As enterprises rush to deploy AI agents across internal workflows, identity security is shifting from background infrastructure to a central business concern, according to Geoffrey Mattson, CEO of SecureAuth.
Speaking on the latest episode of ID Talk, Mattson argued that agentic AI is forcing security teams to rethink how they manage access, authorization and risk, as autonomous systems move beyond chatbot interfaces and begin taking action inside enterprise environments.
“In the past, in security, identity is what my kid would call NPC, non-playing character,” Mattson said. “Whereas now with agentic, it has main character energy.”
The comment reflects a broader shift underway as enterprises experiment with AI agents that can handle workflows, interact with software systems and execute tasks on behalf of employees. Unlike traditional enterprise software, Mattson says these systems combine the unpredictability of human users with the scale and speed of machine access.
“They are as unpredictable as a human, but they are as powerful as a service account,” he said.
That combination creates new security risks that many organizations are only beginning to understand. Service accounts have long been treated differently from human users because they can access large amounts of data quickly and automatically. AI agents, Mattson argues, inherit that power while also introducing reasoning and behavior that cannot be fully predicted or statically controlled.
Why traditional controls fall short
Mattson says many security teams are under pressure from both executives and employees to accelerate agent adoption, even while concerns around prompt injection, excessive permissions and unintended actions remain unresolved.
“Security people are in the position of being doctor no,” he said, describing boards demanding faster deployment while employees independently experiment with agentic tools across the business.
He pointed to one example involving a large financial customer that experienced a data breach after a C-suite executive used AI-assisted coding tools and approved a series of access requests without fully understanding what permissions were being granted.
“It’s a delightful experience to vibe code if you’re an executive,” Mattson said. “They’re getting these consent streams that say, ‘Can we access GitHub?’ and, like, what is GitHub? And just yes, yes, yes, yes, let me get through it.”
For Mattson, the solution is not simply trying to make models safer internally, but enforcing stronger authorization controls around what agents are allowed to do.
“You need to put guardrails not inside them, but around them,” he said.
That means continuous authorization at the transaction level, real-time monitoring of behavior and restrictions that can limit or quarantine an agent when it begins to drift outside expected patterns.
Agentic AI is already here
Mattson rejected the idea that agentic security is a future problem comparable to the long-term threat of quantum computing. Instead, he says the issue is immediate.
“I’ve never heard it’s never going to be a problem,” he said. “Agentic is in your face right now.”
He added that some of the largest financial institutions are moving faster on agentic adoption than startups, driven by both efficiency goals and executive demand for internal deployment.
“Your C-suite’s gonna be using them. Your developers are gonna be using them, and your board’s gonna be asking you to deploy internally,” Mattson said.
As that pressure grows, he believes enterprises will need to stop treating identity as a secondary security function and recognize it as the primary control plane for AI agents operating inside the business. For identity vendors like SecureAuth, that may be the defining enterprise security shift of the next phase of AI adoption.
For more colour on these and further insights, listen to the full episode of ID Talk, available through the media player below or wherever you get your podcasts.
–
By the ID Tech Editorial Team
