Modern DLP is becoming the control layer for AI adoption

At a glance

What this is: Cyberhaven argues that modern DLP is moving from a blocking tool to the control layer for AI adoption because legacy content filters cannot govern sensitive data flowing into AI apps and agents.

Why it matters: This matters because security teams need to let employees use AI without losing control over sensitive data, policy enforcement, or auditability across human, autonomous, and machine-speed workflows.

By the numbers:

• 39.7% of the data employees share with AI tools is sensitive, according to Cyberhaven research.
• Endpoint-based AI agents grew 509% in 2025, according to Cyberhaven research.

Context

Modern DLP in this article means using data loss prevention controls to govern AI use without forcing the business to choose between access and security. The governance gap is that older DLP assumes predictable data paths, while AI tools and agents move sensitive content across endpoints, SaaS apps, and unmanaged surfaces at a pace traditional inspection models cannot keep up with.

For IAM, PAM, and NHI teams, the relevant intersection is not just content security but control over who or what is allowed to move data into AI systems. Once employees, copilots, and endpoint-based agents all become data movers, policy has to follow the data flow rather than the network perimeter.

The article’s position is that organisations already using AI without a clear governance layer are exposing themselves to both leakage and friction. That starting point is now common rather than exceptional.

Key questions

Q: Where does traditional DLP fail in AI adoption programs?

A: Traditional DLP fails when it assumes data will only move through predictable network paths. AI tools and endpoint-based agents can accept sensitive content on the device, transform it, and send it onward before perimeter controls ever see the release. That leaves security with policy intent but no reliable enforcement at the point of use.

Q: How should organisations govern AI programs before scaling them enterprise-wide?

A: Organisations should define ownership, data quality checks, approval gates and retirement criteria before AI use cases proliferate. Governance works best when legal, privacy, data, security and business stakeholders share responsibility for the full lifecycle, from intake to decommissioning. That prevents fragmented models, unclear accountability and compliance drift.

Q: What are the signs that AI usage controls are not working as intended?

A: Common warning signs include sudden token spikes, repeated 429 responses, uneven consumption across users, and budget overruns that appear before the quarter ends. Another indicator is when lower-priority users consume premium model capacity that should have been reserved for higher tiers. If these patterns show up, the policy design is too loose or the thresholds are not being reviewed.

Q: How should organisations balance AI enablement with security oversight?

A: They should permit experimentation, but only within a model that assigns ownership, sets data boundaries, and tracks usage cost and risk together. The key question is not whether AI is allowed. It is whether every AI-enabled workflow has a clear accountable owner and a documented control boundary.

Technical breakdown

Why legacy DLP misses AI data flows

Legacy DLP was built for a slower world of predictable channels, fixed endpoints, and relatively stable user behaviour. It usually depends on content inspection, keyword matching, or network chokepoints, which means it struggles when data is summarized, rephrased, or pasted into browser-based AI tools. It also has limited visibility once content leaves the original channel. For AI adoption, that creates a structural blind spot: the policy decision happens too late, after the data has already entered an external model or agent workflow.

Practical implication: security teams need to evaluate whether their current DLP can actually observe AI usage at the point of data release.

How endpoint visibility changes control over AI agents

Endpoint-based AI agents operate at the operating system layer, not only inside a browser or corporate proxy. That matters because network inspection cannot reliably see what an agent is doing if it is interacting locally with files, applications, or copied content before sending data onward. In practice, the endpoint becomes the enforcement point where data classification, destination, and user intent can be evaluated together. That shifts the control model from perimeter blocking to contextual policy enforcement tied to the device and the session.

Practical implication: teams should treat endpoint presence as a prerequisite for governing AI agents, not an optional enhancement.

Why data lineage matters more than content matching

Data lineage tracks where information came from, how it moved, and where it went after transformation. For AI governance, that is more useful than a one-time content scan because the original file may be summarized, copied, or repackaged before it reaches a model or assistant. Lineage gives security teams the ability to preserve origin and sensitivity context across those transformations. That makes audit, investigation, and policy enforcement possible even when the content no longer looks identical to the source material.

Practical implication: organisations should anchor AI controls in lineage-aware policy, not only in pattern-based content detection.

Threat narrative

Attacker objective: The practical objective is unauthorized exposure or retention of sensitive organisational data through AI tools that outpace existing governance controls.

1. Entry occurs when employees or endpoint-based AI agents move sensitive business data into AI tools that sit outside legacy inspection points.
2. Credential or policy abuse is not the main issue here; the failure is that the control layer cannot reliably distinguish approved from unapproved AI use or trace what data was released.
3. Impact follows when sensitive data, source code, or internal documents enter external AI systems without enforceable governance, creating leakage, compliance exposure, and audit gaps.

Breaches seen in the wild

• Slack GitHub breach 2022: Slack employee tokens stolen via a compromised vendor were used to download private GitHub repositories over the 2022 holidays.

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

AI adoption governance is becoming a data-control problem before it becomes an AI-policy problem. Organisations are not failing because they lack a rule about AI use. They are failing because the control plane cannot see where sensitive data is going once employees move from sanctioned workflows into AI tools. For identity and access teams, that means the decisive question is no longer who may log in, but what data can leave a trusted context and under what policy. Practitioners should treat AI governance as an extension of data and identity control, not as a separate programme.

Endpoint presence is now a governance requirement, not just a detection feature. Browser-only or network-only models assume the tool is visible at the perimeter, which is increasingly false for copilots, local agents, and OS-level workflows. That creates a control gap where policy exists on paper but not at the moment of release. The field implication is that visibility, enforcement, and auditability must converge at the endpoint if AI adoption is to remain defensible.

Data lineage is the named concept that modern DLP makes operational. Once content can be transformed by AI, the original payload is no longer the only thing that matters. Origin, destination, and transformation history become the basis for policy and evidence. That is a more durable model than regex-driven blocking because it follows the data across AI workflows. Practitioners should design for traceable data movement, not static content signatures.

Machine-speed sharing changes the boundary between human and non-human governance. The article’s own figures point to sensitive content moving into AI tools at scale and endpoint-based AI agents expanding quickly. That means security teams have to govern not only employee behaviour but also the non-human systems that now act on behalf of those employees. The practical conclusion is that NHI, IAM, and data security teams need a shared policy model for AI-enabled data movement.

Blocking alone is an operational dead end for AI governance. The article shows why over-restriction creates shadow usage while under-restriction creates exposure. The right model is not permissive by default, but policy-driven enablement that preserves business velocity and audit evidence at the same time. Practitioners should judge DLP by whether it can allow useful AI work without losing control of sensitive information.

From our research library:

• 28% of secrets incidents now originate outside code repositories, in Slack, Jira, and Confluence, and are 13% more likely to be categorised as critical than code-based leaks, according to the State of Secrets Sprawl 2026.
• Read next: NHI Security Platform Buyer's Guide

What this signals

Modern DLP is becoming the policy layer for AI-enabled data movement. The control question has shifted from whether users will try AI to whether security can govern what leaves the endpoint with enough fidelity to prove it later. That makes lineage-aware enforcement more valuable than static blocking because the business now needs both productivity and evidence.

Data lineage is the practical answer to AI-era audit pressure. A control that can trace origin, transformation, and destination gives teams a defensible story for auditors and board reporting. In Identity and access programmes, that same logic applies to non-human actors: if an agent can move data, its actions need a traceable governance record.

Security teams should expect AI governance to converge with identity governance rather than sit beside it as a separate initiative. When copilots, bots, and endpoint agents participate in business workflows, the policy model must describe who or what is allowed to move which data and under what conditions.

For practitioners

• Define AI data classes and release rules Map the data types employees commonly paste into AI tools, then assign explicit allow, prompt, or block decisions based on sensitivity and destination.
• Extend DLP to the endpoint Validate that DLP can observe AI activity on laptops and local workstations, not just network traffic or browser sessions.
• Track data lineage across AI workflows Preserve origin, classification, and destination context so transformed content can still be governed after summarisation or rephrasing.
• Separate approved AI use from shadow AI Inventory sanctioned copilots, unmanaged chat tools, and local agents, then enforce different controls where the business risk differs.
• Build an auditable AI governance record Document which AI tools are permitted, what logging is required, and which business owners approve exceptions for sensitive data use.

Key takeaways

• Legacy DLP breaks down when employees and endpoint-based AI agents move sensitive data through channels that the control plane cannot reliably see.
• Cyberhaven reports that 39.7% of the data employees share with AI tools is sensitive, which explains why simple blocking is not a viable governance strategy.
• Modern AI governance depends on endpoint visibility, data lineage, and policy-driven enforcement that can preserve both productivity and auditability.

Key terms

• Modern DLP: Modern data loss prevention is a policy and enforcement layer that governs how sensitive information can move across endpoints, SaaS apps, and AI tools. Unlike legacy content filters, it combines classification, context, and destination awareness so controls follow the data rather than just the channel.
• Data Lineage: The record of how data moves across systems, applications, and workflows. In security operations, lineage shows where sensitive data propagates, which identities touch it, and how a compromise could spread across connected environments.
• Shadow AI: AI agents, copilots, or connected tools operating without full visibility or governance from security teams. Shadow AI becomes an identity problem when those systems authenticate with unmanaged tokens, service accounts, or OAuth apps that can reach production resources.
• Endpoint Visibility: Endpoint visibility is the ability to observe activity directly on laptops, desktops, and servers so defenders can see malicious behaviour, credential use, and execution in context. In remote work environments, it becomes a primary control because network location alone no longer tells you whether a device is safe.

Deepen your knowledge

NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, machine identity security, and secrets management. It is designed for practitioners who need to connect identity controls to broader security and AI governance programmes.