Industry Leadership
Strategic Initiatives
CSA's strategic programs driving innovation in AI, cloud, and Zero Trust.
A public-interest 501(c)(3) dedicated to secure and trustworthy AI.




Industry Leadership
Strategic Initiatives
CSA's strategic programs driving innovation in AI, cloud, and Zero Trust.
A public-interest 501(c)(3) dedicated to secure and trustworthy AI.

CSAI FoundationChaptersEventsBlog
Join this webinar on November 18 to explore how automated testing, red-teaming, and model scanning identify AI vulnerabilities early and strengthen secure development →

DLP and UEBA Are Asking the Wrong Insider Risk Questions

Published 11/18/2026

DLP and UEBA Are Asking the Wrong Insider Risk Questions
Originally published by Above Security.
Written by Aviv Nahum.

According to IBM's Cost of a Data Breach report, a breach that begins with a malicious insider costs $4.99M on average (the costliest initial vector measured). Mean time to identify and contain it: 260 days.

Ask a security leader whether insider risk matters and the answer comes back immediately. Ask them to describe their insider risk program and the conversation slows down.

That gap is worth sitting with, because it isn't a budget gap. Two mature categories- DLP and UEBA- twenty years of engineering, billions in spend, and insider risk is still handled after the fact. Something structural is going on.

Here is the clearest way to show it:

 

The risk is in the sequence, not individual actions

An account executive. Every action below is authorized, in policy, and typical for the role.

  • Day 1: Signs in to Salesforce, opens a strategic account plan. Authorized.
  • Day 3: Runs the account data through a sanctioned AI assistant. Approved tool.
  • Day 6: Connects a personal application to a work account. Permitted.
  • Day 9: Communicates with an external destination not seen before. Normal for sales.
  • Day 12: Working patterns shift the way they shift before someone resigns. Not an event at all.

No policy was violated, so DLP stayed silent. No single action cleared a tuned threshold, so UEBA stayed quiet. The risk was never in any one row. It was in the sequence, and the sequence is the thing neither category is built to hold.

 

DLP isn't broken. It's answering a different question that was never built for insider risk

The DLP model is coherent and, on its own terms, effective: classify the data, author the policy, monitor the channels, block or log. It remains the backbone of most compliance programs for good reason.

It was designed for a world where data lived in files, files lived on endpoints, and they left through a small number of legible exits like email, removable media, and print. Watching the exits was a fair approximation of watching the risk.

However, DLP acquired the insider risk mandate largely by accident. It was the only control positioned to observe what employees did with company data, so when insider risk arrived on the roadmap, it landed on the tool already standing there. Adjacency, not design. Much of the frustration security teams feel toward DLP traces back to that.

The limit isn't tuning. A rule is a conditional statement evaluated against an observable event, and several of the things that matter most in insider risk are not observable, not events, or not reducible to a condition:

  • Intent: A policy compares an action to a rule. It has no place to store why. Uploading a file to personal cloud storage is one event with two entirely different meanings, and the rule sees one.
  • Sanctioned-channel abuse: Most insider harm travels through approved tools. The CRM export, the Drive share, the direct message. Channel allowed, volume within norms, person authorized.
  • Behavior over time: Access broadens, patterns shift, then data moves. Each step clears policy in isolation. The arc is the signal, and the arc is not an event.
  • Actions that move no data: Access probing, privilege accumulation, credential sharing, a second full-time job on company time. Nothing leaves, so nothing triggers.
  • The classification dependency: DLP protects what has been labeled correctly. Anything unlabeled, newly created, screenshotted, retyped, or pasted into a browser sits outside the perimeter the policy defines.

 

DLP and shadow AI

Verizon's 2026 DBIR report examined more than 850,000 DLP events involving generative AI tools: 45% of employees are now regular AI users on corporate devices, up from 15% in a year, and 67% sign in with non-corporate accounts. Source code is the most submitted data type by a wide margin.

DLP saw all of it. But, what no DLP event can tell you is which paste was an engineer debugging a function under deadline and which was someone assembling a portable copy of the codebase. In the log, those two are the same record.

 

UEBA asked a better question, but deviation still isn't necessarily risk

UEBA started from the right instinct: stop enumerating forbidden actions and start describing normal people. That moved the industry from what is prohibited to what is unusual, which is a real step toward behavior. For compromised accounts behaving unlike their owners, it works well.

The trouble is that people depart from their own history constantly, and almost always for legitimate reasons.

What UEBA scores: identical telemetries and scores, different readings

table

Identical telemetry. Identical score. Two completely different futures. Nothing in the behavioral data separates them, because the separating fact isn't behavioral data. It's context: what this person's role changed to last month, whether their manager sanctioned the project, whether they resigned on Friday afternoon.

So a score is not a finding. It's a prompt for a human to go and produce a finding- check the role, the recent changes, whether the manager knows, the surrounding business context. Twenty minutes on a good day, two days on a bad one. Multiply by the queue, and the arithmetic nobody runs before purchase runs itself: deploy and baseline, alert volume overwhelms, thresholds up and exceptions in, quiet system.

Every tuning decision in that cycle is defensible. The aggregate is a system tuned toward silence, and a quiet insider risk tool is indistinguishable from a working one. Worse: because the readings in that right-hand column are rare, a well-tuned engine learns to treat them as exactly the kind of outlier it should suppress.

 

The premise both DLP and UEBA share

DLP and UEBA look like opposites; one defines the forbidden, the other defines the normal. Underneath, they share an assumption: that insider risk can be specified before it happens.

Write down what each actually answers and the gap becomes obvious.

  • DLP answers: did sensitive data violate a defined policy? A yes or no, about an object, against a rule someone wrote in advance.
  • UEBA answers: did a person behave anomalously? A distance from a baseline, about an individual, against a model built in advance.
  • Insider risk requires: what is actually happening, does it represent risk, why, what else is connected to it, and what should we do?

The third question is not a harder version of the first two. It's a different kind of question. The first two can be answered by a system that compares a thing to a stored definition. The third can only be answered by something that reasons about a situation.

Which is why functioning insider risk programs have historically correlated with headcount rather than tooling budget. The reasoning was always the deliverable, and until now the only thing capable of producing it was a human investigator who understood the organization.

Large banks, intelligence agencies and heavily regulated enterprises could staff that. Everyone else inherited a default: something happens, HR or Legal raises it weeks later, security is asked to look, a timeline gets reconstructed. That's forensics, not risk management. By the time Security is engaged, the decision that mattered was made weeks earlier, and documentation is not prevention.

 

Insider risk is an investigation problem

The missing capability isn't more detection, it's investigation. Who is this person right now, what is normal for them, what changed, why, and does it indicate risk.

The alternative to an alert is not a better alert. It's a finished investigation with a behavioral timeline, the contextual analysis around it, the reasoning behind the classification, and recommended actions, landing where Security, HR and Legal can work from one record. Nothing defined in advance, because nothing about a person's intent was ever available in advance.


About the Author

Aviv Nahum is a founder and technologist focused on applying AI to real-world cybersecurity problems. He has built and led core R&D teams at high-impact security startups, including a company that exited in a billion-dollar acquisition. His service in an elite Israeli Military Intelligence cyber unit gave him a highly specialized perspective on the intersection of cybersecurity and AI today - and where the field is heading. Previously CTO and co-founder of Ctrl, which was acquired by Sana AI (now part of Workday), Aviv is currently the CEO and co-founder of Above Security: a platform focused on AI-driven prevention and detection of insider threats.

Share this content on your favorite social network today!

Unlock Cloud Security Insights

Unlock Cloud Security Insights

Choose the CSA newsletters that match your interests:

Subscribe to our newsletter for the latest expert trends and updates