Share with your CISO
Agentic AI security dominated day one of Black Hat USA 2026, with practitioners converging on a pointed conclusion: autonomous agents in production are outpacing every governance model built for static software. Jon Oltsik of theCUBE Research framed the core tension, noting that organizations managing the threat well are doing so through deliberate strategy rather than tooling alone. The operational consensus settling around agentic AI security is that identity, not perimeter, is now the primary control surface.
What this means for your business
The companies sitting exposed right now aren’t the ones without AI agents, they’re the ones who deployed agents and assumed existing identity and access management infrastructure would hold. Static entitlements, meaning fixed permission sets assigned once and rarely revisited, break the moment an agent starts improvising paths to complete a task. If your non-human identity inventory hasn’t been audited in the last quarter, the governance gap is already open. The question isn’t whether your agents have too much access; it’s whether you’d know if they used it.
Oltsik’s framing of visibility as a prerequisite to control is the right sequencing, and it’s one most enterprises skip. The instinct is to buy a control layer first, a policy engine, a privileged access management tool, and then discover you can’t govern what you haven’t catalogued. Agentic identities multiply faster than human ones because any developer can spin up a new agent without a formal provisioning request. That velocity makes the inventory problem active and continuous, not a one-time audit exercise.
The human-in-the-loop argument coming out of Black Hat deserves more precision than it usually gets. “Keep humans involved” is not a security architecture. The operationally meaningful version is defining, in advance, which agent decisions require human sign-off, at what confidence threshold, and under what time pressure. Organizations that get this right will find that human oversight shrinks over time as trust is earned and edge cases are resolved. Organizations that treat it as a permanent checkbox will just have slow agents and the same residual risk.
Concept deep-dive: Non-human identities
A non-human identity is a credential or access token assigned to software rather than a person, covering bots, service accounts, and AI agents. They’ve existed for years in IT infrastructure, but agentic AI inflates their count dramatically and makes their behavior less predictable. Unlike a service account with a fixed API call pattern, an agent can request access to resources it wasn’t originally scoped for. That dynamic behavior is why traditional identity governance tools, built around stable human roles, don’t transfer cleanly.
Based on reporting from agentic AI security requires human oversight, originally published 2026-08-06 15:32:00.
