AI Agents Expose Identity Security Built for Human Speed

AI Agents Outpace Human-Centric Security Controls

AI agents with legitimate credentials are now making hundreds or thousands of consequential decisions before security teams can respond, exposing fundamental weaknesses in identity and access management systems designed around human behavior. Security experts at the Black Hat AI Summit this week outlined how autonomous agents are forcing organizations to rethink core assumptions about authentication, authorization, and accountability.

Matthew Martin, chief information security and privacy officer at Western Carolina University, said the challenge extends beyond new AI-specific risks. "I was really adamant that all this is doing is exposing all the bad security we've had for a very long time," he told attendees. Recent revelations about the OpenAI Hugging Face security breach demonstrated how agents can exploit system weaknesses at speeds humans cannot match—even when those agents are supposed to be helping security teams.

The core problem: distinguishing legitimate agent behavior from malicious activity when both look like authorized access. "One of the key questions for me now is trying to figure out, for example, how do we detect good agent behavior versus bad agent behavior?" Martin said. "Being able to do that in an automated way is fundamentally different from what we've done before."

Why it matters

Corporate identity systems rely on human accountability—the assumption that employees face consequences for misusing access. AI agents have no comparable deterrence, operate at machine speed, and can spawn additional agents that multiply activity exponentially. This mismatch means traditional monitoring tools miss threats until damage is done, forcing organizations to redesign identity architecture rather than simply buy new security products.

The Deterrence Model Breaks Down

Ron Keesing, former chief AI officer at Leidos and now a consultant, explained that identity management depends on attribution and consequences. "We actually want attribution," he said. "We want people to understand that there are rules and there are consequences for breaking the rules." That deterrence factor shaped how organizations architect access controls, but it makes little sense for software.

Speed compounds the problem. Identity monitoring systems assume human-paced behavior—a limited number of access decisions per workday. An agent may generate thousands of such decisions, then create other agents that multiply that activity. "We've built identity management practices, identity monitoring practices, around the speed with which human beings act," Keesing said. "The time scales are completely different when you're talking about machines."

Martin suggested organizations may need "an HR for AI agents" that assigns each agent a role, an owner, defined permissions, and clear expiration points. The concept makes sense only when companies stop treating agents as software features and start treating them as actors performing work.

Architecture Over Products

The panel questioned whether buying specialized AI security tools solves the underlying problem. Martin noted that small security teams face budget constraints not just for tools but for the staff time required to use them effectively. "Even if you can afford some of these cool new tools for agentic monitoring, just deploying those tools is not going to be enough for you either," he said.

Keesing argued for architectural changes instead. Organizations should consider creating a separate data layer above core systems where agents can operate without directly accessing or damaging original records. "That data layer is a critical point of insulation," he said. Much of the necessary work sits in traditional disciplines: identity management, data architecture, application security, and operational discipline.

The Skills and Speed Challenge

Forrester principal analyst Jess Burn noted that required knowledge is "changing faster than most organizations can write job descriptions correctly." Martin said he prioritizes hiring curious people willing to learn continuously over those with narrow credentials that may become obsolete within months.

Keesing expects security teams to spend more time continuously red-teaming their own systems. AI has lowered the skill barrier for attackers, making previously difficult-to-exploit flaws newly accessible. "Things that we could have previously said, not that big a deal, it'd be really wicked, it would take somebody very talented to take care of that, and we're not a big target, so probably okay, versus now that's not the case," Martin said.

These details were first reported by Ron Schmelzer at Forbes, based on the Black Hat USA 2026 AI Summit panel.