176)
Okay.

Martin Hinton (01:23.698)
All right then. Welcome to the Cyber Insurance News and Information Podcast. I'm the executive editor, Martin Hinton and your host today. Joining us is our return guest, Erin Kenneally who is the co-founder and CEO of dltaskop.ai is joining us for the first time since last October. And last October, she came on to talk about AI risk and the potential of AI to repeat the ransomware issues the insurance in industry faced. And Erin, thanks so much for coming back. How's your day gone so far?

erin kenneally (01:52.898)
Going well. Thank you. Thank you. It's great to be back.

Martin Hinton (01:55.64)
So I I I looked at the podcast from a year ago, almost exactly a year ago. We we released it on October twenty-fourth. And one of the things you said on it is this, and I quote, people, myself included, perhaps underestimate how difficult it is to determine if and to what extent AI is and will be the cause of some of these losses because it's so embedded. So twelve months on from that prediction, that line, how do you think it stands up?

erin kenneally (02:22.636)
You know, I I think it stands up pretty well. that initial study that I had did had done, excuse me, back in I wanna say spring of twenty five, was a descriptive, you know, study of and a little bit of of predictive analysis in terms of the state of the market as it were, and then the retrospective about a year later.

and look the the initial premise, and I think we covered this on our podcast a year ago, was that there was just coverage control and measurement and modeling gaps in the market. And at that time, you know, you you'd get a a little bit of a different response from carriers and brokers in terms of the state of affairs. Some said, you know, nothing to see here, it's all covered under cyber. others would

Quietly tell you that they were very concerned because there's a ton of unpriced risk sitting on books. and then sure enough, you know, circa January of this year, right, with the ISO Varisk CGL exclusions coming out, we saw a lot of a lot of those exclusions getting filed in various states, which I should I should note. and I think some people, certainly people in the industry realize this, but others.

probably don't realize that just because it's file file doesn't mean that it's been implemented. So in terms of what's actually getting excluded from an AI perspective, and what what claims are coming in and and getting paid or not, is a is another story. But the point being that the extent to which the market is actively and definitely I would say there's been a market

market increase in the explicit discussions, whether it's in the mainstream media or coming f from carriers themselves about AI exclusions, you know, I think that that speaks volumes.

Martin Hinton (04:34.226)
Yeah, so you get the th th that's I think it's two essays you've got out this year. One one was in April, if I'm not mistaken. But the argument along those lines is that is that

Martin Hinton (04:46.408)
You've got two new essays out that are that are quite recent. One of them is that that well, they're about the AI exclusion thing. And the idea is that AI is being written out of policies faster than it's being written in. And until there's governance and the ability to observe and prove what AI can do and then what it actually did in any kind of incident, it's really, really important that that those tools come into play. tell me a little more about that.

erin kenneally (05:12.319)
Yeah, I I mean I think the from a high level perspective, the takeaway from that messaging is that and this is actually the genesis of of my company. Most companies haven't bought AI specific policies, right? They they've simply unbeknownst to them lost protections that they assumed that their either their general liability or their cyber policies provided. So and often without notice. and so it's it's this classic

shadow AI problem that's that's impacting the industry. So and I just lost track of your your your question, sorry.

Martin Hinton (05:51.634)
No, but no, well I well well let's just let's take that topic right right there. The the the shadow AI idea. You know, this idea that that you know if you're insuring a company with a thousand employees and you know they all have the capacity to use AI in their personal lives, that they may drift across that sort of, you know, if you will, the old classic church and state sort of barrier that would exist between work and and h home life. We we s we saw that initially

Sort of that barrier dissolved with COVID in a large way and the the remote work reality that came into a lot of places. And now AI's come along as this brilliant tool that, if incorporated with governance and guardrails and all kinds of checks and balances and and clarity about its use and its touch and reach, whether it's agentic or otherwise, a company can say to its insurer, Hey, listen, this is how we're using it. We obviously know that the people

stray outside the margins and the barriers and guardrails are even on the highway not entirely effective at keeping cars on the road. T tell me about that part of this. I mean, how how much of that is, you know, still being assessed? Do we do we have any real idea about its p potential impact at this point?

erin kenneally (06:57.974)
Yeah, I mean I think it might be helpful to think of how the and this was a function of the of the retrospective study that was done a year after the initial artificial insurance paper. So let let's maybe start by understanding how the market responded to both the concrete and the and fears of AI risk being manifest

and becoming you know coming out of the shadows so to speak and from a descriptive standpoint i you know kind of bottom up what the trend that we saw happen in that in that year I would say between what was it May of twenty-five through May of of twenty-six is it it kind of clustered around four tiers, right? So

You had what I would call legacy exclusions. So these were traditional policies that r removed AI exposures that wasn't, you know, clearly contemplated or priced. and and I I I I should make a finer point that, you know, look, insurance is in the business of of making money, right? It's not a charity. So it it could very well be a rational risk management decision.

to write AI exclusions. But we have to realize that it's, you know, it's an ecosystem, right? So it leaves when you write exclusions, so those CGLs are a great example, it's going to leave the buyer of insurance needing some sort of replacement for that protection. Whether it be in a cyber depending on the scenario, E and O or or whatnot. So

So we started to see these exclusions creeping in when carriers admitted that look, we just have in

Martin Hinton (08:54.376)
So so to Er Erin, just one second. You mentioned one of them. Do you wanna do just rattle off the four that you mentioned you touched on there being four. Do you wanna say what the four are and then we can go through them one by one?

erin kenneally (09:02.006)
Sure, sure, sure. So there was so I call it legacy exclusions. Then there's what we would think of as specialty or standalone AI policy, so specialty affirmative coverage for AI risks. and then there was there is coordinated or or what I would say technically evaluated underwriting, and then there's this infrastructure capacity, which is frankly orthogonal, but there was a lot of activity there. And I I'll just give you a quick

drill down into each one of those. So we talked about legacy exclusions. The specialty affirmative coverage, right? These are products that arose that explicitly address and defined AI related losses. So these were, these have been and they're they're useful advances. Don't get me wrong, they're usually pretty narrow and the limits are are can be rather low relative to the risk at hand.

but frankly, rather than kind of looking at the AI label, meaning standalone AI, the real kind of devil in the detail is is understanding how those policies are written from a a trigger perspective, from an excl you know how they define those exclusions and what are the actual limits and retentions and coinsurance associated with that, right? So and that's a small part of the the market.

Then there's this coordinated or or technically evaluated underwriting, we'll call it. and this comes about, I think, in kind of one of two ways in general. and and really what those are about is are ways of connecting technical distinctions between like cyber and AI and the and how contracts and coverages really respond. So, in one regard,

The products you could have like an AI wraparound or an AI liability layer around a cyber or tech ENO policy. And then oftentimes with that, there's some sort of like a predefined allocation rules for how those losses are are distributed between those policies. And then there's also products that are linking of the coverage and the protection to defined AI performance failures and

erin kenneally (11:26.452)
measurable performance data. so those would c would be kind of two buckets there. And then this infrastructure category is is really more traditional insurance property underwriting where insurance is supporting the the the data centers, right? The the buildings and the equipment and the construction and you know all those physical risks that are surround the AI.

in in these data centers, but the point there is that, you know, s to keep it clear, insuring the physical data center is really not the same as insuring what comes out of it. And and you know, certainly that's important, but they're very different beasts in terms of assessing a a physical risk to a data center that harbors AI and insurance has

a long history from NATCAD and and whatnot, and and property underwriting to handle that, versus a very new game, which is ensuring the the models and the outputs of the AI itself.

Martin Hinton (12:39.808)
Yeah, I mean it would seem that the the the the the last of those categories, the the the building that the the data center sits in, if you will, that is the low-hanging fruit when it comes to the problem we're facing here. So you coded a bunch of incidents and the the the the phrasing that I I have in my notes here is the 92% problem. tell me about that.

erin kenneally (12:49.526)
Absolutely. Absolutely, yeah.

erin kenneally (13:01.41)
Sure. Yeah. So what I did was I took the, so there's a a number of databases that record AI incidents, right? So I the one of the big ones that's often cited is this AI ID database. So the AI incident database. I think it is administered partially or entirely out of MIT. what I did is I took, so on the heels of the first report in the retrospective, I wanted to look at, you know, that.

years worth of publicly reported incidents and and see what it said. And then, you know, I coded those for, you know, what what's the category of risk, what was the reported severity. And and importantly, what kind of I added to the mix was the applicability of our standard familiar in the cyber realm, the standard pre-breach cyber controls, just to see how that mapped to some of the things that I, you know,

I loosely say predicted in that in that first paper. and you know, basically what we found was that from a control perspective, about ninety-two in in about ninety-two percent of those cases, there was no standard cyber control that would have been helpful. Meaning it wouldn't have plausibly kind of interrupted or materially reduced.

the harm that was described in that in that incident.

Martin Hinton (14:31.346)
So so if I'm if I'm understanding that correctly, what we're talking about now is that the premise to begin with was the idea that it's gonna be very hard to ensure this until we can assess the risk, create the guardrails. And what you're saying, if I'm understanding you, and please tell me if I'm wrong, is that nothing that existed now could have stopped those things from being a problem, right? So we need a new a new method, a new tool, a new thing needs to come along to deal with this.

erin kenneally (14:56.68)
Yeah, so I would r I would phrase it. You're you're conceptually correct there. look, we can't our traditional cyber control pre breach checklist isn't gonna be sufficient sufficient proxy or sufficient in and of itself for AI risk management, right? I mean, there's this whole industry, nascent industry of AI risk controls. I d I forget the numbers in terms of how big that how large that industry is from a a billion plus perspective, but

I mean, that speaks volumes from a security standpoint relative to the efficacy of security controls to meet the problem of of AI risk. So I'm not meaning to say that, you know, cybersecurity is useless by any stretch, but we need to have a new pre-breach control stack. And the the one the second paper that I had written, now I'm blanking on the name of it, controls bigot coverage, I think, was

Really an attempt to lay out what what's the control stack that insurers should be looking at when they're underwriting and assessing and and pricing these risks. And and maybe to put a finer point on it, the reason why the cyber control stack is not helpful in these instances, and it may be obvious, but

They address and answer a different cr question, right? So think of things like MFA. It it helps establish identity, right? patching. That helps address known vulnerabilities. EDR is another tool that's in that that pre-breach stack that insurers would would look for in their underwriting of cyber risk. And that looks for particular forms of malicious activity, we'll call it. But none of them really established

a s are able to establish that an authorized AI generated decision is accurate, appropriate, or within authority. And though and it's it's those causes of loss from hallucination to adversarial attacks to autonomous activity that are causing the losses in an in an AI from an AI perspective. So so like if you've got a fabricated citation and this is well known in the in the you know legal world like

erin kenneally (17:21.772)
That's clearly generated without without having a breach. So that's not gonna trigger a cyber policy. an agent itself can make an operational error within a session that is fully authorized by the company that's deploying it. And you know, some sort of synthetic impersonation, deep fake, can cause a loss without compromising a network. And so all the cyber controls are really kind of anchored around.

this this unauthorized access and breach of of network access. But that's not where the loss is flowing from for the most part in the AI perspective. Now just real quick, certainly the extent to which AI is accelerating the you know conventional causes of loss in the cyber perspective, then yes, those those, you know, a lot of those access control type

controls in the pre-breach stack would be applicable. So let's not throw it out, but we also have to realize that a new control stack is needed.

Martin Hinton (18:26.98)
I mean I mean I mean I was gonna ask a different question, but one of things that you you said now to to my fifth grader mind, the existing problems were largely external. The word breach matters, right? Something has to penetrate your network, or you have to let someone in through clicking on an email or whatever it might be, react to an invoice that's you know a business email compromise attempt. The failure here is sometimes internal and thus owned, right? It's almost like a product liability issue where, you know, let's say

i it were a car manufacturer and they were making a car that had a flaw that was causing accidents, right? And that's certainly been there's a history for that sort of thing. Is that where the sort of, you know, looking for the risk and and thus the cause of the liability and then the cost of the liability, there's almost a you know, beyond the technical which we've touched on and the the stack to prevent these things, there's a different way to look at the risk this creates. Cause i I mean, am I understanding that correctly?

erin kenneally (19:21.548)
Yeah, and and I think the way I framed it, and I forget which in which paper I I teed this up, but it's this notion of the the three is the is the framing I use, not for sake of coming up with yet another framing, but it introduces this this it helps characterize what this new risk is. And so the three is misuse, misalignment, and mistake as a kind of a high-level abstraction of how to think about the the causes of

Of AI harm that then have implications for coverage and controls and whatnot. Right. So we're familiar with I'll I'll break them down. So misuse, right? So this is when someone would deliberately use or manipulate the AI system and produce some sort of unwanted outcome. Right. And this is this should be very familiar. These are things like you know, adversarial attacks and very akin to a lot of the the cyber attacks that we're used to.

The from a mistake perspective, this is where the system, the AI system, or you know, even the human who is deploying and configuring the AI system does something and just simply gets it wrong, makes a mistake. And we're all familiar with that as well. The the new on the block, so to speak, would be this misalignment, where the the system itself is pursuing an objective.

the objective function in a way that is really diverging from what the humans, the companies, the deployers, the users, what they intend or the the purpose that for which they wanted to use the system. and so that is you know I think a difference with distinction certainly. and and from an insurance perspective, you know, they're they're behaving differently, right? So misuse is going to tell you

you know, you're gonna wanna know like who influenced the system, how was it influenced? was there unauthorized access? You know, things like prompt injection failure modes and credential abuses and malicious tools and whatnot. So I think everyone can kind of understand that relatively easily. the mistake in terms of understanding from a causal perspective, an an underwriter needs to understand that the system failed at the task itself. And this could be through hallucination or

erin kenneally (21:41.346)
you know, some sort of miscalculation or operational failure. And then the misalignment gets sticky, right? Because it's coming from within the model, which is normally opaque. And certainly if you're dealing with a foundational model and you don't know the innards, you don't know how it's been parameterized or or how those decisions are weighted. I mean, this is, you know, for lack of a better description, you know, it's like understanding a it's the neural network.

d we don't even understand how our brains work yet. These these models are are, you know, somewhat, you know, modeled, no pun intended, after after that. so, you know, understanding how the the model is optimizing for the goal that the human programmed is the big problem, at hand. And and yeah, that's I I think that that that's just

Martin Hinton (22:35.098)
Is i are are are are any of those harder to price or does one show up first in claims? you know, is there a hierarchy to the complexity from the insurer point of view?

erin kenneally (22:47.224)
Frankly, and you know, I think most folks who are knowledgeable about this would probably agree that misalignment is is probably the hardest to price. for a couple reasons. One I just alluded to is that it's it's very tough to measure and and and model that all the different scenarios, states.

That misalignment can be manifest through. But the another nuance there as well that's concerning is the apparent success of the model in optimizing can also be the pathway to the loss itself. right, so here's a simple example. GPS. If the if the optimization goal

Of the GPS of the AI-driven GPS technologies to get from point A to point B in the shortest amount of time. Yet the map takes the the drivers over a cliff. it's actually fulfilling its purpose, and it's successful in that regard, but it also becomes the pathway to the loss. and so it's it's it's kind of interesting how how that plays out.

from claims perspectives and from an underwriting perspective, obviously, in terms of

Martin Hinton (24:16.7)
I I I go go ahead, pardon me.

erin kenneally (24:19.586)
Yeah, no, no, I just gonna answer that second part. I think you asked about what's going to show up in claims. caveat here, you know, I'd separate out what I would expect from what the data proved, right? I don't I don't think we have a sufficiently mature body of AI specific insured claims. We certainly have a growing corpus of AI incidents, which is great and super helpful, but

But that's different than insured claims that say empirically what what's gonna come first. I my my expectation though is that I think the the mistake and the misuse causes of loss are gonna appear first under kind of familiar claim descriptions, right? Because they're gonna they're gonna resemble losses that insurers already understand, fraud, professional errors, you know, technology failures, things like that.

Martin Hinton (25:15.668)
Exactly. Yeah. Exact yeah. I mean that th that's a junior lawyer using it to help with a brief and then not proofing the brief effectively enough, or perhaps not at all, or or whatever, you know, any anything, a a memo, a press release, anything that, you know, may not be a huge financial hit, but if there's a a failure to put the wording correctly or have it say what you intended it to say or hoped it would say, those are those are problems and problems are what you have insurance for in in some respects.

erin kenneally (25:43.22)
Absolutely. No, absolutely. And then I think that that misalignment is definitely gonna creep in. I mean, we've already seen it, right? So it's probably gonna arrive disguised as something else to begin with, like some sort of tech EO event or operational error or management failure or whatnot, only to later learn that it was it was actually the cause of of a of a model misalignment. and you know that

Martin Hinton (26:07.272)
Mean the y you you you you touched on the misalignment and it is this is the you know it you use the GPS analogy, the famous one you hear often is the paper ch the paper clip optimizer. You tell an AI to make as many paper clips as possible and you don't give it parameters for what it shouldn't do in trying to achieve that goal and it will eventually start taking, you know, your grandpa's titanium knee, never mind your grandpa, and turning it into paper clips, right?

everything for the for the paperclip cause. It's it's very Machiavellian. These concepts aren't lost in us, right? The ends justify the means. And if you don't realign and if you read The Prince by Machiavelli, it's not all about being a brutal prince. There's an enormous about it. Like you you you gotta have the people on your side, right? You you and and that whole for me the alignment thing is well aligned with who? I mean human beings haven't been aligned since we discovered a stone was better with a sharp edge on it. And we started sharpening our stones.

erin kenneally (26:56.717)
Absolutely.

Martin Hinton (26:59.124)
So it is a i it's a fascinating problem. And and when you put it in the hands of a company, the the sprawl it can create. I mean, we hear about sprawl in all sorts of senses in the in in the cyber insurance world is is significant. I you you touched on this and I I would wonder if you could explore this idea a little more for me. That the telltale, if you will, the canary in the misalignment coal mine is gonna appear as something else. And I think you said ENO. Tell me a little more about what you think.

you know, we we could see there first and what might be an indicator for people who are paying attention that that it's arrived in a real way and there's now, you know, from a positive point of view, an event to use to assess risks and help price policies or at least, you know, make the right decisions about how to move forward with with coverage.

erin kenneally (27:43.596)
Yeah, I mean, first off, you know, one of the one of the ch challenges is this notion of separating out and and this is a little bit of a sideways way of answering your question because it I think this needs to be understood in order to kind of figure out where how the claims are how the coverages are gonna play out and how the claims are are gonna be assessed.

and and the issue is with agentics specifically, because they're off they're acting in an autonomous or pseudo-autonomous manner. the the concern here is that oversight of the model should be more than just trusting what the the reasoning process of the model says, right? And we've seen this in in instances in the last several months with the PyPie.

incident that you talked about as well as with the the hugging face incident right and so we I think I think what's h what's important to understand is that or differentiate is separating out functions in order to understand outcomes, right? So there's there's visibility into the reasoning of the model.

What does the model appear to be doing? Then there's this separate piece, which is, and you you're hearing about it a lot, which is this notion of runtime enforcement at the at the agentic level. and that's the enforcement authority, right? What what is it actually permitted to do? So there's something that it's trying to do probabilistically, what's it permitted to do? That should be deterministic. And then there's the evidentiary perspective, which is what it actually did do, because we know that.

models can fabricate evidence and lead investigations astray and whatnot. So you know I think from and we all this is another term we hear a lot is this notion of governance. what does that mean? I think we need to unpack this this notion. and it's not just a paper policy but it's gotta move inside the enforcement authority. So you can't have the fox guarding the hen ho h hen house

erin kenneally (30:09.55)
from a you know the the standpoint of oversight and understanding the the risk dis and and predicting the risk distribution and severity.

Martin Hinton (30:21.074)
You so we we're we're touching on your second paper that you you referenced and and the the title of it i is governance controls begets coverage, the agentic AI insurance stack. And one of the things you you did in it, if I'm remembering correctly, is you compared the agentic governance, to use your word, to air traffic control with gaps in the radar. Tell me a bit more about that.

erin kenneally (30:45.422)
Sure, you know, so the the the analogy kind of maps in this regard. So you can think of radar and surveillance as the discovery and the observability of the AI, the agents. you can think of the flight identification or the transponder that that allows you to do such as the agentic identity and attribution of the agent, the non human actors. You can think of the flight.

plan and sort of the the clearance around flight plan and paths as the intended authority. then there's this notion of the controllers instructions which I would equate to in the agentic world runtime governance. and then the final two would be you know within this this bounded area of ATC you've got restricted airspace and you know kind of

procedural rules around that. And I would consider that to be the equivalent of like policy boundaries from an agentic perspective. And then from an evidentiary perspective, you got, you know, flight recorders and and radar history and and and specimens such as that, which would equate to the evidence and and the reconstruction. So the it's it's an imperfect analogy, right? Because from an ATC perspective, there's really no

Hard runtime enf enforcement, right? The the the controller can do a lot to kind of bound the aircraft, but it can't reach into the cockpit and you know physically prevent the aircraft from entering certain areas or or engaging, you know, in certain spaces. But this is where from an agentic perspective, we can actually potentially go one step further there. and and specifically, you know, we can intercept tool calls, we can elevate authorities, we can

instrument that authority authority enforcement control to refuse actions or to require human approval or to constrain certain accesses to you know information or resources or or even you know route the request somewhere else. So I I think we can build the execution boundary.

erin kenneally (33:04.662)
for an agent that is stronger than how we can enforce, you know, a relationship between an ATC and a pilot, I guess was the messaging there.

Martin Hinton (33:12.808)
No, I mean so w in in when I was reading that I started to think about filtration. And you might have a filter that lets through large parts and the next layer of the filter lets through parts that are slightly smaller, but and then there are layers of of filtration that get to the end where, you know, eventually every particulate matter is taken out of your drinking water and you you enjoy your freshly filtered water. But th this idea that there are layers depending on the capacity of the agent.

to do harm, right? So if the agent is, you know, s sending invoices and you know transferring money, that's a high high threshold there for for potential liability and financial damage, right? the the the example I I I've encountered, and it's very minor, but I I don't know if I mentioned this to you when we were talking in the last year or so. I had a flight to Ireland that was canceled while I was at the airport. It was actually my son's flight, but I'll simplify the story for the purposes of telling it. And at the airport they canceled the flight.

And I got an email saying, hey, we've rebooked you on the same flight tomorrow night. Didn't think anything of it, took a cab home, kept the receipt, figured I'd submit that to them. The next day, around three o'clock, about the time I'd start to get ready to leave for the airport, I checked the app and I was booked on a five o'clock flight. And this was two hours before that departure. It's an international flight. I'm 45 minutes from JFK in the best case scenario. This is a problem.

So my wife jumps on the phone while I jump in the car. And 45 minutes later, I'm on the nine o'clock flight, no problem. But during the course of the call, the human airline employee said, you know what? AI sends those emails. That's not us. Right. This idea that it was, you know, not our and and you know, rather than just put me on the nine o'clock flight and solve the problem, they resisted. Like they were, they were, you know, not particularly helpful right away.

erin kenneally (34:57.718)
Mm-hmm.

Martin Hinton (35:06.344)
So very small example of Martin's encounter with AI. And and I I guess what what I'm wondering is is is this, you know, for the people who are watching this who may be wondering how how did I wind up on a an AI cyber insurance podcast, is this the sort of place that it exists in your life where it makes those sorts of mistakes? Because we see it do that sort of thing. Those of us who use the agents and have played around with them, they they're like people, right? They make small mistakes. And I guess that makes sense given the the huge amount of what they know is based on us. And

I I wonder is is that a way to contextualize the way these things can sort of have a a very small impact on brand reputation, which obviously matters for companies, that sort of thing. That there is this isn't just a an AI that'll inadvertently you know, mail a check to someone or something that, look, that check for fifty grand, that that was a mistake, now I gotta deal with that. I mean w what about that s sort of microscopic Martin example of an AI agent, at least I'm told, being the cause of a pain in the neck.

erin kenneally (36:04.814)
Sure. I mean I think that happens a lot for our purposes for our purposes of our discussion, what level what rises to the level of a of an insurable harm? and I it it gets back to something that you had said before, which is

What's what's consequential, right? So you're not gonna put a a a ton of guardrails around you know a chat bot that's you know answering questions that are you know whatever it may be. What what's the weather today? Versus an AI system that is making consequential decisions for for a company. And I mean, look, again, we saw

Let's take the hugging face example, or even the Pie Pie example. I don't know that they r rose to the level of insurable incidents, but what you did is you had AI escaping sandboxes and attacking other companies. And we saw this with, you know, I think the Australian government was attacked and and whatnot. When you're dealing with consequential impacts with agentic AI.

it's important I to differentiate and to implement a deterministic enforcement next to the models which are operating from a you know kind of a a probabilistic scoring perspective in terms of you know what they expect the outcome can be. you know it's funny because I think of that the Air Canada case when you describe what happened with you, and that involved

you know, dollar loss. So that ended up going to a small claims court and Air Canada was was held responsible in that case. But I I think to your point the consequence of the probabilistic act from or the decision from the model you know that is gonna determine

erin kenneally (38:20.216)
the extent to which and the robustness of the guardrails and deterministic runtime enforcement that you put those agents within.

Martin Hinton (38:31.666)
Yeah, yeah. You you you s you I think you you tested eleven governance tools. Wha w what did you find?

erin kenneally (38:40.78)
Yeah, I mean for the most part from those from that tool perspective, what I tried to do is lay out again kind of, you know, what is the what did the what's kind of the functional equivalent of what we're asking from the cyber pre-breach control stack and and then I, you know, tested, I'll I'll I'll use that term lightly. I I operated, it wasn't a technical test where I, you know, ran the tools next to each other in a in a sandbox.

I operated off of publicly available information and and and then collected that and and whatnot. Look, what I found at that time, and I think the market has changed since, because what we're seeing is a lot of convergence where tools that were initially stood up to address maybe the identity piece or the evidentiary piece that's required are now adding more functionality and capability.

and so there was a little bit disparate there. I didn't find any of those tools that covered kind of the full range of what you would want to address from an an underwriting perspective. And any we even see some of the a lot, I should say, of the conventional, you know, cyber tools adding AI risk control capabilities as well.

You know, the the point I think now is in like well, like I said, I think some of that has changed. There's a lot of talk on the street and you know, CISOs and companies don't want to have to buy fifteen different tools. So we're necessarily seeing convergence of those capabilities within existing tools. Seeing a lot of those those native AI or cybersecurity tool toolings and companies acquire a lot of those.

you know, some of those AI risk control companies. But so the market is the market is changing. and it's gonna be it's gonna be interesting to see, you know, from my perspective, what I the kind of the boulder that I've been pushing up the hill ever since I I was working on cyber insurance go to market with those cyber insurance companies, is the extent to which the insurers are going to kind of walk side by side as it were with the AI risk control vendors.

erin kenneally (41:02.08)
in terms of, you know, how they're assessing the efficacy of those controls, how those controls impact loss ratios, what is the right, you know, combination what what are the right questions to be asking even from an underwriting perspective with regards to AI risk? so it's a whole new world there. And you know, I'm messaging

One of my messages is, you know, you gotta get ahead of it early. if you sit back and wait for those claims to come in, you're already behind the ball, the eight ball.

Martin Hinton (41:35.614)
So s speaking of the the the brave new world, tell us about the new company. Is it Spe rem remind me of the name? I I wrote it down but I can't read my own handwriting.

erin kenneally (41:45.038)
Sure. Yeah, no, no, no, no problem. DLTASKOP. So it's DLTASKOP without the two E's. yeah, look, we're we're positioned, DLTASKOP is positioned as this AI risk insurance intelligence layer, right? So we're not an MGA, we're not a standalone insurance company, we're not an you know, we don't underwrite itself, but I it it's really born out of the unmet need that I've seen in the market.

which is the ability to identify AI risk, coordinate the signals of the risk and the controls related to the risk, and then translate those into evidence for underwriting and pricing. so yeah, I mean it it really is a a practical extension of the problem that I've been, you know, researching and writing and experiencing from an advisory perspective.

that look, this this technical AI risk evidence does exist. it's really scattered, it's uneven, it's rarely organized around decisions that insurance needs to make. and that the genesis then of the name is that, you know, delta from the Greek term for for gap, that AI risk lies in that delta. It stems from these gaps between

what the AI technology capabilities are, and then what are our expectations of how those capabilities impact our individual and our organizational rights and interests. And so, you know, the byline that I like to use, I put on the website, is, you know, from a a problems to be solved perspective, is frankly, AI insurance is being bought and sold blind to that delta. And so we aim to to fill those gaps. and what yeah.

Martin Hinton (43:42.504)
No, no, I I I I  listen, I  think ⁓ again, my  I'm I'm a I'm a simple person. You you you you've identified a problem and there's there's money to sell the solution. And I and I it it's what everyone talks about while trying to not fall behind in being able to provide what people think they need to ensure and protect their business. So it's i it it sounds really interesting. You touched on the website, we've touched on a few papers, just so everyone knows, there'll be links in the show notes wherever you're listening or watching. So you'll be able to find Erin, the papers we've been referencing and that sort of thing.

Erin, we've we've been talking a little bit now. I wonder if we could just start to wrap up. W if I'm a risk manager deploying agents within a company, what's something I should do before the end of this quarter? W what's something I need to start asking?

erin kenneally (44:21.388)
Yeah, I think I think there's some some real concrete actions that can be taken that will help move the ball forward and help really kind of close this growing protection gap that we have and penetration gap as well. So for one, I would advise to look audit, get with your broker, audit your cyber, your CGL, your DNO, your ENO.

policies for for these AI exclusions. right, those those endorsements that those ISO endorsements that took effect and see if that's actually landed in your in your policy. Secondly, again, from more of a a technical CISO perspective, make sure that you inventory all of your your agents, your shadow AI, right? The the AI that is not managed. And we've seen in

increasing the number of numbers of reports that a lot a good portion or material share of those enter enterprise AI use of of AI is unmanaged, right? Because people can spin up models and and agents and systems and frameworks trivially. so that's a one area where we can reduce some exposure and then

presumably help with with pricing and and coverage conditions. I would put briefly touched on this earlier, I I would put binding deterministic enforcement on high impact actions, right? So decisions, you know, in a medical system by way of example, where an AI is you know assisting from a decision support perspective, a doctor or a clinician in making

determinations that affect people's health. we need deterministic enforcement on top of these probabilistic models that are are doing the scoring and you know trending and doing great stuff but again when push comes to shove there's got to be a kind of a rule based deterministic enforcement where consequences are high. The fourth thing I would say is

erin kenneally (46:38.936)
Claims grade logging, right? So logs are not logs are not logs, right? So make sure that your logs contain traces and spans of your agents' activities, d you know, what decisions they took, make sure that the evidence is tamper-proof. that's gonna be key because I I actually think that there's gonna be a premium placed on the the what I'll call the kind of the pre forensic readiness of companies.

in the underwriting assessment. Because look, carriers aren't going to want to spend a ton of money. And it's not always obvious, right, if AI is at play, if it's the cause of loss, and then and then whether it's covered or not. They're not going to want to spend a lot of money on defer only to learn right that you know it's not an AI cause of loss. So this is where having those pre forensic insurance grade evidence

Capabilities in place are going to be key. And then let's see, the last one I would say is, and I push this boulder up the hill with cyber. I I I don't think it's going to be a nice to have. I think it's going to be a necessity if we're going to create a sustainable AI market, which is to say that companies need to leverage the existence of having these provable controls and sharing of

telemetry, whether it's metrics or you know traces or you know solid logging as leverage for better premium and coverage terms. Like th this is the whole incentive problem. carriers will tell you that they want and it was an aspiration with cyber. Hey, we need inside the firewall data. It was somewhat implemented. It's they're still largely relying on external scan data.

And there's no equivalent in the AI space that's gonna give you the diagnostic credibility and reliability to write some of these, you know, much more uncertain AI claims. So I think they're necessarily the carriers are gonna have to, you know, I say require. the only way they can require it is if they give policyholders proper incentives, you know, lower premiums.

erin kenneally (49:04.684)
better coverage terms, you know, higher limits, lower deductibles, whatever it may be, in exchange for proof that they've got AI specific controls.

Martin Hinton (49:17.69)
AI specific controls. The printed on a t-shirt. Erin, I don't have anything else. Is there anything we didn't get to or anything you'd like to say some more about?

erin kenneally (49:26.51)
gosh, no. I just I guess my closing thought would be it's it's very easy to find yourself sort of chasing your tail because gosh things are moving so quickly. but you know, I threw out a lot of kind of fundamentals. and I think if you stay grounded in in fundamentals and apply those to the issue at hand, you can prevent from from kind of being in in quicksand and

engaging in analysis paralysis and and that's what we're trying to do at DLTASKOP as well is really we're steeped in fundamentals of observability and coordination and then and then translation of those into how how our insurers think.

Martin Hinton (50:10.484)
put well put. Well, Erin Kenneally CEO and co-founder of DLTASKOP.ai Thank you so much for joining us again. a year later, with it seems like a hundred years worth of AI news to fill that 12 months. as I mentioned already, but I'll say it again in the show notes. There'll be links to Erin so you can find her and a variety of other things that we've referenced and touched upon today. If you've got a question or a comment, please leave it and

⁓ share, subscribe. All that helps us reach more people with this important information. again, I'm Martin Hinton executive editor of Cyber Insurance News. Thank you very much for tuning in, listening, and watching. Enjoy the rest of your day.