The Urgent Need for Zero Trust Implementation in the Age of AI Agents

As AI agents rapidly proliferate within enterprises, a fundamental review of security approaches is becoming necessary. Andre Durand, CEO and founder of Ping Identity, argues that Zero Trust, a security architecture approach, should be implemented immediately as a prerequisite for AI agent operations rather than as a long-term goal. Zero Trust is a security model that does not automatically trust users, devices, or systems, and repeats authentication and verification for every action.

Traditional enterprise security has been designed on the assumption that humans operate systems. It was common to authenticate once at login and then maintain sessions with relatively broad permissions. However, AI agents operate at speeds completely different from humans. Durand states that "human unauthorized access proceeds over minutes to hours, or in some cases days, but at agent speed, 1,000 actions can be executed in 5 minutes," pointing out that the time axis of risk has been fundamentally compressed.

Permission accumulation is one structure through which risk grows. Every time an employee approves an AI agent's request for access to internal drives, databases, or code repositories, the enterprise relinquishes a small measure of control. While individual approvals appear routine, as thousands of agents repeat thousands of requests, their accumulation becomes a significant risk that existing security designs cannot adequately measure, Durand explains.

What Zero Trust emphasizes is simultaneously narrowing two variables: the "scope" and "duration" of access rights. Using Durand's words, it is the idea of "granting only the minimum necessary permissions at the exact moment needed," shifting the focus of security from one-time confirmation at login to continuous judgment with each action. This is an approach fundamentally different from the traditional model of maintaining broad permissions over extended periods.

Furthermore, what concerns Durand is how agent "identity" is handled. Currently, it is common for AI agents to reuse human login credentials or operate under shared service accounts across multiple systems. However, he is skeptical of this practice, stating that "each agent should have its own ID. Rather than impersonating humans, the correct form is to receive explicit permission delegation from humans." The idea is that clearly distinguishing human operations from agent operations is a prerequisite for making Zero Trust function.

Additionally, dependence on shared "secrets" such as API keys is raised as an issue. Many service accounts still rely on such mechanisms, and in environments where agents operate at high speed and scale, the damage range from a single compromised key could expand to an unprecedented degree compared to traditional systems.

As AI agent adoption accelerates, the review of security design has emerged as an urgent task that cannot be postponed. The principle of Zero Trust—assigning unique IDs to agents and verifying permissions with each action—is positioned as a mechanism to address risks that existing frameworks designed for human-centric systems cannot handle. As AI agent adoption becomes more widespread, it can be said that the design of identity and access management will increasingly determine the overall strength of enterprise security.