When your most powerful users aren’t human: Managing AI and NHIs for compliant privileged access

The identities of the most powerful users and administrators in many organizations aren’t people. They’re non-human identities (NHIs). As the number of NHIs within organizations continues to rise sharply, a growing proportion is no longer tightly bound to direct human configuration or oversight.

Others, namely AI agents, execute high-risk functions at different levels of autonomy. They perform tasks that range from analyzing data to deploying code, at a volume and velocity far beyond human capabilities.

Yet many privileged access security programs are still designed for human administrators. This mismatch leaves gaps in the attack surface, with identities and logins falling outside governance controls. This is where identity-based attacks exploit misconfigurations and compliance gaps. Privilege is expanding faster than it is being controlled – and NHIs are the reason.

Why the riskiest logins no longer come from human identities

Access is still being granted via standing privileges, giving agents always-on access to repositories and resources. Of course, this unlocks productivity gains, but it also means any misconfiguration, exposure or incident is magnified and replicated at significant speed.

This stood out in the Cloudflare outage late-2025, where a change in database permissions led to the Bot Management system automatically generating excessive outputs. The scale of the symptoms was likened to a hyper-scale DDoS attack.

What’s more, these machine identities come at a scale that outnumbers human workers by up to 82:1. When tools are fragmented and policies are inconsistent, CISOs have no way of accurately understanding and mitigating exposure to AI and NHI-based attacks. IT and security leaders need unified, real-time visibility for this new, fast-growing digital workforce.

The new privileged majority: AI and NHIs

Digital entities include service accounts, APIs, machine identities, containers, RPA bots, AI agents and workloads. They are used by software, applications and other resources for authenticating and gaining access to systems.

The risk is not simply that NHIs have access, but that many hold persistent or elevated privileges across critical systems. When that access is always-on, poorly scoped or not regularly rotated, it undermines Principle of Least Privilege (PoLP) and Just-in-Time (JiT) access strategies. The 2024 Internet Archive breach illustrates the point: Failure to rotate exposed API keys gave adversaries unauthorized Zendesk access to 800,000+ support ticket details.

When credentials are exposed, deletion is not remediation. Removing the files or repositories that contain secrets may reduce visibility, but it does not invalidate the credentials themselves. The Nx “s1ngularity” incident showed why this matters: Malicious packages harvested GitHub tokens, npm keys, SSH keys and AI tool credentials, then published the stolen data into public GitHub repositories. Effective response requires more than cleanup. Organizations need to revoke and rotate exposed secrets, confirm where those credentials are used and close the access paths they enable. Without integrated visibility across repositories, pipelines, vaults, workloads and identity systems, this becomes slow, manual, incomplete and expensive.

Why traditional PAM and stitched-together stacks fall short

Picture the typical vault-centric privileged access management (PAM) approach, where an administrator authenticates an access request, lets the vault release and allows the hardcoded secret to be used. That request may come from a static user identity and be time-bound for a process designed for human administrators to review, approve, deny or revoke.

All those actions and more were already overloading busy service desks, even when requests were primarily from employees. Now, add dynamic NHIs and fast-evolving AI systems with continuously evolving identities. It becomes harder to maintain a single view, leading to fragmented audit trails and policies that aren’t built for hybrid or cloud-first environments.

This governance gap increasingly contributes to compliance problems, not just security ones.  ISO/IEC 42001 is the world's first international standard for AI management systems, and it specifies requirements for establishing, implementing, maintaining and continually improving how organizations govern AI responsibly. AI governance fails if it only governs the model. ISO/IEC 42001 pushes organizations to look across the AI lifecycle, and PAM extends that lifecycle view to the identities and privileges that make each stage accessible. Where traditional PAM treats AI agents as little more than service accounts, ISO/IEC 42001 demands structured risk assessment, accountability and lifecycle controls mapped to every AI system.

Organizations still relying on vault-centric PAM will struggle to demonstrate compliance with this standard, because traditional PAM was never designed to govern AI at machine scale.

It’s clear that modern stacks and their bolt-on convergence need a different form of integration. Where identity security and the identity lifecycle are managed consistently across different directories, with security aligned to AD and Entra ID. Traditional, vault-centric PAM was built for human administrators. It was never designed to control privilege at machine scale.  Modern PAM is required to protect all identities – human, non-human and AI – with session recording and access mapped to all individual identities, privilege and accountable ownership.

Naturally, there’s not enough time to manually view sessions and log files generated by multiple NHIs. This is another reason why a different, modern strategy with several core capabilities is needed.

As noted in KuppingerCole analyst research, “As identity increasingly becomes the primary control boundary within distributed infrastructures, the management of privilege takes on a central role in enterprise security. PAM solutions provide the mechanisms for governing the most powerful permissions within an organization, ensuring that elevated capabilities, whether exercised by people, machines or automated processes, are controlled, monitored and fully auditable.”

What compliant privileged access and governance really looks like

To govern NHIs that are constantly evolving and adapting, management must be similarly dynamic, real-time and consistent. ISO/IEC 42001 requires a structured framework for governing AI systems, covering risk identification and mitigation, transparency, accountability and a process for reviewing AI performance and refining governance strategies.

ISO/IEC 42001 focuses attention on responsible AI management, risk, impact, transparency, accountability and lifecycle control. But AI systems do not act in isolation. They rely on non-human identities, service accounts, API keys, secrets, workloads and integrations to access data and perform actions. This is where ISO/IEC 27001 and ISO/IEC 42001 overlap most directly. If an AI system can access production systems, sensitive data, identity stores, ticketing systems, code repositories or administrative consoles, then privileged access is part of the AI impact surface.

 KPMG In practice, aligning to both modern PAM and ISO/IEC 42001 builds an identity platform that is:

• Continuous: New NHIs and AI agents must be discovered and classified in real-time. This adds a layer of security that goes beyond rigid intervals, which can widen the attack surface through the risk of orphaned and unmanaged identities.
• Policy-driven: Machine identities must be granted privileges with granular attributes based on least privilege and JIT. This helps to control and enforce access based on risk signals, behaviors and attributes.
• Proactive: For NHIs, enforcement of strong authentication and secret management, with rotation by design and segmentation based on connections instead of sessions, passwords or credentials. By enforcing at this level, identity threat protection can be proactive rather than reactive and post-breach.
• Machine-driven: Monitoring and analytics must be mapped to machine identities, building up a baseline for quickly recognizing and mitigating against anomalous behaviors. Any deviations must trigger context-based responses, from requesting extra credentials to revoking access.
• Evidence-based: Consistent visibility across environments, both cloud-based and on-premises, showing NHI owners, functions and access review dates is essential. Machine identities can then be governed with full accountability while replacing persistent privilege with controlled, evidence-based access.
• Unified Identity: To manage and control the risk of identity sprawl from so many NHIs, monitoring must span IGA, PAM, AM and AD security. Otherwise, organizations are vulnerable to visibility gaps that come from silos within their ecosystems.

The practical roadmap for CISOs

Step 1: Build inventory

It starts with scoping which identities you need to secure and protect. This means creating an inventory of all NHI and AI identities with privileged access, including the ability to access sensitive data, issue credentials or elevate permissions to move within the network. Then, these items must be ranked in terms of risk, across parameters such as environment location, authentication type and access level.

Any orphaned privileged identities must have an owner and purpose. AI-based identities need classifying based on their level of autonomy and level of access to sensitive and downstream systems, including any NHI that may be triggered or involved in the process. This inventory is the essential starting point for demonstrating compliance. Under ISO/IEC 42001, organizations must define and operate a documented AI system impact assessment process, executed at planned intervals or when significant changes occur.  

Step 2: Assign ownership

Each AI system and non-human identity must have a named owner, regularly certified by the business, who stands accountable for ongoing governance. That means every service account, API token, workload identity or AI agent must have a defined purpose, access scope, review cycle, secrets rotation policy and retirement path. Modern PAM tooling provides the mechanisms to manage these highly dynamic controls at scale.

This overarching visibility ensures identities aren't left without owners and defensible policy controls, with any orphans automatically isolated or deleted. The lifecycle must have specific stages mapped, from initial request and approval to deprecation and decommission. ISO/IEC 42001 requires organizations to proactively manage AI risks through structured, repeatable processes — and to demonstrate compliance through audit-ready documentation and performance evaluation.

Step 3: Harden access

NHIs and AIs marked with high-risk privileges must be moved to a JIT policy-based model far from standing privilege scenarios and toward real-time enforcement with conditional access and auto-expiration. This is where hardcoded credentials and secrets are replaced with identity-based systems for authorization and access.

The objective is to have zero identities with permanent privileged access. This limits sprawl and allows for greater granularity and context when managing access based on roles, policies or other attributes, leaving humans to focus more on the highest-risk privileges where they still need to manually approve requests.

Step 4: Monitor and detect

With the inventory as a foundation, the business can monitor and detect machine-based behavioral patterns. Because NHIs are primarily used for repeatable processes, these are more predictable than human activities. For example, the frequency of API calling and the endpoints involved. So, any small deviation from the established baseline, such as an unexpected privilege escalation, can be identified faster and remediated automatically.

Over time, the inputs and outputs can be fed back into the platform used for monitoring and detection, improving automated outcomes and reducing false positives.

Step 5: Automate audits

More automation gains come from making audits repeatable and consistent. Any compliance-related evidence can be gathered on an ongoing basis, including reports, attestations and configuration baselines. These help organizations to answer fundamental questions around audits, such as who owns the NHIs, what systems the NHIs can access, why they have access to that level, and the date of most recent review.

 ISO/IEC 42001 uses Annex A reference controls to support AI risk treatment across areas such as policy, accountability, impact assessment, AI system lifecycle, data management and third-party relationships. Certification depends on demonstrating that relevant controls have been selected, justified, implemented and evidenced through the AI management system.

A unified identity security platform: The first line of defense

IT and security leaders face an emerging and dynamic challenge as more NHIs enter environments. Privileges are no longer vaulted and visible from one centralized location where access can be monitored manually. Legacy PAM solutions are no longer suitable in the face of NHI and AI growth, with machine identities as a powerful majority for many businesses.

That’s why a new system is needed, spanning IGA, PAM, AM and AD security. This system should be unified, with core capabilities around continuous discovery and classification, policy-driven PoLP and JIT strategies, plus consistent, standardized, scalable lifecycle and governance models. It needs monitoring and analytics that bridge silos to create reliable and robust behavior baselines.

This starts with the 5-step roadmap outlined above, which is designed for strategic questions around compliance and assurance that demand board-level answers.  

AI and non-human identities are already operating across your environment with privileged access. If you’re not controlling that access dynamically, your identity strategy is already behind.

You can find out more about the topics, discover how to assess your NHI exposure and explore how to quickly close the gaps in tackling non-human identity (NHI) threats. The on-demand webinar features industry experts discussing the biggest NHI risks, how to avoid the pitfalls and how to protect identities. The high-level content will appeal to CISOs, security and IAM decision-makers along with senior IT leaders and is available now for you to watch.