AI agents should receive narrowly delegated authority instead of inheriting a user’s full identity, according to a security framework discussed in a new InfoQ interview. The proposal, called DPACT, organizes controls around delegation, policy, auditability, context and time.
The distinction matters as agents move beyond chat and begin operating software, handling data or making transactions. If an agent simply impersonates a person with a long-lived token, a mistake or compromise can expose every resource that person can reach. Delegation instead creates an explicit grant for a task, with a defined scope and duration.
Under DPACT, policy determines which actions are permitted, auditability records what the agent attempted, and context narrows decisions to the relevant user, data and situation. Time limits prevent an old credential from remaining useful after the task ends. The approach also makes it clearer that an agent acted on behalf of someone rather than as that person.
The framework is guidance, not a product standard or proof that a system is secure. Its practical starting point is incremental: inventory the agents already operating in an organization, add visibility into their actions, then replace broad credentials with bounded grants. That can reduce exposure without requiring companies to halt every agent deployment while a complete governance system is built.