In This Episode
Traditional IAM Wasn’t Built for AI Agents
Michael Krigsman: Jo. Identity and access management systems were designed for people. We had and have defined roles in an organization, and the permissions match those roles.
As you said, agents spawned subagents. They have, metaphorically speaking, a mind of their own. They have that intelligence, as the Hugging Face incident revealed. They can go down unexpected paths, and that creates a problem because we don’t know in this probabilistic world what the agent is going to do.
The fastest path to get a handle on this is number one: get an inventory. You need visibility into what agents are doing inside the organization. You’ll come across agents that are unknown. We don’t know the priority, the importance of the work that they’re doing. We may not know who owns those agents.
So you’ve got to watch those agents, get telemetry, establish the identity, figure out what’s going on, and then narrow down the access, prioritize based on business case, based on risk, based on ownership, and at least you’ll start getting a handle on what’s taking place inside your organization.
Jo Peterson: So true, and you know it is moving so fast, right? I read sort of the post mortem on that Hugging Face situation. It was 17,000 attempts over a weekend that were unsupervised, as you’re probably aware, and what I found interesting about it was the depth of technical expertise or cleverness—cleverness maybe that’s the right word—that the agent exhibited by bringing together, daisy chaining a couple zero days.
And I think, wow, I’m a technical person, and I get what they’re doing. And so, my next part of my question is, do you think the non-technical C-suite has a grasp of the organizational security concerns that come along with this agentic workforce?
Do Business Leaders Understand the Risk?
Michael Krigsman: How is it possible that a non-technical business leader can understand the implications and the ramifications of a particular agent? As you said, the agents identified zero days and strung them together.
You know, we think of agents as being drivers of efficiency, speed, cost savings. We forget that agents have another crucial attribute, and that is they are repositories and storehouses of knowledge, of all the knowledge that’s out there.
And so we have to make the assumption that an agent has intelligence because that’s what it’s designed to do. And for anyone, whether it’s a technologist or a non-technical business leader, to make the assumption that they can know and predict and quantify the exact behavior that a swarm of agents and spawning agents will undertake is a fool’s errand.
So the short answer to your question is: Do non-technical business leaders understand the nature and the depth of the risk? Not a chance.
Jo Peterson: Yeah, and that kind of leads me to my next question. And by all means, I’m applauding those organizations that have put together dedicated AI governance committees. In fact, 50% from what I’d read of large enterprises now have these dedicated committees, but is that structure actually working, or is it sort of creating the appearance of governance, maybe without the substance?
An AI Governance Committee Isn’t Enough
Michael Krigsman: It’s an interesting question. Policies and reports are the perfect fodder for books that sit on shelves, and they’re also great for press releases.
What are we actually doing with the knowledge that we gain from these reports? To what extent are we changing our infrastructure? Are we adapting our infrastructure to meet the needs of agent governance? To what extent are we keeping track of our agents, maintaining that inventory and that accountability for ownership?
So the real test is not the committee, and it’s not the report. The real test is the execution, and we all know that there’s a large gap between the intention, which is the committee, and the execution, which is actually making changes on the ground to enforce the policies.
And that’s where the rubber meets the road. And this is not a technology problem, and it’s not specific to AI. This is an organizational problem.
Jo Peterson: Yeah, and you know, fill in the blank. Any technology that comes along, right? I was around when cloud got started, and people’s cloud bills were going crazy. And it was like, well, no, this person’s in charge. Nope, that person’s in charge. Nope, the person that’s in charge really doesn’t care that the cloud bill went crazy, right?
So it’s this ownership piece that you alluded to that’s kind of really key. So I walk in the door and I’m a new CISO, and I come in and I look around and I’m like, oh, they’re using AI. Hmm, I’m not seeing some of the things I need to see in place, and I don’t want to get behind.
So I’ve got 30 days to put in some fixes, and when I say I’m not seeing some of the things I need to see, you alluded to it earlier. Inventory, no agent-specific controls, no monitoring. Where do I start?
A New CISO’s First Move: Get Visibility
Michael Krigsman: If you’re a new CSO and you’re walking in the door and you’re realizing that this is out of control, don’t lose heart because it means you’re among everybody else.
Jo Peterson: Yeah.
Michael Krigsman: So visibility is the first step. As someone from Palo Alto Networks who I interviewed said, without visibility, you cannot control.
So create a table. Who’s the owner? What is the extent of the reach? What’s the control that’s needed for these agents? You’ve got to identify what’s actually going on.
In every large organization right now, there are going to be agents running amok. We call this shadow AI. Somewhere in the organization, there is a bright person in marketing or accounting who realized that they can use Claude Code to accomplish their job better and faster than they ever could themselves.
You need to be friends with these people. You don’t want to be the owner of no, right? You don’t want it to be like shadow IT in the past, where the CIO defaults to no.
You want to say yes, but you need that control. That means visibility. That means discovery. So that’s where I’d begin there.
Jo Peterson: Yeah, I think that’s good. Somebody said something funny to me the other day, and I just chuckled. You know, we’re all going to become so proficient in AI that we’re going to forget to make PowerPoints. So I don’t know. Maybe that’s so. But any final thoughts that you want to share with the audience?
Start Treating AI Agents Like a Workforce
Michael Krigsman: Agents are powerful, but agents are like—you have an employee, and that employee is fully controlled from a governance standpoint, and everything is—the identity and management all is set up right.
But then the employee is a real go-getter, and the employee says, “You know, I’m going to bring in my cousin, my cousin Joe or my cousin Mary, because Mary’s really good at this.” And Mary says, “Well, my cousin Bob is good at that,” and before long, you’ve got the whole neighborhood working for you.
And that’s sort of what happens with agents. It’s great to have the workforce, but you do have to recognize that agents are not the same as employees.
And it can be helpful to name your agents. So agent number one is Bob, and agent number two is Mary, and start thinking about your agents as contractors, freelancers, or employees.
And at least if you do that and take your inventory, you’re on the right path.
Jo Peterson: That’s good advice. It’s a brave new world, that’s for sure. So it’s been lovely chatting with you. Thank you for taking the time to visit today. And audience, I hope you enjoyed the insight. And if you’re not following Michael and listening to his podcast—okay. Thank you, guys. Thank you, Michael.
Michael Krigsman: Thank you.