Gidi's Substack Articles

The Architecture of Data Security is Starting to Break

Written by Gidi Cohen | May 15, 2026 12:16:50 AM

Original article published on Substack on March 17, 2026.

How AI workflows are redefining what data security must control

Traditional Data Security

For more than two decades, enterprise data security programs have been organized around a familiar and largely effective approach.

Most initiatives have focused on protecting:

  • data at rest
  • data in motion
  • data in use

Supported by familiar control layers including discovery and classification, labeling, identity and access controls, policy enforcement, and detection.

Critically, the approach combined content awareness with identity and access controls, reflecting an environment where humans were the primary actors interacting with data.

This architecture made sense for the world it was designed to protect. Data was largely human-created, human-accessed, and human-shared. Information typically moved in discrete objects - files, records, messages - across relatively predictable workflows.

In that environment, improving visibility and tightening policy coverage delivered meaningful risk reduction.

But the underlying assumptions behind this approach are beginning to shift.

In the previous piece in this series, I described how enterprises are encountering new categories of AI risk, from Shadow AI to what might be called Shady AI, where approved systems behave in ways that violate policy or business intent. These emerging failure modes highlight something deeper: they are not simply operational issues. They reveal a structural mismatch between how traditional data security architectures evaluate risk and how AI-driven systems actually generate and assemble information.

From Human-Dominated Flows to Hybrid Actors

Traditional data security implicitly assumed that humans were the primary actors in the data lifecycle.

Users created documents.
Users sent emails.
Users uploaded files.
Users made mistakes.

Today’s environments look different.

Enterprise data flows increasingly involve a mix of:

  • human users
  • copilots embedded in productivity tools
  • AI features inside SaaS platforms
  • semi-autonomous agents
  • fully automated workflows
  • API-to-API AI interactions

The data flow is no longer human-dominated.

Systems designed primarily for human users are now being asked to reason about copilots, agents, and autonomous services operating at machine speed. This shift begins to stress architectures optimized for user-driven events rather than machine-mediated decision loops.

The Meaning of “Data in Use” Is Changing

In earlier architectures, “data in use” referred both to data being actively processed in application memory and to human interaction with information at endpoints.

This framing was appropriate for environments where processing was largely deterministic and predominantly human-driven.

For example:

  • opening a file
  • copying content
  • attaching a document
  • pasting into a browser
  • sending an email
  • or protecting sensitive data while it resided in memory during application processing

AI systems introduce a materially different pattern.

AI models and agentic systems now routinely:

  • decompose content
  • retrieve fragments
  • recombine information
  • summarize sensitive material
  • infer new attributes
  • generate net-new outputs

Data is no longer only accessed. It is continuously reconstructed during interactions between users, AI systems, and automated workflows.

As explored in the previous piece on the missing dimension of AI data security, sensitivity is increasingly contextual rather than purely intrinsic. The same underlying data may be safe in one interaction and problematic in another, depending on who the information refers to, who is requesting it, and how it is being assembled.

This dynamic behavior begins to expose structural limits in security approaches built primarily around static classification and discrete access events.

Detection and Prevention Under New Pressure

Traditional DLP and related control architectures evolved around a practical balance:

  • broad visibility
  • acceptable noise levels
  • selective inline enforcement
  • post-event investigation when needed

In largely human-paced workflows, this trade-off was often workable.

AI-driven environments compress that margin.

Generation is instantaneous.
Automation amplifies mistakes.
False positives break workflows.
False negatives can propagate at machine speed.

As AI systems increasingly participate directly in content creation and distribution, prevention decisions must become both more precise and closer to the moment of generation.

This is less about adding more policies and more about improving the fidelity of enforcement in highly dynamic flows.

Modern Platforms Have Moved the Bar, But the Ground Is Still Shifting

Modern DSPM and next-generation DLP platforms have significantly advanced data visibility, coverage, and posture management across cloud and SaaS environments. These improvements have meaningfully strengthened many enterprise programs.

The next phase of AI-driven workflows, however, is beginning to stress even these improved security architectures.

The reason is structural.

Where the Architectural Break Is Emerging

These shifts can be understood by comparing how the design center of data security is evolving across three stages: traditional data security, modern cloud-era platforms, and what may be emerging as AI-native data security.

Dimension

Traditional Data Security

Modern Data Security (DSPM and Next-Gen DLP)

AI-Native Data Security

Unit of protection

Files and repositories

Sensitive data elements and stores

Dynamically assembled information in context

Primary actors assumed

Human users

Humans plus SaaS applications

Humans, copilots, agents, and autonomous workflows

Meaning of data in use

Memory processing and user activity

Broader user and application access patterns

Real-time machine-driven generation and assembly

Data behavior model

Largely static

Continuously discovered and scanned

Continuously recomposed and generated

Role of context

Primarily content-centric

Improved metadata and ownership context

Deep entity, relationship, and intent context

Detection vs prevention

Detect-first with coarse prevention

Improved detection with selective inline controls

High-precision, in-flow decisioning required

Precision requirements

Moderate noise tolerance acceptable

Reduced false positives

Near-real-time, high-confidence decisions required

Risk creation point

When data moves or is accessed

During sharing and exposure events

During generation, recomposition, and agent execution

Control architecture

Channel-specific, siloed controls

Broader cross-SaaS visibility

Coordinated multi-channel, workflow-native enforcement

This shift does not invalidate earlier approaches. It reflects a change in where risk is most likely to emerge.

Why Multi-Channel Architecture Now Matters More

The traditional model often treated controls for data at rest, in motion, and in use as largely separate domains.

In AI-driven environments, these boundaries are increasingly fluid.

A single workflow may span:

  • retrieval from a data store
  • processing inside an AI model
  • enrichment via external services
  • generation of new content
  • distribution through collaboration or communication channels

Risk frequently emerges across these transitions, not neatly within one control plane.

As AI-driven workflows expand across email, SaaS platforms, copilots, and agent ecosystems, effective governance increasingly depends on coordinated multi-channel enforcement rather than isolated point controls.

The Design Center Is Shifting

Taken together, these changes point to a deeper conclusion.

AI is not simply increasing data risk. It is changing where and how that risk is created.

Security models optimized for:

  • static objects
  • human-paced actions
  • and discrete movement events

are being asked to reason about:

  • continuously generated content
  • machine-mediated decisions
  • and highly contextual information flows

This is the architectural break.

What Comes Next

As AI adoption accelerates, the gap between traditional control models and real-world data behavior will become increasingly visible.

Forward-leaning security teams are already asking new questions:

  • How do we enforce policy at the moment AI generates content?
  • How do we maintain accuracy when context determines sensitivity?
  • How do we govern workflows that span humans and autonomous systems?
  • Where does the effective control plane for AI-driven data actually live?

Answering these questions will define the next phase of data security evolution.

Closing Thought

The familiar categories of data at rest, in motion, and in use still matter. But what “in use” means is evolving rapidly.

In environments where humans, copilots, and autonomous systems continuously assemble and generate information, data security can no longer focus only on where data resides or where it moves.

The architectural center of gravity is shifting from protecting where data lives to governing how information is created.