The Identity Problem at AI Scale: Why Agentic AI Demands More From OAuth - Source: securityboulevard.com

AI agents aren’t theoretical anymore. They’re active in our enterprises—making decisions, chaining API calls, interacting across clouds, and operating autonomously at machine speed. But as organizations move from predictive AI to proactive, agentic AI, our identity systems hit a wall.

AI agents create entirely new challenges for identity and access management. At Strata, we’ve seen this first-hand. Let’s break down the core problems that drive the need for a modern agentic identity layer—one powered by OAuth but extended for Zero Trust and machine-speed operations.

The Problems Driving the Need for Each Agentic Identity Feature

Delegation Chaos: The Need for On-Behalf-Of (OBO)

AI agents rarely act on their own authority—they act on behalf of users or other systems. Without a standards-based way to represent these delegation chains:

There’s no clear, auditable link between agent actions and who authorized them.
Security teams can’t enforce or review delegation policies in real time.
Accountability breaks down during investigations or compliance reviews.

We need OBO to securely and transparently bind agent actions to their delegators.

Cross-Domain Gaps: The Need for Token Exchange

AI agents don’t operate within a single cloud or API domain. They cross trust boundaries constantly. Without secure, standards-driven identity propagation:

Trust breaks down between systems, creating integration gaps.
Agents get stuck at cloud or API borders, or worse, get over-permissioned to compensate.
Identity silos undermine Zero Trust architectures.

We need token exchange to propagate identity securely across clouds and APIs.

Token Theft Risk: The Need for DPoP

AI agents operate in high-churn, distributed environments. Tokens are transmitted frequently, creating an expanded attack surface. Without proof-of-possession protections:

A stolen token can be replayed by attackers anywhere.
Token misuse becomes invisible until after damage is done.

We need DPoP to cryptographically bind tokens to agent keys and prevent replay attacks

Untrusted Agent Flows: The Need for PKCE

AI agents often can’t store client secrets—especially in public or dynamic environments. Without a secure way to handle authorization code exchanges:

Agents are vulnerable to interception and code injection.
Public clients become a weak point in the identity chain.

We need PKCE to secure agent flows without relying on static secrets.

Static Authorization Fails: The Need for CAEP

AI agents operate at machine speed, and their context changes constantly. Relying on static token lifetimes:

Leaves organizations exposed between issuance and expiry.
Prevents real-time adaptation to risk changes (e.g., abnormal behavior, compromised agent).

We need CAEP to dynamically enforce authorization and revoke access the moment risk conditions change.

Oversimplified Access: The Need for Attribute-Based Authorization

Traditional scope-based access control is too blunt for agentic AI. Without attribute-based decisions:

Policies can’t account for task-specific purpose, context, or dynamic conditions.
Agents end up over-permissioned, or critical workflows get blocked unnecessarily.

We need attribute-based authorization to apply fine-grained, intent-aware Zero Trust policies in real time.