Contributed By Brian Muskoff VP of Product Perforce
Software development has always generated sensitive data in unexpected places. However, the rise of agentic AI has pushed that problem into new territory. AI agents are not just accelerating the software development lifecycle (SDLC), they are evolving the process by touching data at every stage in ways that teams may not fully see. The concern is that it is interacting with potentially sensitive data without a specific request being made. The scale and speed at which this can occur has the ability to go beyond the governance frameworks of many organizations.
The encouraging news is that this is a solvable problem. Teams that build strong data governance practices designed for machine speed and autonomous systems and not just human workflows, will comply more confidently and innovate faster. This in turn will support the ability for organizations to build AI they can trust.
Test data management was relatively well understood only a few years ago. Managing the data prior to a release was typically performed without concern. Unfortunately, sensitive data now appears everywhere across the SDLC, including development sandboxes, CI/CD pipelines, model training datasets, feature stores, regression testing environments, and AI agent memory stores. It is present in every environment and across every model, throughout the development lifecycle.
This matters because of the volume of code and test data are growing sharply. With the rise of agentic AI, fewer humans are involved in the coding and more are instructing AI agents to do it. That shift puts greater emphasis on the need for test data because as more code is being generated, more of it needs to be tested. Agentic AI is increasingly capable of autonomously driving this process. As a result, many organizations are reporting that AI adoption is accelerating faster than their data privacy strategies can keep up.
Non-Production Environments and Known Risks
There is a persistent blind spot in how organizations think about data security. Production and non-production environments are treated very differently, even when both hold large volumes of sensitive data. Production environments are equipped with SOC monitoring, strict access controls, and incident response protocols. Non-production environments on the other hand include development, test, staging, and user acceptance testing (UAT). These environments were simply not built to withstand the same level of threats as production data, which is precisely why allowing real customer data, financial records, or health information to flow into them is high risk.
The problem is compounded by the economics of convenience. DevOps culture has rightly encouraged environment proliferation, such as the spinning up of multiple clones, refreshing data on a regular basis, and accelerating delivery pipelines. More environments mean more copies of data. Because it is easier to grab a production dump than provision masked or synthetic alternatives, that is frequently what teams do. There is a DevOps paradox at work, where the easier it is to do something, the more one is inclined to do it. When shortcuts feel low-risk, they become the default. Conversely, when properly governed, masked data is just as frictionless to access, teams do the right thing. The answer is not to restrict the data, it is to make compliance the path of least resistance.
Data governance frameworks were built for human workflows and allow for manual reviews, approval committees, and periodic audits. That model was already straining before AI arrived. With autonomous agents now capable of making hundreds or thousands of data requests per hour, it is simply incompatible with the new reality.
Governance needs to function as a service, with automated controls that enforce policy at the point of data delivery, in real time. Data compliance is increasingly going to be executed at runtime and at the moment of execution. That puts much greater pressure on organizations to know exactly what kind of data they are working with, meaning that classification and data intelligence needs to be embedded into the pipeline, not bolted on as an afterthought.
Building Governance for the Agentic Era
DevOps best practices do not become obsolete in the age of agentic AI. In fact, they become more important. The 2026 State of DevOps Report reinforces that mature DevOps is foundational to AI success. The same holds for data governance. Here are the practices that matter most:
- Embed compliance controls into the data pipeline itself, not as a downstream review step. Compliance logic should execute automatically at the moment data is requested or delivered. This is what transforms governance from a bottleneck into a service.
- Replace production data copies with virtualized, masked, or synthetic alternatives. Virtualized environments allow teams to spin up production-like database copies in seconds without moving raw sensitive data. Synthetic data can be generated to exact specifications, including edge cases that do not exist in production at all.
- Use Model Context Protocol (MCP) to connect LLMs with test data environments via natural language. MCP allows AI agents and developers to interact with data infrastructure through conversational prompts, removing the need to log into multiple systems or wait on infrastructure teams. When requesting a properly governed data copy is as easy as cloning a production one, teams will do the right thing by default.
- Shift from audit-based governance to runtime enforcement. Classify, tag, and enforce data policy at the moment data is accessed or generated, giving teams real-time visibility into what kinds of data are in use and where.
Two scenarios illustrate how this plays out in modern engineering organizations. In the first case, a testing agent runs regression tests overnight and discovers it needs a fresh copy of a payments database, masked for PCI compliance. No human is available to approve the request. The agent calls a data API, receives a virtualized, masked copy within 90 seconds, completes its tests, and tears the environment down without a compliance ticket ever being raised.
In the second situation, a QA agent needs to test how a payments system handles 10,000 simultaneous expired credit cards during a leap year. That scenario does not exist in production data. The agent generates a synthetic dataset with exactly those characteristics, runs the tests, validates a fix, and closes the defect before the team’s morning standup. No real customer data was involved at any point.
What both scenarios share is a design philosophy of being compliant with production-quality data on demand. All this is done through an API or natural language interface, with policy enforcement built into the delivery mechanism, rather than applied as a gate afterward.
The SDLC has never been more productive, and it has never exposed more sensitive data to more systems, more agents, and more environments simultaneously. The window to get governance right is narrowing, as regulatory frameworks like the EU AI Act raise the bar for compliant AI development and data breaches in non-production environments continue to make headlines. The organizations that will navigate this well are not those that build manual compliance processes. They are the ones that redesign governance as infrastructure: automated, embedded, real-time, and built for a world where the primary consumers of data are autonomous systems operating at machine speed. That is not a future state, it is the architecture that engineering leaders need to be building toward
Brian Muskoff has built his career around leading global product teams to create solutions that help large enterprises safely and intelligently streamline the flow of value through the software delivery pipeline. At Delphix he leads the Data Platform business line that is focused on providing market-leading DevOps Test Data Management offerings.
Guest Column: The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of Autonomy Magazine.
Leave a Reply