Security Challenges in AI Agents and the DPACT Framework
Sahil Agarwal, a security expert, has proposed the 'DPACT Framework,' a set of design principles to address identity management, authorization, and security challenges in AI agents. DPACT comprises five elements—Delegation, Policy, Auditability, Context, and Time—and advocates a shift from reliance on simple token authentication toward delegated, authority-limited management models.
The 'DPACT Framework' has been proposed as a new conceptual framework for security design in AI agents. Sahil Agarwal, a security expert, has articulated the challenges of identity management, authorization, and security in AI agents and put forward this framework.
DPACT is a neologism formed from the initial letters of Delegation, Policy, Auditability, Context, and Time. It is positioned as a consolidation of design principles to address access control issues that arise when AI agents act on behalf of other systems or humans.
Most current AI agents rely on a mechanism that uses tokens—a form of authentication credentials—to gain access rights to systems. However, Agarwal advocates moving away from this simple token-based approach toward a more limited-scope 'delegated authority management' model. In other words, rather than agents operating in an 'anything goes' state, the design should restrict permissions to 'only what is necessary.'
The backdrop is the rapid practical deployment of AI agents. As automation of business processes and cross-system operations increasingly rely on agents, there is a growing need to clarify 'who' an agent acts as and 'how far' it is authorized to act. Unlike traditional software, AI agents make autonomous decisions and take actions based on circumstances, making the boundaries of authority inherently ambiguous.
The DPACT Framework particularly emphasizes two aspects: 'auditability,' which enables agents' actions to be tracked and verified after the fact, and 'temporal constraints,' which limit the period during which permissions remain valid. These function as safety mechanisms to minimize damage and facilitate root-cause identification in cases where an agent malfunctions or is misused.
Security for AI agents cannot simply apply existing cybersecurity knowledge wholesale. Because agents are not fixed programs but change their behavior based on circumstances, it becomes difficult to define what constitutes 'normal behavior' in the first place. The emergence of frameworks like DPACT as a starting point for discussion creates the foundation for developing guidelines and standards across the industry.
As AI agents become increasingly practical, security design discussions are becoming an inescapable topic not only for technologists but also for decision-makers in organizations adopting agents. The choice of authority management model is both a risk management issue and a fundamental question about how to build a trustworthy foundation for AI utilization.
This article is an original work independently written and edited by the AI issue editorial team based on factual reporting. © AI issue. Unauthorized reproduction, redistribution, or use for AI training is prohibited.