# Derek Fisher on AI Governance, AI Agents & MCP Risk

An episode of [AI Security, AI Cloud Ops & Infrastructure | ClearTech Loop](https://invisiblenetwork.ai/podcast/1823980617), 1 Jul 2026, 14 min.

Audio: https://www.buzzsprout.com/2248577/episodes/19429925-derek-fisher-on-ai-governance-ai-agents-mcp-risk.mp3

AI governance is no longer just a policy conversation. As AI moves into business workflows, sanctioned platforms, employee tools, local models, agents, and third-party services, organizations need to understand where AI is being used, what it can access, who approved it, and who owns the outcome when something goes wrong. In this episode of ClearTech Loop, Jo Peterson speaks with Derek Fisher, founder of Securely Built, cybersecurity educator, author, and Director of Temple University’s Cyber Defense and Information Assurance Program. Derek brings a practical security lens to AI governance, AI agents, and third-party MCP risk. The conversation covers why governance needs clear ownership, how organizations should think about AI agents as non-human actors with access and authority, and why MCP servers and AI-enabled services should be evaluated through a third-party risk management lens. This episode is especially relevant for security leaders, technology leaders, compliance teams, and business executives trying to move AI from experimentation into controlled, accountable use. In This Episode: • Why many organizations still do not know where AI is being used • Why AI governance needs executive ownership, cross-functional standards, and business accountability • How AI agents create new access control and auditability challenges • Why agents should be treated more like privileged non-human actors than simple tools • What organizations should ask before adopting third-party MCP servers or AI-enabled services • Why AI governance is not about slowing the business down, but making the approved path usable enough that people follow it Key Questions: • How do we operationalize AI governance, and who is legally accountable when an AI agent makes an unauthorized decision? • How do we prevent agents from executing actions the user should not be allowed to perform? • How do organizations verify the authenticity and security of third-party MCP servers and services? Featured Guest: Derek Fisher Founder, Securely Built Director, Cyber Defense and Information Assurance Program, Temple University Derek Fisher is a cybersecurity leader, educator, author, and speaker with experience across product security, secure software development, governance, risk management, regulatory compliance, incident response, and cybersecurity education. Host: Jo Peterson CIO, Clarify360 Chief Analyst, ClearTech Research Additional Resources: • Securely Built https://securelybuilt.substack.com/ • The Application Security Program Handbook https://www.manning.com/books/application-security-program-handbook • Derek Fisher on SecureWorld News https://www.secureworld.io/industry-news/author/derek-fisher • Your AI Coding Assistant Has Root Access—and That Should Terrify You https://www.secureworld.io/industry-news/your-ai-coding-assistant-has-root-access Watch More ClearTech Loop: Subscribe to ClearTech Research on YouTube: https://www.youtube.com/@ClearTechResearch Stay in the Loop Follow ClearTech Research for more conversations on cybersecurity, AI governance, cloud, enterprise technology, and emerging risk. 🎧 Listen: In Buzzsprout Player ▶ Watch on YouTube: https://www.youtube.com/@ClearTechResearch/videos 📰 Subscribe to the Newsletter: https://www.linkedin.com/newsletters/7346174860760416256/

## Transcript

[0:07] SPEAKER_00: Hey everyone, thank you so much for joining today's edition of ClearTech Loop.
[0:12] SPEAKER_00: I am Joe Peterson.
[0:14] SPEAKER_00: I'm the CIO of Clarify 360 and the chief analyst at ClearTech Research.
[0:18] SPEAKER_00: And I'm here today with Derek Fisher, founder of Securely Built.
[0:22] SPEAKER_00: Hi, Derek.
[0:23] SPEAKER_01: Hi, how are you doing?
[0:24] SPEAKER_01: Thanks for having me on, Joe.
[0:26] SPEAKER_00: Thanks for taking time to visit.
[0:29] SPEAKER_00: In case you're not following Derek, you should.
[0:32] SPEAKER_00: Derek and his company provide advisory services to mid-market firms for strategic planning, risk management, and regulatory compliance.
[0:41] SPEAKER_00: He designs and implements security programs, including bone management, secure SDLC, and incident response in order to help firms strengthen their resilience.
[0:51] SPEAKER_00: And he delivers cybersecurity education and training awareness to advanced from I should say education and training to advanced technical skills.
[1:01] SPEAKER_00: With over 30 years of cybersecurity experience, Derek serves as director of Temple University's Cyber Defense and Information Assurance Program.
[1:13] SPEAKER_00: So welcome, Derek.
[1:15] SPEAKER_01: Hey, thank you very much.
[1:18] SPEAKER_00: As you guys know, we do three rapid-fire questions that are of the moment that security professionals are thinking about.
[1:26] SPEAKER_00: So let's not deviate from the norm.
[1:29] SPEAKER_00: Let's get on with question one.
[1:31] SPEAKER_00: Derek, how do we operationalize AI governance and who is legally accountable when an AI agent makes an unauthorized decision?
[1:42] SPEAKER_01: So that's, I mean, that's tricky because I think one of the biggest challenges is that a lot of organizations don't know where the AI is, where AI is being used.
[1:52] SPEAKER_01: So there's, you know, whether it's being used within sanctioned platforms, you know, Salesforce, if you're creating uh uh workflows within there to automate and and create uh um features, and that's great.
[2:05] SPEAKER_01: But then there's also you know, shadow AI where people are able to, whether you're going directly to ChatGPT or Claude or you're creating your own local um, you know, uh agents and and you know utilizing local LLMs.
[2:18] SPEAKER_01: I mean, it this this is one of those things where it's very difficult to put the toothpaste back in the in the bottle, right?
[2:25] SPEAKER_01: Or the or the tube because once these tools are are in the hands of of everybody, it's very hard to get a uh uh a grip on where it's being used, how it's being used, and more importantly, how is the company data being used in those systems?
[2:41] SPEAKER_01: Um, you know, one of the podcasts I was listening to recently, I think it was yesterday, they were talking about how, you know, when we had these massive changes in technology, like let's take nuclear technology as an example here, you know, as a collective, as a society, we all decided like, hey, this is very powerful technology.
[2:59] SPEAKER_01: We really need to get our hands around it, make sure that it's not, you know, nobody can build a nuclear reactor in their home, right?
[3:06] SPEAKER_01: And so that's a little different now because we have a technology that could and could have very far-reaching implications, and yet it's pretty much in the hands of everybody.
[3:16] SPEAKER_01: And so being able to get your hands around that and be able to make sure that we understand how decisions are being made and how we can get um you know the appropriate controls around it can be difficult.
[3:29] SPEAKER_01: But what that really starts with is having an AI governance board uh that's in um that is cross-cutting.
[3:37] SPEAKER_01: So you don't want to just have it in the technology side, but you also need it uh, you know, members of that governance board that are part of uh HR, that are part of legal compliance, um, and have that like cross-cutting across the organization visibility into, you know, what is it the organization uh really um, you know, what are the impacts of this technology?
[3:59] SPEAKER_01: Uh above that uh AI governance is really an executive uh leadership.
[4:04] SPEAKER_01: Now it depends on the organization who that executive or or what that executive leadership looks like, but they should be the ones that are in control of the actual risk appetite of the organization as it relates to AI.
[4:17] SPEAKER_01: So the executive at the top uh owns the the that risk appetite, the governance board below that really uh decides the standards and um and the way that uh any like exception use cases and things like that that are that need to approval uh go through that AI governance board.
[4:36] SPEAKER_01: But down at the in the trenches, you really have uh product owners.
[4:40] SPEAKER_01: Um the product owners are the ones that are managing the use case outcomes.
[4:45] SPEAKER_01: So for instance, like if you have um, you know, if you're utilizing AI in order to, you know, create um, I don't know, create better reporting or something like that uh within a certain business unit, you know, the product owner of that particular area needs to have the the, at least in the lower levels, the the ownership of the outcomes from that AI.
[5:11] SPEAKER_01: Um, but again, you have above that the governance board, above that the risk uh being owned at the executive leadership.
[5:17] SPEAKER_01: So that's sort of the way in a nutshell uh that that should be divvied up.
[5:22] SPEAKER_00: Yeah, and that's fair, and that's very thoughtful in terms of the approach.
[5:26] SPEAKER_00: Um so I just heard MCP referred to as the confused deputy, which I thought was funny.
[5:34] SPEAKER_00: Um, how do we prevent agents from executing actions that the user should not be allowed to perform?
[5:42] SPEAKER_01: So, you know, I I think sometimes about AI agents, specifically as as, and I know this is like an oversimplification, but as a privileged account, right?
[5:52] SPEAKER_01: It's an account, it's a it's a non-carbon user um that has the ability to um, you know, make changes in the environment.
[6:01] SPEAKER_01: I mean, we're not that it's not that far removed from like a service account, uh, with the exception that you know these are far more powerful, have far more uh access, can do more.
[6:13] SPEAKER_01: Uh, you know, with with the service account, you sort of have a a you know, single direction or bi-directional type of interaction that you can really kind of shape that what that traffic looks like and that interaction looks like.
[6:24] SPEAKER_01: With um, you know, with with uh MCP in particular, where you have um clients and you have you know servers or clients and and uh services, you know, that the MCP is is you know making that connection to, it's a little bit different because tools can be added on the fly.
[6:42] SPEAKER_01: Uh so let's say, you know, if we're if we're uh if you take uh Claude as an example, that that just happens to be my my tool choice, um, you know, I can ask it to generate a Word document out of uh a text, you know, that that it output.
[6:57] SPEAKER_01: It's utilizing MCP in the background to you know connect to a service to generate a Word document.
[7:04] SPEAKER_01: Um, you know, that's a lot different than what we would typically see in like a service account type of activity.
[7:11] SPEAKER_01: But it really comes down to like shaping that uh access and putting the guardrails around it.
[7:17] SPEAKER_01: So, you know, having the pro appropriate RBAC and control, you know, rule uh rule-based uh access control or role-based access control.
[7:26] SPEAKER_01: Um, having also uh a registry of allowable services, like here are the services that this, you know, um, that uh this client is allowed to be able to access in terms of like, you know, you can't just create, you know, connect to a service uh on the fly.
[7:44] SPEAKER_01: Like it has to be in an allowable set of services that that MCP uh is allowed to connect to.
[7:51] SPEAKER_01: Um, you know, putting those those guardrails around in terms of how it interacts with those services, uh, and also having auditing in place.
[8:00] SPEAKER_01: So you know that, like, okay, here's what it did, you know, here's what it connected to, um, here um was the way it interacted.
[8:09] SPEAKER_01: Um, and so that you're able to at least peel back the onion at some point, should something go wrong, that you're able to at least audit that interaction.
[8:18] SPEAKER_01: Um, and then really just having that transparency around how the MCP, um, you know, how it's making those connections between the services and the client, and um knowing what types of um, you know, again, that sort of ties into that auditing, but like what is that, what is it, how is it thinking about making those connections?
[8:43] SPEAKER_01: How is it thinking about um, you know, uh meeting the goals of of what the client is asking?
[8:50] SPEAKER_01: So it really comes down to again, shaping that connection, uh, having uh a solid uh auditing uh around you know what it's doing and then having that visibility into it, and then really just you know having uh those guardrails uh in place, shaping the the type of traffic that it's allowed to do.
[9:08] SPEAKER_00: That's a great answer.
[9:10] SPEAKER_00: Thank you.
[9:10] SPEAKER_00: Um, one of the things I think about, and I wonder how we solve for is there's more and more MCP servers coming into everyone's environment somehow, some way.
[9:23] SPEAKER_00: How do organizations verify the authenticity and security of these third-party servers coming in?
[9:31] SPEAKER_01: And this comes back down, and I don't mean to like sort of like tie back down uh into like what's sort of uh been security classic or or whatever, but I think you know, it comes down to third party management, like we like we've been doing, third-party risk management.
[9:47] SPEAKER_01: Um, you know, the first thing is, you know, what it what business need is this particular um, you know, what business need is this is this really solving?
[9:57] SPEAKER_01: Like what problem are we solving with this?
[9:59] SPEAKER_01: Yeah.
[9:59] SPEAKER_01: Um, because you know, again, we would ask that question in a third-party risk assessment is like, do we need this?
[10:05] SPEAKER_01: Like, do we not already have this capability?
[10:07] SPEAKER_01: Is this something that we really need to go external for uh or use you know another uh service uh before onboarding them?
[10:14] SPEAKER_01: Um so really, you know, what is the what is the value and what is the problem that this that this is uh that this particular uh case is solving?
[10:24] SPEAKER_01: Um and then it comes down to um you know understanding um do we have an internal team that's actually capable of being able to manage this?
[10:36] SPEAKER_01: Uh, because again, looking at like a third-party risk assessment, like, okay, we can get this tool and bring it in, but do we have people that have number one, the knowledge to be able to uh manage this tool or utilize it appropriately?
[10:49] SPEAKER_01: Um, do we have the ability to uh provide proper oversight for it and and so forth?
[10:54] SPEAKER_01: And then also, you know, compliance and regulatory uh um, more importantly, like the regulatory aspects of utilizing that service, um, you know, is there uh uh concerns in that in that aspect?
[11:07] SPEAKER_01: Do we have uh concerns about how our data is being used?
[11:11] SPEAKER_01: Do we have concerns about um, you know, the way they're, you know, the way they're processing the data, the information we're getting back, are we, you know, doing critical uh or making critical decisions based on on the interaction that we have there?
[11:24] SPEAKER_01: So, you know, those are sort of again, uh not the not to sort of put put it in the same bucket as like three uh third-party risk assessment, but it sort of is that, right?
[11:36] SPEAKER_01: We we should be treating it in the same way.
[11:38] SPEAKER_01: Um AI is not AI is a novel way for us to sort of tackle technical problems, but at the same time, it's a it's a tool, it's a technology tool that uh we we should be leveraging as much of our kind of learned experience and learned knowledge uh to be able to uh apply in this case.
[11:59] SPEAKER_01: Um it yes, it's more powerful, yes, it can do more things, but in essence, we should be utilizing what we already know from a security standpoint uh on on how to better uh you know get our hands around it.
[12:11] SPEAKER_00: Right.
[12:12] SPEAKER_00: And I asked that, and you went exactly where I think you needed to with that one, is I want to draw out the idea that I don't think people are asking the question.
[12:23] SPEAKER_00: Right?
[12:24] SPEAKER_00: It's just for whatever reason there's some sort of disconnect.
[12:28] SPEAKER_00: And sure, large companies have lots of talented folks, but the smaller you get in terms of companies, the less bench you normally have.
[12:40] SPEAKER_00: And not that those people aren't smart, but they're overworked and they're they just may not have enough resources.
[12:47] SPEAKER_00: And I don't think sometimes the question's getting asked.
[12:49] SPEAKER_00: What do you think?
[12:51] SPEAKER_01: Yeah, I agreed.
[12:52] SPEAKER_01: I mean, you know, I I sort of think about, and I was thinking about this the other day, that we've been complaining for so long, you know, in in cybersecurity and technology and in general, that like we have people that are just burned out.
[13:09] SPEAKER_01: You're working many hours, you're never getting ahead, you're never catching up.
[13:14] SPEAKER_01: Um, you know, this is our moment to utilize these tools, whether it's you know, AI, particularly AI, to be able to hopefully um reduce that cognitive load and and hopefully reduce that amount of of stress and burnout in the people that we uh that we employ.
[13:31] SPEAKER_01: And, you know, hopefully AI is is going to be a solution to that.
[13:36] SPEAKER_01: Um and I I hope that that is the the direction we're heading.
[13:40] SPEAKER_01: Um sorry, I hope that sort of answers your question.
[13:42] SPEAKER_00: It did.
[13:43] SPEAKER_00: It was perfect.
[13:44] SPEAKER_00: It was it was a perfect answer, and you and I are thinking about it the same way.
[13:47] SPEAKER_00: So it's that was great.
[13:49] SPEAKER_00: Well, it's been great to have you visit.
[13:51] SPEAKER_00: Thank you so much for taking time to visit with me, and guys, we'll catch you later.
[13:56] SPEAKER_01: Thank you.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.