COMMENTARY: A few months ago, I asked a security leader at a large organization a simple question: do you know where your most sensitive data actually lives today? The honest answer was no. There were policies in place dictating where the data was supposed to live. However, nobody could point to the real, current location.
This knowledge gap is close to universal, and it is the real story behind why artificial intelligence (AI) is suddenly making boards, risk committees, and compliance leaders nervous.
The existing data sprawl problem
Not knowing where data lives, or what is known as content sprawl, is not a new problem. All organizations face the slow, invisible accumulation of files, folders, and shared drives that nobody quite owns anymore. When employees leave, they often leave behind orphaned sites or libraries. Almost every project spins up a chat channel that outlives the project.
[
SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]
None of this happened because people were careless. Collaboration tools are built to make sharing frictionless, and most organizations have nothing in their operating rhythm that makes stopping sharing equally easy. Access accumulates and rarely gets cleaned up. Employees often move to the next project without thinking about the data they left behind.
For years, this data sprawl was tolerable, because obscurity functioned as a de facto control. A finance folder from years ago might be technically accessible to half a department, but finding it required knowing it existed and digging manually through hundreds of files. This obscurity was never real security. It simply raised the effort required to exploit oversharing high enough that almost nobody bothered.
AI is now finding old data
If you ask an AI assistant to summarize everything the organization knows about a given account, it will search everything it is permitted to see and return a clean, confident answer, including material only accessible because permissions were never cleaned up. The assistant is not hacking anything. It is using its access exactly as intended, at machine speed, applied to old or poorly understood permissions.
A new data sprawl source: AI itself
AI working as intended is creating a new source of data that needs to be secured. Security teams have spent two decades monitoring how data leaves through email, instant messages, and cloud uploads. Traditional data loss prevention programs are designed to protect those channels. AI assistants have opened another door most organizations do not yet realize is there.
Related reading:
A routine prompt can locate, aggregate, and reproduce sensitive data from a dozen sources in seconds. The output is pasted into an email or presentation, or shared with someone who never had access to the original files. Because no file was downloaded, traditional checkpoints stay silent even though the information moved.
The reverse happens too: employees regularly paste customer records, contract language, or source code directly into prompts, not to exfiltrate data but to get work done faster.
Getting ahead of the problem
Once aware of the AI risks, security teams’ first instinct is to restrict usage of AI tools. However, that tends to backfire. Employees want to work efficiently and try new tools, so organizations that restrict too much often push usage into unmanaged personal accounts and lose visibility entirely. Instead, organizations need to think of shifting their viewpoint on data to catch up with new AI tools.
First, treat all AI interactions as a monitored channel, on par with email or endpoints. This is not about surveilling individual prompts. Instead, it is about acknowledging that an organization cannot govern a data flow that it refuses to recognize. If a data loss prevention program covers email, endpoints, and cloud storage but has nothing for AI interactions, its view of ways data may leave is out of date.
Second, prioritize the data that matters most rather than trying to secure everything at once. Perfect visibility across every file overnight is not realistic, and attempting it often stalls the effort before it starts. A more workable approach identifies the small fraction of data, often 5% to 10%, that would cause the greatest damage if exposed and secures that data first.
Third, prepare for the future of AI. The scenarios above assume a human is in the loop, typing prompts and deciding what to do with the answers. Agentic AI introduces systems that access data and act on their own with no one watching each step. The governance work done now, while a human is still in the loop, is the foundation needed before autonomous agents operate at scale.
Organizations that start the AI visibility work now, while AI use is still growing and evolving, will be in a better position to control their data than those organizations that wait for a cybersecurity incident or news headlines to force the issue.
