DPACT Framework Offers Blueprint for Securing AI Agents Through Delegation and Bounded Authority

As AI agents evolve from passive conversational interfaces into autonomous systems capable of taking real-world actions—accessing sensitive data, modifying records, making decisions—the industry faces a critical security gap. Traditional identity and access models built for human users break down when autonomous actors enter the picture, creating risks that most teams racing to deploy agents have yet to fully address.

The Challenge: Moving Beyond Token-Based Security#

The shift toward agentic systems represents a fundamental change in how software operates. Rather than humans invoking discrete actions, agents now function as delegated actors that can chain multiple operations, exercise judgment, and interact with production systems autonomously—all while potentially accessing credit card data, customer information, and proprietary knowledge. Yet many organizations have deployed agents with the same permission models designed for human authentication, leaving significant security blind spots.

The core distinction that matters: agents should act on behalf of users, never in place of them. When an agent impersonates a user—claiming to be that person rather than representing their delegated authority—it bypasses the accountability structures that prevent privilege escalation and unauthorized data access. This distinction separates responsible automation from a security nightmare.

DPACT: A Framework for Bounded Agency#

To address these challenges, security leaders are adopting the DPACT framework, a five-pillar approach to agent governance:

• Delegation: Clearly identify on whose behalf each agent operates and ensure agents never masquerade as human users
• Policy: Define both capabilities and absolute restrictions—what an agent can do and, critically, what it cannot
• Auditability: Maintain comprehensive logs explaining why agents took specific actions, in what context, under what authority
• Context: Draw clear operational boundaries; for example, allow an agent to draft content but prohibit sending without human review
• Time: Make authority expiration a first-class feature; avoid perpetual access grants tied to long-lived API keys

This framework treats each component as essential from the start, rather than layering security reactively. Delegation anchors the entire model—without clear delegation, the remaining controls lose effectiveness.