As AI agents and SaaS platforms now exchange data autonomously, security leaders must rethink what anchors data protection in 2026 and beyond.
This article was originally published on Cybersecurity Insiders.
‍
For decades, security teams operated under a simple assumption: enterprise data lived inside the perimeter. Defenses were built around that assumption, fortifying networks and endpoints as if all the information worth protecting sat neatly on laptops, servers, and a handful of carefully procured cloud apps.
‍
Today, the most critical data lives inside dozens—or hundreds—of SaaS and cloud platforms. Yet many security controls were built for different assumptions. Endpoint agents, VPNs and web proxies, and traditional DLP tools still hark back to a time when “data protection” meant banning USB dongles.
‍
According to IBM’s 2025 Cost of a Data Breach Report, nearly half of all breaches now involve data stored in cloud environments. Enterprise data has officially escaped any single, defensible perimeter.
‍
The rise of SaaS and AI has created an new ecosystem, where data flows freely across cloud applications, often without any human in the loop.
‍
Legacy controls were designed for a traditional hub-and-spoke model: data left an application, passed through an endpoint the enterprise owned, and went somewhere else. Today, that model has given way to an interconnected web of machine-to-machine communication. APIs and non-human identities are constantly moving data between cloud systems, frequently outside of direct enterprise visibility or control.
‍
Gartner projects that by 2026, non-human identities will outnumber human identities by a factor of two, amplifying the challenge of securing these automated exchanges. As one security veteran recently put it, “It’s all just TLS connections between machines—and we know how to monitor that.” That may be true in theory. But when those machines are SaaS workloads running in cloud accounts owned by vendors, the opportunity to meaningfully observe (or influence) that traffic has effectively left the building.
‍
This is what makes the current moment different. The most consequential data movements are no longer initiated by a person on a managed device. They are automated, persistent, and increasingly decoupled from the controls security teams historically relied on. Our technology hasn’t failed us. Our assumptions have.
‍
A CISO recently told me, “We’ve got offboarding covered—we disable the user and lock the laptop within 15 minutes.” That’s a comforting thought, but it’s a relic from 2000. Today, laptops are often little more than remote controls we use to access SaaS and AI applications running on the internet.
‍
SaaS platforms like OneDrive, GitHub, Salesforce, and Google Workspace contain the files, code, customer data, and spreadsheets that fuel modern work. Even modern software development often happens almost entirely within the browser, with AI agents writing, reviewing, and deploying code without ever touching a local environment.
‍
In this SaaS-first world, risk is no longer defined by where data exits the perimeter, but by how information is stored, accessed, shared, and acted on across systems you don’t own.
‍
As endpoints and networks lose their role as primary control points, one risk surface has quietly expanded in their place: non-human identities.
‍
SaaS platforms rely on OAuth grants, API tokens, service accounts, and delegated permissions to move data between systems. These non-human identities are essential to modern automation—but they’re also long-lived, frequently over-permissioned, and rarely revisited. Unlike humans, they don’t get promotions, go on leave, or get offboarded.
‍
The recent Salesloft-Drift breach is a telling example. Attackers didn’t need a zero-day exploit or custom malware. They simply abused stolen OAuth tokens—one of the most common building blocks of SaaS and AI connectivity. The incident underscored just how fragile and over-trusted today’s web of SaaS and AI integrations has become.
‍
This is where many security teams are now blind—not because data is hidden, but because trust was extended once and allowed to quietly compound.
‍
If data no longer lives in a single place, protecting it can’t depend on controlling a location. The traditional traffic-cop model of security—waiting for violations to cross your line of sight—simply doesn’t work.
‍
Security teams can’t realistically stand in the middle of every encrypted connection between cloud platforms, nor do they control the infrastructure where much of this activity occurs. But they can influence the access relationships that make those connections possible. OAuth grants, API permissions, service accounts, and corresponding delegated AI capabilities function like roads and on-ramps—durable decisions that determine which paths exist long before any traffic begins to flow.
‍
What’s emerging instead is a model centered on relationships—between users, applications, services, and data—rather than traffic flowing through a defined boundary. There is no “perimeter” enforced at the network edge. Data protection becomes less about directing traffic in real time and more about designing a better map: deciding which roads should exist, where they lead, and when they should be closed.
‍
Security teams preparing for this shift aren’t waiting for a perfect replacement for perimeter controls. Instead, they’re adjusting how and where they apply influence.
‍
First, they treat access decisions as security events. Granting an integration, authorizing an AI assistant, or creating a service account is no longer a low-risk administrative task—it’s a decision that can enable ongoing data movement long after the initial context is forgotten. The moment to intervene is the moment access is requested—when context is richest and course correction is easiest.
‍
Second, they focus on visibility across identities, not just applications. Understanding which users and non-human identities can access sensitive data often matters more than knowing where that data is stored.
‍
Finally, they accept that enablement is now part of security’s role. In SaaS-first environments that security teams don’t own or operate, sustainable risk reduction comes from guiding better decisions, not enforcing boundaries.
‍
Native controls within SaaS platforms remain inconsistent and incomplete. When SaaS and AI providers pushed decision-making to end users, employees became active participants in data protection. Employees are not just passive rule followers—they are decision makers whose choices create new data flows and new forms of risk.
‍
The perimeter isn’t gone—it’s just multiplied. And if we want to protect data in this new world, security leaders will need to stop guarding the gates, and start guiding the flow.
