Zero trust data security now has to govern AI systems and NHIs

At a glance

What this is: This is an analysis of zero trust data security showing that access control must now govern AI systems and non-human identities alongside people, workloads, and applications.

Why it matters: It matters because IAM, NHI governance, and data security teams need one access model that evaluates sensitivity, context, and behaviour instead of trusting network location or prior authentication.

Context

Zero trust data security is the shift away from perimeter trust toward evaluating each access request against identity, context, and data sensitivity. In this article's framing, that model has to extend beyond people to service accounts, workloads, APIs, and AI systems because sensitive data now moves across cloud, SaaS, on-premises, and remote environments.

For IAM and NHI programmes, the governance gap is not authentication alone. The harder problem is deciding whether an identity, human or non-human, should reach a specific dataset under current conditions and whether that access remains appropriate as behaviour, permissions, and risk change.

Key questions

Q: What breaks when zero trust stays focused on the network instead of the data?

A: Zero trust becomes incomplete when it treats network location or prior authentication as proof of trust. Sensitive data can still be exposed through cloud apps, SaaS, APIs, workloads, and AI systems unless access is evaluated against identity, context, sensitivity, and current risk. The missing control is resource-level authorization tied to the data itself.

Q: Why do service accounts and AI systems create extra risk in zero trust models?

A: Because their privileges are often broader, inherited, and less visible than human access. When those identities can reach sensitive data without task-specific limits, compromise or misuse can expose more information than a user-centric model expects. The key issue is effective access, not just configured entitlements.

Q: How do you know if zero trust is actually working for data?

A: Zero trust is working for data only when the organisation can show that sensitive information stays within defined use boundaries after access is granted. The signal is not just successful authentication. It is whether copying, sharing, training, and derivative use are constrained and observable across the environment.

Q: When should organisations prioritise data classification over broader security tooling?

A: Prioritise data classification first when the main challenge is inconsistent protection across different data types. Classification gives every later control a target, including access rules, encryption, retention, and backup decisions. Without it, teams often apply controls unevenly, overprotect low risk data, and underprotect sensitive records that need tighter handling.

Technical breakdown

Why perimeter trust fails for data access

A perimeter model assumes that once an identity is inside the environment, subsequent access is broadly trustworthy. Zero trust data security rejects that assumption and evaluates every request at the resource level. The decision should combine identity proof, device or workload posture, requested asset, data sensitivity, location, behavior, and current risk. That matters because a valid login tells you who authenticated, not whether that identity should read a particular dataset, export it, or pass it to another system. The control objective changes from network admission to continuous authorization around the data itself.

Practical implication: Treat network access as insufficient evidence of trust and move authorization decisions closer to the dataset.

How least privilege changes for non-human identities and AI systems

Least privilege in a zero trust model applies to users, applications, service accounts, workloads, and AI systems alike. For non-human identities, privilege is often inherited, overbroad, or left in place long after the task changed, which creates exposure even when authentication is strong. The important shift is from static entitlement thinking to task-scoped access with ongoing review of what the identity actually uses. That is especially relevant for AI systems and agents because their access patterns can be broader than the human process that created them.

Practical implication: Inventory effective access for non-human identities and remove permissions that exceed the specific task or dataset.

Why continuous monitoring is central to zero trust data security

Zero trust is not a one-time policy decision. It depends on continuous monitoring of data activity, permission changes, unusual access, and risky sharing because access that was appropriate at issuance can become excessive later. Monitoring also creates the feedback loop needed to spot anomalous usage by humans and machines, including downloads, lateral data movement, and unexpected access to sensitive records. In practice, this is what turns zero trust from a static control set into an operating model that reacts to changing risk around the data.

Practical implication: Watch for permission drift and unusual usage patterns, then trigger revocation or step-up controls when risk changes.

Breaches seen in the wild

• Kubeflow cryptomining attacks 2020: Exposed Kubeflow dashboards and a notebook's mounted Kubernetes service account let attackers run cryptominers on tens of ML clusters.
• DeepSeek database exposure 2025: An unauthenticated DeepSeek ClickHouse database exposed over a million log lines with plaintext chat history and API keys in 2025.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

Zero trust data security fails if it stays network-centric: The old assumption was that authenticated users or devices inside the perimeter could be trusted to behave appropriately. That assumption breaks when sensitive data sits in cloud services, SaaS platforms, on-premises stores, and AI workflows that are accessed by both humans and non-human identities. Practitioners should treat data itself as the control boundary.

Least privilege is now a cross-actor governance problem: Users, applications, workloads, service accounts, and AI systems all need access boundaries that reflect actual task need. Static privilege models fail because non-human access is frequently inherited, broad, and difficult to see in ordinary IAM reviews. The governance implication is that effective access must be measured, not assumed.

Identity-aware data context is the new zero trust differentiator: A request to a low-risk resource and a request to regulated or proprietary data cannot be governed with the same confidence level. Zero trust becomes materially stronger when identity, sensitivity, and activity are evaluated together. That is where human IAM, NHI governance, and data security finally converge into one programme decision.

Continuous authorization is the real control model, not one-time authentication: Authentication answers who entered. Zero trust asks whether access should continue, for what data, and under what conditions. For practitioners, the control problem moves from admission to ongoing entitlement validity, which is a much harder but more accurate way to govern modern access paths.

Named concept, data-aware zero trust: The article's central message is that zero trust only becomes operational when data sensitivity is part of the access decision. Without that layer, teams can tighten identity controls and still leave sensitive information broadly reachable. The practical conclusion is that data classification must inform authorization, not sit beside it.

From our research library:

• 59% of organisations say they lack viable alternatives to standing privileged access for NHIs and AI agents, according to Delinea research.
• Read next: Ultimate Guide to NHIs

What this signals

Data-aware authorization is the decisive shift: Zero trust programs now need to evaluate access against the sensitivity of the resource, not just the identity requesting it. That change matters because service accounts and AI systems can otherwise inherit reach that was never intended for the data they touch.

Non-human identity governance and zero trust now overlap: Once workloads, APIs, and AI systems are treated as first-class access subjects, NHI inventory, permission review, and remediation become zero trust controls rather than adjacent hygiene. According to the Ultimate Guide to NHIs, NHIs outnumber human identities by 25x to 50x in modern enterprises, which means this is now a scale problem as much as a policy problem.

For practitioners

• Map sensitive data to actual access paths Build a current inventory of where sensitive data lives and which identities, human and non-human, can reach it through direct, inherited, or indirect permissions.
• Tighten privileges for service accounts and AI systems Review effective access for service accounts, workloads, applications, and AI systems, then remove permissions that exceed the task or dataset they need.
• Use data sensitivity in authorization decisions Apply stronger controls when access requests target regulated, personal, financial, or proprietary data, even if the identity already authenticated successfully.
• Monitor for permission drift and abnormal usage Track downloads, access spikes, unusual sharing, and permission changes so that risky behaviour triggers reassessment before exposure expands.
• Automate remediation for excessive access Create workflows that revoke stale entitlements, reduce broad group access, and assign excessive permissions back to owners for correction.

Key takeaways

• Zero trust data security only works when access decisions are tied to data sensitivity, context, and behaviour instead of inherited trust from the network.
• Non-human identities and AI systems widen the blast radius because they often carry broad or poorly reviewed access to sensitive information.
• Practitioners need one access model that unifies IAM, NHI governance, monitoring, and remediation around the data itself.

Key terms

• Data-centric zero trust: A zero trust model that treats the data itself as the primary control boundary. Rather than relying mainly on network location or device trust, it asks whether a subject, workload, or AI system should access specific data for a specific purpose under current policy.
• Effective Access: The actual permissions an identity can exercise after inheritance, nested groups, delegation, and object-level controls are evaluated. In Active Directory, effective access is more useful than direct membership because it reveals the true operational reach of a service account.
• Non-Human Identity (NHI): A digital identity assigned to a non-human entity such as a software application, service account, API key, bot, machine, or AI agent that enables it to authenticate and interact with systems without direct human involvement. NHIs now outnumber human identities in most enterprises by 25 to 50 times.
• Continuous authorization: Continuous authorization is the practice of rechecking access as a session unfolds instead of trusting a single login decision. It matters for AI workflows because the request, context, retrieved data, and downstream action can all change between prompt and execution, making static approval too blunt.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity lifecycle are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are responsible for identity security strategy or NHI governance in your organisation, it is worth exploring.