Key Takeaways
- Organizations now host an average of 109 non‑human identities for every human employee, yet most security leaders lack visibility into how many AI agents are actually running.
- 82 % of companies have discovered shadow AI—agents spun up without formal approval—that quietly accumulate permissions and retain them long after their useful life ends.
- Only 21 % of enterprises possess a reliable method to shut down an unauthorized agent, leaving a large gap in lifecycle management.
- Identity is the single continuous thread that follows an AI agent from creation to decommissioning, making it the most effective point of control for mitigating risk.
- Token Security helps enterprises gain visibility, enforce least‑privilege access, and automate the safe retirement of AI agents through an identity‑first platform.
- Founder/CTO Ido Shlomo brings elite offensive‑defensive experience from Israel’s Unit 8200, positioning Token Security at the forefront of machine‑identity security.
The Scale of Non‑Human Identities
Modern enterprises operate in a world where machine identities vastly outnumber human users. Recent industry data indicate that for every employee there are roughly 109 non‑human identities, encompassing service accounts, API keys, robotic process automation bots, and increasingly, AI agents. This explosion is driven by cloud‑native architectures, micro‑services, and the rapid adoption of generative AI tools that can be instantiated with a single command. Security teams, however, often lack comprehensive inventories of these entities, resulting in blind spots that adversaries can exploit. Without a clear denominator, risk assessments become speculative, and traditional controls built around human credentials prove insufficient for managing the sheer volume and dynamism of machine identities.
The Rise of Shadow AI
A striking 82 % of organizations have uncovered shadow AI—agents that were spun up without formal approval, oversight, or documentation. These agents frequently emerge from developer sandboxes, CI/CD pipelines, or business‑unit experimentation, where the pressure to deliver quickly outweighs procedural checks. Once deployed, they begin to collect permissions as they interact with data stores, APIs, and other services, often inheriting excessive privileges through overly permissive role‑based access controls. Because their creation is ad‑hoc, they frequently evade change‑management processes, leaving security teams unaware of their existence until an incident or audit reveals anomalous activity.
Permissions Creep and Persistent Access
One of the most insidious characteristics of unmanaged AI agents is permissions creep. As agents perform tasks, they request additional rights to fulfill new functions, and many platforms grant these requests automatically to avoid friction. Over time, an agent originally intended for a narrow data‑enrichment job may accumulate broad read/write access across multiple domains, effectively becoming a privileged insider. Worse, when the agent’s original purpose is fulfilled—or the project is abandoned—its access is rarely revoked. Studies show that only 21 % of companies have a reliable mechanism to shut down an unauthorized agent, meaning the majority retain lingering, over‑privileged identities that can be hijacked or abused long after their useful life has ended.
Challenges in Decommissioning AI Agents
Decommissioning an AI agent is far more complex than disabling a user account. Agents may be embedded within orchestration frameworks, tied to event‑driven triggers, or replicated across multiple environments for resilience. Identifying all instances, understanding their dependencies, and safely revoking credentials without disrupting legitimate workflows requires deep contextual awareness. Additionally, many agents generate ephemeral tokens or short‑lived certificates that are difficult to trace once issued. The lack of standardized deprovisioning workflows, combined with insufficient logging of agent‑to‑resource interactions, creates a situation where security teams must resort to manual hunts or accept residual risk.
Why Identity Is the Critical Control Point
Despite the chaotic lifecycle of AI agents, identity remains the one immutable attribute that follows an agent from its birth to its retirement. Every agent, regardless of how it is created, must present some form of credential—be it a service account, API key, OAuth token, or certificate—to interact with systems. By centering security controls around these identifiers, organizations can enforce consistent policies: least‑privilege assignment, continuous monitoring, automated rotation, and timely revocation. An identity‑first approach also enables correlation across disparate logs and telemetry, providing a unified view of agent behavior that would be impossible when focusing solely on network traffic or application‑level anomalies.
How Token Security Addresses the Problem
Token Security provides a platform designed explicitly for securing Agentic AI and non‑human identities in hybrid and cloud environments. The solution begins with continuous discovery, automatically cataloguing every agent, service account, and machine credential across AWS, Azure, GCP, Kubernetes, and on‑premises directories. Once identified, the platform applies dynamic policy engines that enforce just‑in‑time (JIT) access, ensuring agents receive only the permissions they need for the duration of a specific task. When a task concludes, the system automatically expires or revokes the associated credentials, eliminating lingering privileges.
Beyond provisioning, Token Security delivers real‑time behavioral analytics that detect anomalous permission usage, privilege escalation attempts, or data exfiltration patterns indicative of compromised agents. Alerts are enriched with identity context, allowing security analysts to trace the activity back to the specific agent, its owner, and the originating pipeline. Finally, the platform integrates with existing ITSM and SOAR tools to orchestrate automated remediation—such as disabling a rogue agent or initiating a forensic snapshot—thereby reducing mean‑time‑to‑respond (MTTR) from hours to minutes.
Real‑World Enterprise Implementation
Several Fortune 500 clients have adopted Token Security to curb shadow AI and regain control over their machine‑identity sprawl. In one case, a global financial institution discovered over 12,000 undocumented AI agents operating within its cloud estates, many of which held privileged access to transaction‑processing databases. After deploying Token Security’s discovery engine, the security team mapped each agent to its responsible business unit, applied JIT policies, and retired 8,500 agents that had been idle for more than 30 days. The result was a 68 % reduction in over‑privileged machine identities and a measurable decline in anomalous access alerts. Another client, a healthcare provider, used the platform’s behavioral analytics to detect an AI agent exfiltrating patient records via a misconfigured API; the automated response revoked the agent’s token within two minutes, preventing a potential breach.
Author Bio: Ido Shlomo
Ido Shlomo is the Co‑Founder and Chief Technology Officer of Token Security, where he leads the company’s mission to secure Agentic AI and non‑human identities in modern hybrid and cloud environments. Prior to founding Token Security, Ido served in Israel’s elite cyber intelligence unit, Unit 8200, where he honed both offensive and defensive cybersecurity expertise. His background equips him with a deep understanding of adversary tactics, techniques, and procedures (TTPs) as well as the rigor required to build resilient defenses. At Token Security, Ido focuses on solving one of the field’s most overlooked risks: unmanaged, interconnected, and over‑permissioned machine identities. He is a recognized voice in the cybersecurity community, frequently publishing insights on emerging threats, identity‑first security strategies, and the evolving attack surface.
Conclusion and Call to Action
The proliferation of AI agents and other non‑human identities has outpaced traditional identity‑management capabilities, leaving organizations exposed to significant risk. By recognizing that identity is the constant thread linking an agent’s inception to its decommission, security leaders can shift from reactive patchwork to proactive, policy‑driven control. Token Security demonstrates how continuous discovery, just‑in‑time access, automated deprovisioning, and identity‑centric analytics can close the visibility gap, eliminate shadow AI, and ensure that machine identities remain least‑privileged and short‑lived. Enterprises seeking to strengthen their security posture in the age of Agentic AI should prioritize an identity‑first strategy, leverage purpose‑built platforms like Token Security, and enforce rigorous lifecycle governance for every non‑human identity in their environment.