We Spent Over a Decade Building EDR to Catch What Employees Did Not Know Was Running. We Are About to Need the Same Thing for AI Agents.

By Bhaskar Tallamraju, CTO & Co-founder, Privify Inc.

In the mid-2010s, most enterprises had no real idea what was running on their endpoints. Shadow IT was everywhere: unsanctioned apps, unmanaged devices and browser extensions nobody signed off on. The security industry's answer was EDR (Endpoint Detection and Response): instrument the endpoint, establish a behavioral baseline, flag deviations and give security teams a single pane of glass over what was actually happening on machines they thought they controlled.

It took years to get right. I spent a good chunk of my career on exactly that problem: building intrusion detection and prevention systems, running high throughput packet inspection at network scale and shipping browser based threat detection to more than ten million endpoints. The common thread across all of it was not the specific technology. It was a discipline: you cannot secure what you cannot see. Visibility had to come before control, every time, because you cannot write a policy for a thing you do not know exists.

We are now watching the exact same blind spot open up again, this time with AI agents.

The Pattern Is Repeating, Almost Exactly

Walk into most enterprises today and ask a straightforward question: how many AI agents are actually running across your environment right now? Not the ones procured through official channels — the ones an engineer wired into a CI pipeline, the ones running inside a browser extension, the ones an analyst plugged into a spreadsheet macro because it saved four hours a week.

Almost nobody can answer that question with confidence. The honest ones will tell you they know they cannot.

This is shadow IT's second act. Except this time the stakes are structurally higher, and the discovery problem is, if anything, harder. AI agents do not always look like traditional software. They can be a script, an API call, a browser plugin or a few lines embedded in someone's workflow, with no binary to fingerprint and no installer to flag.

Why "Visibility First" Still Applies, Maybe More Than Ever

Every mature security discipline eventually converges on the same sequence: discover, observe, then govern. Skip step one and step three becomes theater — policies written for an incomplete picture of reality, compliance frameworks bolted onto infrastructure nobody has actually mapped.

We are at the point in the AI adoption curve where most of the industry conversation has jumped straight to step three. Everyone is talking about AI policy, AI compliance and responsible AI frameworks, important conversations all of them, while skipping past the more basic and less glamorous question of what is actually out there.

That is the mistake shadow IT taught us not to make twice. A governance policy is only as good as the inventory it is built on.

What This Means in Practice

If you are responsible for security, risk or compliance in your organization, the useful first question is not "what is our AI policy?" It is "how would we even find out if that policy were being violated right now?" For most organizations, the honest answer is that they would not. That gap, between the policies being written and the actual runtime visibility needed to enforce them, is where the real exposure sits.

Twenty years of watching this exact cycle play out with endpoints and networks convinced me it is worth solving early rather than retrofitting later. It is a large part of why I have spent the last stretch of my career building tools to close that specific gap for AI: discovery and observability, before governance, in that order.

The technology changes every cycle. The discipline does not.

---

Bhaskar Tallamraju is the CTO and co-founder of Privify Inc., an AI governance and compliance company. He has spent over 26 years building security products at Cisco, IBM, McAfee and Symantec/Broadcom.