Is data leak prevention built for AI?

Sponsored by

AI agents are starting to do more than access systems. They can invoke tools, execute workflows, and take action across business applications. As that autonomy grows, IT needs visibility into what each agent is allowed to do, what it actually did, who owns it, and how to revoke access when needed. Learn how JumpCloud helps govern AI agents alongside human and device identities.

Data leak prevention (DLP) products—tools that monitor and send alerts when sensitive data leaves a company—had it so much easier in their early-2000s heyday, when employers could box employees into one office and sensitive data could only exit via a few chosen avenues, like email and USB drives.

Now with AI, personal data can get absorbed into a prompt or one of many available cloud services, which leaves IT pros debating whether DLP is built for these automated times.

“There are many different doors now,” Hillary Baron, AVP of research at the not-for-profit Cloud Security Alliance, told us. “A single employee can really touch a dozen different apps in a day.”

And so much for that “P.” In a survey of more than 433 global IT pros by the Cloud Security Alliance (CSA), released on October 13 and done in collaboration with AI-native DLP company Jazz, only 16% of reported detections were “blocked by DLP before data left.” (Just over a third of incidents were caught after data departed, according to the report.)

“The vast majority of the time, it’s being used as a detection and post-event forensics tool rather than a preventative measure,” the report read.

DLP products monitor network activity (email and messaging, for example), endpoints (like computers, servers, and mobile devices), and cloud environments and repos. After high-priority data is classified, the DLP tools send alerts for unauthorized data transfers from the network—whether it’s a malicious insider or even just a well-meaning but policy-ignorant CEO trying to send something to their personal email.

The CSA study found that “generative AI tools” led data-exfiltration risks for respondents, ahead of cloud storage uploads, outbound corporate email, and personal or unmanaged devices. Despite citing GenAI as a top concern, the report also found that only 29% of polled IT pros could routinely produce a 30-day report showing how many times employees sent sensitive data to AI tools.

A separate October report of 5,000 global adults from the National Cybersecurity Alliance revealed that just under a third (32%) of workplace AI users have shared sensitive information with an AI tool without their employer knowing, and over four in 10 (43%) reported that their employer hasn’t clearly communicated what’s approved to share.

What about the agents? Andrew Bales, director analyst at market-intelligence firm Gartner, sees traditional DLP outmatched by AI-driven workflows, where non-human agents can connect to a data source via common interfaces like the Model Context Protocol. AI services, integrated into workplace platforms, can gain access to sensitive details.

“Historically, it’d be users accessing information. They download that information. They share that information they just downloaded to an external recipient. That’s no longer the case,” Bales said.

Gartner, in a September report, still projects the global DLP market to grow at a 20% compound annual growth rate. DLP vendors, the firm noted in the same study, are leveraging behavioral analytics and both small and large language models to identify anomalous data movements.

If DLP isn’t ready for the AI era, what’s the solution? Bales recommends additional measures alongside the products, including mechanisms closer to the data itself, like restricting people and AI access to specific environments. An IT pro, for example, may have to carefully configure the source platform itself—say, SharePoint, for example.

Other important defenses include real-time data protections like encryption, AI usage controls, and monitoring at the browser and endpoint level to identify sensitive data heading into AI systems.

Neither Bales or Baron are shutting the door entirely on DLP. As long as humans share data that they shouldn’t be sharing, the tech will have a place in the enterprise, Bales said.

“I think it’s more relevant than ever. It’s just that data has far more ways to leave than it did in 2009. Not fewer,” Baron told IT Brew.