In cybersecurity, the most persuasive arguments are often about cost.

In 2025, the global average data-breach cost to companies was estimated at $4.44 million per incident, according to research by IBM and Ponemon Institute. In particular, healthcare breaches averaged $7.42 million per incident and took the longest to identify and contain. These costs are occurring at a time when artificial intelligence (AI) systems are expanding the number of human and machine actors interacting with sensitive data.

Costs matter, but so do frameworks. One of today’s dominant cybersecurity frameworks is Zero Trust Architecture (ZTA). Moving beyond the older "castle-and-moat" assumption that actors inside an organization's network perimeter are inherently trustworthy, ZTA is based on continuous verification of users, devices, applications and workloads.

While ZTA can require significant upfront investment, the IBM-Ponemon research has shown these kinds of stronger identity-centric security controls and faster breach containment can help lower breach costs.

And so zero trust has become a foundational architectural response to cloud and identity-centric security challenges. But what happens when the tools an organization uses to protect identity, devices, networks and applications fail and data moves beyond its intended boundaries?

This is an increasingly plausible scenario in the AI era and it points to the need for data-centric controls that protect sensitive information even after it leaves the environments where access was granted.

Growing cybersecurity risk factors

Even as ZTA has matured, the domains it was designed to protect have become more complex.

First, the number of entities touching sensitive information throughout an organization now goes far beyond humans to include services and autonomous or semi-autonomous software agents. The IBM-Ponemon research notes that AI oversight is a growing risk factor for companies, governments and other organizations, and governance gaps around AI can raise their exposure to cyberattacks.

A widely used taxonomy for applications using large-language models (LLMs – the neural networks underlying generative AI) highlights recurring risks such as prompt injection and accidental disclosure of sensitive information. These risks can turn legitimate access into unintended leakage at machine speed.

In practice, this means AI-enabled workflows could expose an organization’s internal data or spread its sensitive information across systems in seconds.

A second reason for growing complexity in this area is the expansion of what counts as sensitive data to include streams of rich media. As computer vision advances, techniques such as pixel obfuscation may provide less protection than once assumed.

Recent research shows progress in restoring faces from blurred imagery and recovering pixelized facial information. The upshot is weakened privacy protections for people whose identities were intentionally obscured, whether they are victims or witnesses of crimes, participants in political demonstrations, or clients or patients in sensitive medical settings.

In response to these developments, privacy-enhancing technologies have gained renewed attention. Broadly speaking, these tools aim to reduce exposure of sensitive data during analysis, sharing or collaboration. For example, they may allow organizations to analyze customer, health or financial data without fully exposing the underlying information to every participant.

But many focus on protecting computation or governed data exchange, rather than also protecting data that moves beyond the environments where access has been approved.

A different kind of cybersecurity protection

The continued, potential exposure of sensitive data shows the need for a complementary layer of defense: data-centric access control through encryption. Most zero-trust controls operate at system boundaries. Data-centric encryption shifts enforcement to the data itself. The idea is to embed access policy into data that is encrypted so that the right to decrypt travels with it.

One standards-recognized, privacy-enhancing mechanism for embedding access control policies into encryption is called attribute-based encryption (ABE). Instead of focusing on computation or data exchange, ABE addresses the problem of enforcing access rules for data that may be copied, stored, re-shared or leaked. Under ABE, for instance, a leaked sensitive document, medical record or surveillance video would remain accessible only to authorized users.

Academic and industry research over the past two decades has advanced data-centric encryption approaches, although they have received less attention in mainstream cybersecurity discussions than many other privacy and security technologies.

As with other technologies, effective deployment of data-layer encryption depends on sound design and administration. A 2023 NIST analysis of ABE-based access control highlighted issues such as key management, authority design and policy misconfiguration. In practice, advanced cybersecurity becomes less a single-product decision than an ongoing operating model.

A portfolio approach to data privacy tools

Zero trust can no longer be framed as a network and identity story only.

In an AI-shaped economy, cybersecurity leaders need to take a portfolio mindset – one that accounts for non-human identities, autonomous workflows and increasingly sensitive data. Some privacy-enhancing technologies can support secure collaboration and analytics, while data-centric cryptographic enforcement can help contain exposure when data moves beyond its original controls.

There is no single best privacy technology. Instead, there are layers that can work together to minimize breaches and costs. In an AI-driven environment, where data is moving faster and farther than traditional controls were designed to manage, enforcing security at the level of the data itself deserves serious consideration.