

All the talk about AI threats today surrounds agents, which are quickly being adopted into enterprise workflows with or without adequate guardrails. AI agents largely work autonomously, reading data and calling APIs, and triggering actions across applications with little human involvement. In addition to autonomy, these nonhuman identities are used in numbers that significantly outweigh human users.
Service accounts and machine identities were traditionally designed to perform fixed, predictable tasks. AI agents act differently. They can initiate actions independently and connect to several systems while collaborating with other agents in a single workflow, and operate with varying levels of autonomy.
With identity a central focus of cyberattacks, the tidal wave of agentic AI creates a novel threat vector. Security teams need a grasp of what agents can access and what actions they can perform. They also need to know whose authority agents operate under. Existing identity frameworks provide organizations with a foundation, but they also may need to evolve so that they can answer these questions as AI adoption grows.
Nonhuman identities (NHIs) are digital credentials that work on behalf of software platforms and automated processes. Common use cases include service accounts, API keys, tokens, and certificates.
AI agents extend the definition of NHIs and the cybersecurity implications that come with them. They can act for users or systems, which raises new questions for identity and access management (IAM). And they may operate through several access models.
Some agents are assigned a dedicated identity, while others run under a shared service account or borrow delegated user permissions. Many call systems with API credentials. Single agents can combine more than one model within one workflow. In short, they present a complex case for identity-centered security.
Security teams can use authentication information as a starting point for tracking down agents who exceed their required authority. Authorization and context then determine whether a given action is permitted. Even an authenticated agent has the potential to have broader access than necessary to complete a single task.
By aligning these systems together, security teams gain knowledge about what applications, APIs, and sensitive data an agent can access. This information is valuable to security teams because it can help identify where a poorly configured agent, or one being used maliciously, could cause damage.
Agent identity becomes much more powerful when the rights granted to the agent are tied to an intended purpose. This provides security teams with the ability to verify that the right to perform a particular action is correct for the tasks assigned to the agent. It also allows security teams to track all actions taken by the agent.
Excessive permissions let an agent do more than its task requires. Persistent credentials, inherited user rights, and broad API scopes are common sources. Wider access increases the potential impact of any mistake or misuse. Limiting access to the task at hand can help reduce that exposure.
Research from Zscaler shows how quickly weaknesses can appear. In red team testing for its ThreatLabz 2026 AI Security Report, researchers found critical flaws in 100% of the enterprise AI systems analyzed. The median time to first critical failure was 16 minutes. These results come from controlled adversarial testing, so real-world outcomes may vary. They suggest that narrowly scoped access can matter when systems face pressure.
Agent autonomy raises the stakes. A compromised credential, a manipulated instruction, or an unintended action may spread across several systems before anyone notices. The OWASP Excessive Agency guidance names excessive functionality, excessive permissions, and excessive autonomy as typical root causes of damaging agent actions.
Least privilege offers a practical response. Task-specific permissions grant access for one job and can expire when the job ends. Short-lived credentials and narrowly scoped API tokens support this approach. The same OWASP guidance also suggests human approval for high-impact actions, which can add a useful checkpoint.
Agents can delegate tasks to other agents. Each handoff may cross applications, data stores, and trust boundaries. Attribution, authorization, and accountability can blur along the way. Security teams need visibility into the whole chain of access.
Consider a multi-agent workflow. An orchestrating agent calls a second agent, which calls a third agent that queries a database. Logs can show a record that ties back to the final identity. But that log makes it perplexing to determine which one initiated the request and whether the steps it took stayed within its authorized boundaries.
The NIST NCCoE concept paper on AI agent identity and authorization explores how organizations may identify, authenticate, and authorize agents that act inside enterprise systems. This work suggests that governance may need to follow delegation from end to end. Chain-level visibility can help teams verify permissions at each step and assign responsibility when outcomes differ from intent.
The first step in managing agent identity is recognizing that they exist. An organization needs to identify its agents, their credentials, privileges, and the systems they can access. Once this is done, continuous authorization, credential lifecycle management, periodic access reviews, and behavior monitoring can ensure that a system user or application has ongoing and proper access.
Continuous discovery includes both authorized and unauthorized agents. As stated above, the relationship between users and applications may span multiple applications. Perform authorization on a per-request basis rather than granting one-time authorization.
Establish clear identification and rotation/retirement processes for all credentials used by each agent. Periodic access reviews are necessary to determine whether each agent is still required to perform the functions associated with the access rights assigned to them.
These practices build on established principles. Zscaler's emphasis on zero trust is based on an approach that verifies every access request and limits access through least privilege. Existing IAM programs provide a strong foundation, and agents become another identity type for those programs to cover.
Identity governance is no longer limited to employees and machine identities, as we now have a growing number of autonomous AI agents handling the work. Securing these identities involves understanding who or what an agent is, what it is authorized to do, and how it uses that access.
Zscaler’s efforts in zero trust and AI security are a testament to how cybersecurity innovation is contributing to a larger industry movement toward modifying Identity and access controls to accommodate agentic environments. Organizations that link together their identity management systems, permissions systems, and behavior monitoring systems might be able to effectively govern these agents as they complete more tasks.