AI Leadership Insights: Successfully Governing AI with Ken Huang

Hello, and welcome to the latest edition of the Techstrong AI video series Army host Mike Bizzo. Today we're with Ken Wong, who is Chief AI Officer for distributed apps ai. And we're talking about exactly how are we gonna manage all this governance and compliance that exists around AI applications.

'cause I think we're finally starting to understand the scope of the challenge. Ken, welcome to the show. Oh, thank you, Mike.

Wait, when you talk to customers, what's the relative level of maturity about this? I think, you know, it feels like to me we're so excited about building the app in the first place that we didn't really think about how it was gonna be managed and a after the fact. And now people are asking tough questions about like, what's in this thing and how does it work?

What's the level of transparency? So where are we on the journey? Yeah, I think we are still in the, uh, early stage, especially with generative ai.

There is some kind of established practice with, uh, the legacy ai, if I can call, or predictive ai. Uh, but with generative ai, I think, uh, everything changed because of an undetermined this tech nature of the ai. So who ultimately is gonna be responsible for all this?

Is it the same folks that have been doing this all along? Or will there be, um, new teams with different tools who are specifically taking this challenge on? Yeah, I think this may be a good question.

Uh, that's why at the Cloud Security Alliance we have the, uh, AI corporate responsibility working groups. Uh, we trying to, uh, building a matrix to map who is responsible for what. Like, uh, do you, if you have the chief data officer or you have CTO, uh, now, like you, you have the new laws, uh, from the Biden administration.

It's, uh, kind of mandate from the executive order, right? The AI exact order that each federal agency has A-C-A-I-O or Chief AI Officer, uh, laws. Uh, so the, the matrix we're building, uh, we actually apply to different white papers that we are working on.

The first one is the core responsibility, uh, which is really if you build your own model or if you fine tuning the model, uh, that's the model and also where you get the data, right? So the model, the data, and also the vulnerability in terms of ai. So that's the first white paper we published in April.

And in, uh, after six months in October, uh, just, uh, a month ago, we published a stick in white paper, which is more close to, related to what we are talking. It's the GRC, uh, the governance risk management compliance, and also cio. Uh, so that's the most recent one.

Currently. Uh, we have the third white paper in public review. Welcome everyone to jump in, uh, to review it.

It's more on the downstream of the model applications, uh, that's including the applications, tools, frameworks, uh, even, uh, the supply chain. So, uh, this is the odd for public review. Are we ultimately gonna have AI agents that manage those tasks and affect the AI manages the ai?

Or how do you see that all plan out? Yeah. Uh, I, I hope this is the kind of the, uh, future, the foundation, right?

But nevertheless, the human always have to be in the loop. Uh, currently everyone is talking about, uh, using the AI agent, uh, for the, uh, SOC or a circulator operation center, right? And also for the defensive, or even the offensive circulator, rather teaming some of, for the company, uh, kind of like even open AI or Microsoft as well, uh, in their rather teaming practice, they actually, I think especially Google, it's public in terms of how they doing it.

They have the red teaming agent, blue teaming agent, and have the judge to judge what they doing. Right? So, um, yeah.

So this, this is a ongoing, uh, kind of, uh, area of research and also some practice. It is very lacent. Uh, it's just a beginning.

Do you have any thoughts is as that develops, one of the things I've seen already is there's a lot of AI agents running around and I'm, I'm scratching my head, trying to figure out, well, how will they all get orchestrated to create something that feels like a workflow that might span governance compliance and or is that the next frontier of research? Okay. Yeah.

This is a, certainly, I, I believe with you, Mike, that's the new frontier of research. And one, uh, of the, uh, candidate framework that we propose. I work with us 40 other peoples just under the oasp, uh, umbrella open web application security project.

Uh, they release the top 10 for a LM or will recently release a candidate framework for the agent, uh, inject ai, uh, in terms of, uh, access control orchestration, uh, confused the DePue. So certainly just like, uh, when you're talking about as agent, that's, uh, a kind of, uh, a attack surface en large, right? You can have in the planning kind of phase, uh, you may have some attack surface there.

And also in the reasoning, reflection and, uh, eventually the exclusion phase. So each phase there is a potential vector of attack and potential vulnerabilities. So I think the, this US is trying to do it.

And, uh, uh, in the, uh, cloud circulator lines, uh, we do have another working group called AI working Group that I'm co-chair. We're trying to build as the, the matrix, right? Uh, given different control we're using, uh, the cloud control matrix as a baseline and add more AI related control, including the, including the agent ai, right?

And then trying to see who is responsible for what, who is impacted, and what kind of vulnerability it could be. Uh, this is a huge matrix. We're building it, we're almost done.

Uh, it should be in publicly review, uh, early next year. Do you think the regulators and auditors are starting to figure all this out and they're gonna be asking tougher questions soon? Or are they still like pretty far down the road from figuring any of that out yet?

So the Python administration has the, um, ai, uh, safety Institute, right? It's, uh, had, uh, uh, kind of spearheaded by nest now. Uh, it's very good.

I actually visit the, the, uh, innovation lab. It's, it's big. It's huge.

It's, uh, very, very good. Like, uh, right? So they did do a lot of work, but they, they are not regulators.

They, of course, they help, uh, in terms of, uh, uh, to one of the requirement in the exact ward that is to build the testing, uh, uh, leather teaming practice, the evaluation matrix. And then this has the AI IMA for our risk management framework, uh, focus on how do you measure the risk? How do you map all this risk, and then how you manage and govern, right?

And we actually leverage this in our, uh, uh, uh, c the GRC, uh, responsibility, uh, white paper as well. Uh, were not replicate what already have it's really good document. Uh, but we leverage what it has, reference it, and create our own kind of matrix there.

As we go forward. Is it really just a matter of time before we have some sort of crisis and then maybe we learn from the crisis? 'cause that seems to be the way we've done everything in the past.

So will AI be any different, or can we be more proactive? Yeah, I, I think for previous question, I, I didn't give you a clear answer about regulate, right? So, uh, I, I think I, I do still need mention that, uh, because Trump is reelected, uh, one of his, uh, kind of initial, um, action will appear all the executive order from Biden administration.

So potentially that could be right, a, uh, right, it's, uh, and I think, uh, it's, uh, currently there is some slight majority in Senate and also in the house, uh, in, uh, from the Republican, there may be potential to have a clear regulation. Uh, maybe you know that in California, uh, the 10 47, uh, bill 10 47 was vetoed. It's goes through both false, uh, chamber of the, uh, like in the state legislature, right?

But a vetoed, it's because, uh, the reason that, uh, uh, governor Newsom said, uh, that's is more like a too futuristic looking. We have to regulate now. But for regulat, now, maybe current CCPA or other regulation may be applied.

So, um, so given the divided, uh, government in terms of register body, uh, it's really ha hard to actually, from federal level to coming up with some regulation. But there's a quite a few state actually already have the, uh, AI regulation, right? Uh, so it's more like a potential, it could be state level, it's not a federal level, right?

Uh, so now give you question, uh, about if there is a disaster, right? Then it may promote the, the, the tough regulation, um, not necessary, you know, uh, in the, uh, crypto space, there's a, uh, I, I think you may also heard FTX collapse, right? It's kind of a disaster for most people, but we don't see as actually any official regulation, right?

Uh, and, but there is some of, uh, rule making, uh, from the SEC, right? And some lawsuit, but eventually, you know, uh, SEC chairman we're stepping down, right? So, so that's, yes, that's hard to say.

Larry said, I think I, I, I think the European Union may be the one who's driving most of this conversation at this point, right? Oh, yeah. You, AI act is actually, uh, uh, enforced, right?

Uh, so the most of provision will be in full, in effect, like, uh, in, in two years, uh, in October, August, I think August, 2026. But some of for provision will gradually, uh, become applicable. Yeah.

So they, they have very tough regulation. So, uh, you have prohibited use. You cannot use, like, for biometric identification, uh, using ai unless you are law enforcement, you have to apply for exception, right?

So there's a positive high, medium, low, right? So, and then you, you basically, everything from medium to high, uh, you have to do the red teaming testing. You have to, uh, place the governance structure.

So, so people said that, uh, US is for innovation, right? So, uh, EU is for, for the, uh, regulations, yes, there is some controversial opinion about EU AI act. Uh, but, uh, some company have to deal with it.

There will always be debates about ai. And one of those debates is, um, so should organizations have like just the center of excellence for AI initially and appoint the chief AI officer? Or do we need to kind of infuse this down into our workflows so that, you know, the folks who are running GRC just see AI as yet another thing to be managed.

And, um, you know, we're not creating this isolated thing. 'cause what happens is if there's a special group doing it, everybody just assumes that they're handling everything. Yeah.

Actually, I don't know if you know about that. I recently published a book, uh, it's called The Handbook of Chief AI Officer. In that book, actually, I lay out why a, an organization need actually a specific law for chief of AI officer.

Uh, so one of the main reason is because of a technology innovation, especially in the AI space, it's really hard for a, like a, a chief Data Officer of CTO or CIO, because they have their own other daily jobs, right? For them to really focus on that, right? Uh, so one reason is really is the technology innovation, the speed.

Another is really the business, right? The business side. So how do you really map the business with, uh, this, uh, innovation?

And also with all this regulation, security, safety, you do need a, a big organization need A-C-A-I-O, I said, and medium, maybe small, you can have virtual CIO, right? You can hire someone more external to be a virtuous, uh, CAIO. So yeah, I agree with you.

So the answer is a little bit of both. 'cause, you know, otherwise we need, we basically need everybody to get involved, is just a question of who's at the front of the proverbial sled. Oh, Yeah.

Yeah. I think, uh, uh, one thing I have to emphasize, there must be one in, in the organization to be responsible for ai. So that's the, whatever the title it is, right?

It's, it's, yeah, you have to kind of have the, uh, ai, um, chief AI office. So maybe there's a good title. So what's your best advice to folks in the converse of that is what's that one thing you see folks doing today that just makes you shake your head a little bit and go, we're heading for trouble?

Yeah. So the advice, I think, uh, it's very important, especially the, for some established business, they already have the security program in place, right? That will not work in the, this deterministic AI or larger land model or multi models world.

So you have to basically change the, or augment the security program, right? Uh, to look at, uh, the full stack. Like how do you look at the, how do you select the model, right?

How do you do achieve augmented generation, protect the vector database if you do the age agent, or do you protect the workflow of agent? So this is not in the playbook, in the process, procedure, or policy of whatever the security programs the enterprise has, right? And I actually publish a book on this as we published by spring, uh, this, uh, April o it's called the Generative AI Security.

Uh, so in this, I lay out, uh, the way how company can build their own security program, uh, not from scratch, but argument because of this, uh, new right, uh, uh, uh, generative ai. All right, folks, you heard in here. Hey, there's a couple of books worth reading for sure.

But it sounds like to me that the AI Governance playbook is, it's a work in progress, and we're all gonna learn together. Hey, Ken, thanks for being on the show. Oh, thank you, Mike.

It's a great talk To you. Yeah, take care. And thank you for all watching the latest episode of the Textron AI video series.

You can find this episode and others on our website. We invite you to check 'em all out. Until then, we'll see you next time.