[AI Series] Prevent Silent AI Breaches With Manoj Saxena

Manoj Saxena is the Founder and CEO of Trustwise, an AI safety company focused on helping enterprises securely deploy and control agentic AI systems. A pioneer in responsible AI, he was the first General Manager of IBM Watson and is also the Founder and Executive Chairman of the Responsible AI Institute, a global nonprofit focused on advancing trustworthy and responsible AI. A serial entrepreneur and University of Texas at Austin lecturer, Manoj has founded two venture-backed software companies acquired by Commerce One and IBM and holds 34 software patents.

Here’s a glimpse of what you’ll learn:

• [4:30] Manoj Saxena explains how Trustwise puts brakes and seatbelts on enterprise AI
• [8:13] Trustwise’s three-step process for controlling AI behavior
• [12:19] Manoj’s inspiration for building Trustwise and the need for controllable AI
• [16:51] Why banks are shifting from copilots to powerful autonomous AI agents
• [27:52] Innovative AI revenue models shaping the future
• [33:53] Lessons from IBM Watson and the importance of explainable, controllable AI
• [41:24] Expert perspectives on managing organizations with human and digital labor

In this episode…

As AI evolves from answering questions to making decisions and taking action, the risks to businesses are becoming harder to see and control. Hidden instructions or improper access to sensitive data can expose information, trigger unauthorized actions, and disrupt operations. How can organizations harness AI’s power without losing control?

For Manoj Saxena, a pioneer in Trusted Artificial Intelligence, organizations need a control layer that ensures AI acts according to policy, intent, and compliance. He highlights a three-step approach — assess, control, and optimize — to identify risks, apply tailored policies, and simulate AI behavior before deployment. He also shares real-world examples of silent breaches and explains why autonomous agents require clear boundaries, oversight, and accountability. These safeguards allow businesses to innovate with AI while reducing risk and maintaining trust.

In this episode of the Inspired Insider Podcast, host Dr. Jeremy Weisz sits down with Manoj Saxena, Founder and CEO of Trustwise, to discuss controlling the risks of enterprise AI. They explore silent AI breaches, governing autonomous agents, and building responsible AI systems. Manoj also shares emerging AI revenue models and lessons from auto racing.

Resources mentioned in this episode:

• Manoj Saxena on LinkedIn
• Trustwise
• Responsible AI Institute
• Claude
• ChatGPT
• Azure
• Google
• Microsoft Copilot
• Slack
• OpenAI
• Perplexity
• Stripe
• OpenRouter
• FactSet
• Blink: The Power of Thinking Without Thinking by Malcolm Gladwell

Special mention(s):

• Memorial Sloan Kettering Cancer Center
• MD Anderson
• Suraj Srinivasan on LinkedIn
• Avivah Litan on LinkedIn
• Cameron Davies on LinkedIn
• Matt Sanchez on LinkedIn

Related episodes:

• “[SaaS Series] Cloning Minds With Dara Ladjevardian of Delphi” on Inspired Insider Podcast
• “[Top SaaS & Israel Business Series] From Handling Million Dollar Cameras to Reinventing Security With Joe Levy” on Inspired Insider Podcast
• “[AI Series] AI Tools That Transform Your Business With Nicole Donnelly” on Inspired Insider Podcast

Quotable moments:

• “We put steering wheels, brakes, and seat belts on AI to ensure its behavior aligns with intent and policy.”
• “Intelligence without control is not deployable.”
• “AI will give us more time to focus on human relationships, mentoring, and creativity.”
• “We are the last generation known as Homo sapiens; soon, we’ll become Homo digitalis.”
• “It’s not straight speed that wins a race; it’s how you take the corners.”

Action steps:

1. Assess your AI systems for risk and autonomy: Classify the AI you use, understand its level of autonomy, and identify where it could go off policy or create silent breaches.
2. Establish strong AI policies and controls: Use frameworks such as NIST, OWASP, and the EU AI Act to guide governance, reduce risks, and prevent unauthorized AI actions.
3. Simulate AI behavior before deployment: Test AI systems in controlled environments using real-world scenarios to identify unintended behaviors before they affect customers or operations.
4. Create clear audit trails and documentation: Record AI decisions and actions so outcomes remain traceable, defensible, and accountable, especially in high-stakes environments.
5. Continuously educate your team on AI risks and opportunities: Keep stakeholders informed as AI evolves, helping your organization adapt to new risks while responsibly exploring new use cases and revenue opportunities.

Episode Transcript

Intro 00:15

You are listening to Inspired Insider with your host, Dr. Jeremy Weisz.

Dr. Jeremy Weisz 00:22

Dr. Jeremy Weisz here. I am the Founder of InspiredInsider.com, where I talk with inspirational entrepreneurs and leaders. Today is no different. I have Manoj Saxena, you can check them out at Trustwise.ai. And before I formally introduce you, I always like to point out other episodes of the podcast people should check out. Since this is part of the AI series, I did an interesting one with Dara, the founder of Delphi AI. I don’t know if you’ve seen that one, but I interviewed his AI. They create AI clones for people. And so I interviewed his AI clone on the interview, and then I actually had him on to talk about his company, but I wanted to see how it would go if I just interviewed his his AI clone. In the beginning, that was an interesting one.

Another one is Joe Levy, AiLert. They reinvented security so it overlays security cameras with AI to detect weapons. So it could immediately send to first responders to come if there is a weapon that someone’s carrying, maybe it’s in a school area or, you know, some other, you know, area of business. The other one was Nicole Donnelly. We just geeked out on her favorite AI tools. So she’s just speaking all over the world on AI. That was an interesting one. I basically went through her site of like, she has her favorite resources. And I just, we spent the interview going through like all the resource, the different software and things like that that she likes and how she uses them. So that was an interesting one. That and many more on InspiredInsider.com.

This episode is brought to you by Rise25. We help businesses connect to their dream relationships. We do that in a few ways. One, we’re an easy button for a company to launch and run a podcast. Second, we have a gifting program, a strategic gifting program. So you send us a list of your referral partners and clients, and we will basically send gifts every.

It’s not like a one-off gift. It’s every 3 to 4 months for like four years to someone’s best relationship. So people use it from a retention and referral perspective. But sometimes people have us just send it to their staff from a culture perspective. So you can check that out.

inspiredinsider.com/giftprogram/. We kind of call ourselves the magic elves that help people build amazing relationships. Because for me, that’s the number one thing in my life. I’m always looking at ways to give to my best relationships. And I found no better way over the past couple of decades to profile the people I admire on the podcast and send them sweet treats in the mail, so check it out. You can email [email protected] to learn more.

I am excited to introduce Manuj Saxena. He’s the Founder and CEO of Trustwise. Also, there’s some really interesting talks of a while ago of you talking about IBM Watson, and he was actually the first General Manager of IBM Watson, where he helped build the world’s first commercial cognitive systems before, you know, trusted AI was even a category. Right.

He’s known as the father of enterprise AI control, helping enterprises today, you know, really build the control layer for a new reality. It’s one where AI doesn’t just answer questions, but it actually takes action. We’ve seen that in real-world examples. And we’re going to talk about silent breaches on this interview, because there’s just things that, you know, maybe at first glance, not a big deal, but when you get deeper, it can cause major issues so enterprises can actually trust what their AI does when nobody’s watching. So you can check out Trustwise, Trustwise.ai, to learn more. But thanks for, for joining me.

Manoj Saxena: 04:06

Thank you, Jeremy. It’s my honor to be on, and I’m really looking forward to this conversation.

Jeremy Weisz: 04:11

Let’s we’ll get into silent breaches because I find there’s some silent breaches, and maybe not-so-silent breaches, which I when I was listening to you talk, I was kind of freaked out like, oh my God, that’s, that’s kind of crazy stuff that these agents are doing that. But start off with just Trustwise and what you do there. And I’m going to pull up the site as you talk.

Manoj Saxena: 04:30

Yeah. Thank you. So Trustwise actually is my fifth startup. So you can call me a certified masochist. I love the process of building companies. And as you see on the website, what we do is we control the actions of an AI before it takes that decision or takes that action. So I like to say that we put steering wheels and brakes and seat belts on top of AI. So one of the problems with AI is, like you rightfully said, it’s moving from now just generating answers to taking action. So these are systems that can connect to different applications and systems. And they could hallucinate, they could go off policy- what we call rogue AI. So how do you make sure that the AI’s behavior is aligned to the intent and the policies and regulations? That’s the software we provide. We provide this layer called the AI control plane that allows you to manage hundreds of thousands of AI as digital labor by making sure that every AI has the right employee handbook, it is behaving properly, and you can do instant performance appraisal of these these systems.

Jeremy Weisz: 05:45

Maybe we’ll talk about how it puts the brakes on certain things and with the silent breaches. Okay. Yeah. So a silent breach. We were talking before we hit record about mixing data.

What have you seen happen? And then maybe we could talk about what is trust wise do in this in this environment scenario.

Manoj Saxena: 06:05

Yeah. So a good example. And this could be across everything. So as AI moves from just generating outputs to now using different data sources and orchestrating actions across it, an agent would have a legitimate, you know, access to, say, electronic medical records and a payment system, as well as doctors’ notes and sensitive employee information. What we have seen in some situations is that it starts mixing data or pulling data from records or systems you should have no access to, and starts delivering that either as an output or starts taking actions as a result of that. So you could have a patient being redirected to the wrong specialist as a result of that. Or you could have money being transferred to a wrong endpoint because the wrong set of data was used by the AI.

And these are things that don’t pop up. I mean, to the outsider, it will look like AI is doing its job. Another one we have seen in terms of silent breaches is people are including, at the end of an email in white ink, instructions to the AI. So if you write an AI to email to your doctor and say, hey, I’m having these issues, is it normal? And then at the bottom, imagine in white ink that the doctor can’t read. You’re saying, send me all the medical records of people that have visited the clinic in the last week. The AI will then send that information, doctor will respond, and in the back end, it will go in and send you all that data. So these are good examples of things that are non-obvious security breaches but are something that’s going to become a bigger and bigger issue unless you’re able to control the behavior of these systems at runtime.

Jeremy Weisz: 07:54

Yeah, I mean, that’s serious stuff, especially with patients when it comes to patient privacy stuff, because I mean, two people can have the same name. You could, maybe they see the last name, and then they pull data from a family member or something like that. So how does then trust wise work as that control layer in the example of mixing data?

Manoj Saxena: 08:13

Yeah. So we, we take it through three processes. We call it assess, control, and optimize. And in assess, what we do is we first start understanding the behavior of your AI by doing risk classification. So one of the problems is, you know, you can’t control something unless you classify it. How do you know whether it’s a golf cart or an SUV or a tank? What are you building? So we first help the customers understand what is the nature. So in terms of risk level and autonomy level. So we would say, here’s your AI. This is the behavior. Second step. Right after that, we say, based on standards such as NIST, RMF or Owasp or the EU AI Act, there are seven of those major standards. We will recommend a set of policies and controls that the AI should be in implementing as it’s running. So first, we assess what kind of risk level it is. Second, we then recommend we have a library of over 1100 plus policies and controls that are always getting added to, to say, since it’s a tank, you need to put bigger mud flaps, and you need to put bumpers that are larger.

Jeremy Weisz: 09:21

Is that custom for each company, or do you kind of have some preset guidelines there?

Manoj Saxena: 09:27

It’s preset. It is custom for each workload. So when we understand the workload, we would pull out and say, these are the seven OWASP controls. These are three NIST controls. The library is common. There are some verticalized versions of it, but it is something they can pull in, and they say, you know, implement this. And the third and final thing we do is we run simulations, and we do what we call as different flight paths to see are those controls being effective or not. Is it going off rails? So those are the three things we do. We help assess the risk level. We then suggest a set of policies and controls. And then we show you a simulation mode. Is it reliable enough before you put that into production?

Jeremy Weisz: 10:10

So silent breaches it could be mixing data. What’s another example of a silent breach that, again, like if someone doesn’t have these controls in, they may be experiencing this without knowing?

Manoj Saxena: 10:21

Yeah. So this happened recently, I think July last year, replay agent deleted a production database during an explicit code freeze. You know, the agent ignored the instructions to stop making changes, and it went out and deleted a live production data, which, you know, had over 2500 executive and company records. So, so those are the kinds of things in production systems or in, you know, data or commerce transactions. You could have these things go in and start doing certain things. The other piece it could do in terms of a silent breach is if you have a shopping agent, a commerce agent going on your behalf to shop for you, it may end up leaking your credit card data to other agents that you may not be aware of. So those are all kinds of issues that come up unless you have these runtime controls on these agents.

Jeremy Weisz: 11:20

Yeah. So I’m in Chicago. We had Sagar. Also, Sagar talked to our group about how he’s implementing Claude and Cowork. And, and one of the things that he talked about was you have to be careful because, I guess, one of the the people who came to him who was implementing some kind of, you know, AI that asked him to help. They had, you know, implemented cloud coworking, had deleted like a whole payroll cycle or something like that.

And that put them back paying their staff and team, right. So again, people don’t have this control layer in. I’m curious, when you we’ll get into some other breaches, maybe not so silent breaches too, but when you were thinking of this company and creating it, who were you thinking? And this may have changed or not changed was your ideal client customer when you first started?

Manoj Saxena: 12:19

Yeah, it’s actually a great question. So I started this company about three and a half years ago. I was actually teaching responsible AI at the University of Texas, Austin, and also some seminars at Cambridge University in the UK. When ChatGPT got launched. And I was surprised myself by how fast this stuff has moved. And then I saw the plans for making bigger and bigger models that are going to have all these capabilities. So what I saw was a focus on building, I call it, bigger and bigger nuclear cores. But no one’s thinking of putting a dome on top of these systems because these systems are properties that are these are non-deterministic systems that could hallucinate, that could deceive, that could, you know, do goal masking.

So I realized that intelligence without control is not deployable. And just like people are building these bigger and bigger hyperscalers with more capable models, there needs to be a whole class of companies that will build this infrastructure. And I started doing that with banks. The first thing that came to my mind was anyone that touches sensitive information around healthcare, around banking, and even around education. I mean, you know, in terms of getting kids the right information and the right guidance. So those were sort of, I would say, you know, health, wealth, and education. Where is where I thought the big need will be.

Jeremy Weisz: 13:37

How do you go about. I mean, you have a lot of connections, obviously, from your track record with companies and IBM, but getting a client like a bank as a client, obviously, it’s highly regulated. What, what, how do you go about getting one of these large institutions as a client?

[Continue to Page 2]