By: Nick Apostolu, Manager, Product Marketing, Industries, Okta
Federal agencies already struggle to govern identities distributed across cloud platforms, legacy systems, contractors, and mission partners. The rise of agentic artificial intelligence (AI) is making that problem more urgent by introducing autonomous software that can retrieve data, call application programming interfaces, and take actions on behalf of users – often with limited human supervision.
These agents can automate complex workflows and support faster service delivery. But they add a new category of identity to environments where agencies may already lack a complete view of who or what can access sensitive information.
The National Institute of Standards and Technology’s National Cybersecurity Center of Excellence (NCCoE) is examining how existing identity standards and practices can be applied to software and AI agents. Early this year, the center published a draft concept paper on software and AI agent identity and authorization.
The potential NCCoE project would explore standards-based methods to identify agents, manage their permissions, authorize their actions, and connect those actions to an accountable person or organization. Although the work is in its early stages, it reflects growing federal recognition that autonomous software cannot be governed as an invisible extension of a human user.
Unlike a conventional application or service account, an AI agent may independently choose among tools, retrieve records, generate content, deploy code, or initiate transactions. If an agency cannot identify the agent, determine who owns it, understand its delegated authority, or see which data it can reach, the agent becomes another poorly governed privileged identity.
The Cybersecurity and Infrastructure Security Agency and international partners highlighted the governance risks associated with agentic AI on May 1 with guidance on its careful adoption. The guidance warns that agentic AI can create expanded attack surfaces, privilege creep, behavioral risks, and gaps in activity records. It recommends beginning with controlled deployments, limiting access to sensitive resources, and bringing agents under existing cybersecurity and risk management processes.
Agentic AI can amplify identity sprawl
For agencies, agentic AI is not a separate security problem. It exposes and intensifies weaknesses in existing identity governance.
Human and non-human identities are already scattered across directories, cloud platforms, privileged-access systems, application-specific accounts, and on-premises infrastructure. That fragmentation – often described as identity sprawl – can leave security teams with an incomplete view of access privileges, ownership, and emerging threats.
Separate identity systems can also produce inconsistent policies and records. Security teams may need to correlate logs from multiple environments before determining whether an account has been compromised, retained access after a personnel change, or accumulated privileges that are no longer necessary.
From our work with public-sector organizations at Okta, we see the identity challenge becoming more difficult as agencies add cloud services, automated workloads, and AI-enabled tools. Federal agencies cannot effectively manage access risk when they cannot consistently identify, govern, and monitor the entities operating across their environments.
The problem extends beyond authentication. Agencies must manage the full identity lifecycle, including provisioning access, monitoring behavior, modifying permissions, reviewing entitlements, and promptly removing access when an employee, contractor, application, service account, or agent no longer needs it.
Federal zero trust and digital identity policies already establish identity as a core security control. Critically, it must be continuously managed. If it is not, an agency may strongly authenticate a user while continuing to grant access through obsolete group memberships, duplicate accounts, or permissions retained after a job change. The same problem can apply to an AI agent whose initial access was approved but whose role, integrations, or capabilities have since expanded.
Reducing identity sprawl requires agencies to maintain an authoritative record for every user, device, workload, service account, and AI agent, with a defined owner, purpose, privilege level, and review or expiration point. Provisioning and deprovisioning should be automated where possible, and access policies should apply consistently across cloud and on-premises environments.
AI agents add several specific requirements. Agencies need to define which systems and datasets each agent may access, whether the agent can delegate authority, which actions require human approval, and how the agent’s activity will be logged and reviewed.
Privileges should be narrowly scoped and reassessed whenever an agent’s mission, model, tools, or integrations change. An agent should not inherit a human user’s full range of permissions simply because it performs a task on that person’s behalf.
Agencies also need sufficient records to reconstruct an agent’s actions after a security incident or unexpected result. Those records should show which identity initiated the activity, what authority was delegated, which systems were accessed, and whether the agent remained within its approved boundaries.
The practical test is whether an agency can quickly determine what an entity is, who is responsible for it, which resources it can reach, why that access exists, what actions it has taken, and how its access can be suspended.
As federal systems become more autonomous, the ability to answer those questions will determine whether agencies can adopt new capabilities without creating unmanaged risk faster than their teams can address it.
To learn more about overcoming identity sprawl, read Okta’s e-book.