Agentic AI systems are designed to operate autonomously, moving across networks, requesting data, and using tools to complete tasks without constant human direction. That behavior doesn’t fit neatly into traditional Zero Trust models that focus on least-privilege access for human users and devices.

According to Intelligence Community CIO Douglas Cossa, this new class of AI has effectively turned classic Zero Trust “on its head.” Instead of starting from a stance of minimal or no access, organizations often need to give AI agents broad access so they can operate independently. That creates tension with long-standing security principles.

To adapt, the Intelligence Community is reshaping Zero Trust around two pillars:

• Identity as the foundation: Treating AI agents as first-class identities, not just background processes, and giving them clearly defined, verifiable digital identities.
• Fine-grain policy enforcement: Using detailed entitlements and attributes to control exactly which data and functions each AI agent can reach, rather than broad, static permissions.

In this reimagined model, Zero Trust becomes less about blocking activity and more about enabling the right AI-driven functions to access the right data, under tightly controlled conditions.

The Intelligence Community is exploring the idea of a “digital birth certificate” for AI agents as a way to establish and manage their identities across agencies.

Today, there is no unified identity system for non-human users like autonomous bots. As AI agents begin to request, store, manipulate, and process data at scale, that gap becomes a core security risk.

A digital birth certificate would:

• Uniquely identify each AI agent from the moment it is created.
• Record key attributes such as its purpose, owner, and authorized environments.
• Serve as the basis for permissions—what data it can access, what tools it can use, and what actions it can take.

Cossa’s office is investing in an enterprise identity management service to support this approach, with plans to pilot and test tools in operational environments as the Intelligence Community moves into fiscal year 2027. The goal is to make identity the starting point for any decision about what an AI agent is allowed to do.

US Special Operations Command (SOCOM) is rethinking how it defends networks in an environment where both attackers and defenders are using agentic AI.

Adm. Frank Bradley emphasized that future defenses cannot rely on humans manually reviewing logs or reconfiguring trust settings during a crisis. Instead, SOCOM is working toward networks that can:

• Detect compromise in minutes, not months, by continuously monitoring for anomalies.
• Incorporate context—such as device health, location, and behavior—into access decisions.
• Respond automatically, enabling what Bradley called “agentic defense against agentic offense.”

At the same time, SOCOM expects adversaries to focus more on human frailty—lapses in discipline, protocol failures, or simple exhaustion—rather than purely technical flaws. To design for that reality, they are emphasizing:

• Layered defenses that don’t rely on a single control.
• Compartmented access so that one mistake doesn’t expose everything.
• Need-to-know restrictions enforced at the data level to contain the impact of human error.

Together, these shifts show how defense organizations are using automation and fine-grained controls to reimagine Zero Trust for a world where both machines and people are active participants in cyber operations.