Shadow IT and its Nemesis: DLP

What Shadow IT Looks Like Now

Shadow IT is any app, cloud account, or integration used without IT's approval. It's almost never malicious — someone needed to move faster than procurement allowed and found something that worked.

The definition hasn't changed. What it contains has. Most organizations are dealing with three versions at once, while their controls only cover the first:

Shadow SaaS — unsanctioned file sharing in Dropbox or personal Google Drive accounts, unvetted project tools.

• Data sits somewhere it shouldn't
• Bad, but slow: there's a file to find and a share link to revoke

Shadow AI — customer data pasted into ChatGPT, Claude, or Gemini.

• One paste moves thousands of records in seconds
• Desktop and CLI clients never touch your network monitoring
• Consumer tools have no BAA and may keep what you sent
• Nothing to find afterward, and nothing to revoke

Shadow agents — custom GPTs, OAuth-connected AI apps, self-wired MCP connectors, AI features quietly switched on inside SaaS you already pay for.

• Not a one-time paste — a standing identity reading your data on a schedule
• Authorized by someone who clicked "Allow" without reading the scopes

Same pattern throughout: data leaving through a door you can't see. It just gets faster and less reversible each time.

📖 Going deeper: the differences between these matter more than the labels. See Shadow AI vs. Shadow IT: What's the Difference?

Why It Matters

Nearly 7 in 10 organizations were compromised via shadow IT between 2021 and 2022, according to IBM Security's Randori report — and that was before generative AI removed what little friction adoption used to have.

The pattern in real incidents is consistent:

• Kaiser Permanente (2024) — over 13 million people affected through tracking code operating outside deliberate review
• Change Healthcare (2024) — 4TB of data stolen and a $22 million ransom, with unmanaged access at the root
• Sony Pictures (2014) — still instructive: attackers didn't defeat the controls on well-managed systems, they went to the data that had none

Compliance breaks the same way. HIPAA, PCI DSS, SOC 2, GDPR, and CCPA all assume you can account for regulated data — where it is, who touched it, how to delete it. An unknown app processing customer records makes that impossible, and "we didn't know the tool was in use" has never been an accepted answer.

Discovery Isn't the Fix

Plenty of tools will hand you a list of unsanctioned apps. Then what?

• The list doesn't stop anything. Most teams discover far more than they can act on, and not one finding prevents a leak on its own.
• Traditional discovery looks in the wrong place. Network and SaaS logs catch web apps people log into — and miss the desktop AI client, the CLI script, and the browser upload. That's exactly where the newest risk lives.
• Blocking everything backfires. Block the tool people need and they find another one. Your list gets shorter. Your actual exposure doesn't.

📖 Going deeper: for the five channels discovery actually needs to cover, see What is Shadow IT Discovery and Discover AI Agents.

✨ DLP Protects the Data, Not the App List

Here's the inversion that makes DLP shadow IT's nemesis: stop trying to enumerate every destination, and watch the data instead.

An unsanctioned tool being present stops being an emergency. What matters is whether regulated data can reach it.

That takes four things:

1. Detection that reads everything

• PII, PHI, and PCI across documents, spreadsheets, and images
• A screenshot of a card caught like any other record

2. A graduated response

• Redact or mask — keep the content usable in Slack, Gmail, and Zendesk
• Block — when the risk is genuinely too high
• Warn and coach — so people learn in the moment
• Revoke — kill public and external shares on Google Drive, SharePoint, and Box

One blunt action for everything is what drives usage underground.

3. Coverage where shadow IT lives

• Sanctioned SaaS is table stakes
• Browser and endpoint are where the unsanctioned upload and the AI paste actually happen
• Data lineage keeps a corporate file identified as corporate even after someone renames it

4. Controls for AI and agents

• Inspection before submission to ChatGPT, Claude, or Copilot handles people
• Agents need more: redacting what an MCP-connected agent returns in its tool responses, and monitoring what it accesses

None of this needs keystroke logging or screen recording. Watching destinations and data events answers the security question — and leaving employee surveillance out is usually what gets the rollout approved.

📖 Going deeper: Shadow AI Monitoring: See AI Usage Without Surveillance covers where that line sits and how to defend it internally.

✨ How Strac Handles It

Most tools do discovery or enforcement. Strac runs both on the same agent, so the thing that finds the problem is the thing that fixes it — no second product, no policy translated between two consoles.

What that looks like in practice

An employee exports a customer list from Salesforce, saves it locally, renames it notes.xlsx, and pastes the contents into a free AI summarizer.

• A CASB misses it — no sanctioned SaaS login happened
• A filename-based tool misses it — the file was renamed
• A network tool misses it — the paste went to a browser tab

Strac catches it three times over: data lineage knows the file came from Salesforce no matter what it's called, endpoint DLP sees the local copy, and browser DLP inspects the paste before it's submitted — redacting the PII, or blocking it, or warning the employee and logging the justification.

Where Strac is different

• Discovery and enforcement in one agent — the gap between "we found it" and "we stopped it" is where most programs actually leak
• Lineage, not just pattern matching — corporate files stay identified as corporate through renaming, editing, and copying
• Agent-layer control — redaction inside MCP tool responses, not just OAuth scope review. Very few tools reach this at all yet
• Images and unstructured content — a screenshot of a driver's license is detected like any other record
• No keystroke logs, no screen recording — the reason endpoint rollouts get approved instead of stalling in a works-council review

What you get out of it

• One classification engine, so a PCI rule means the same thing in Slack, on a laptop, and in an agent's tool call
• One audit trail covering HIPAA, PCI DSS, SOC 2, ISO 27001, GDPR, and CCPA
• API integrations plus a lightweight endpoint agent — deployed in days, not quarters

📖 Going deeper: Generative AI DLP in 2026 breaks down how enforcement works across browser, endpoint, SaaS, and MCP.

Where to Start

1. Discover continuously — across browser, endpoint, MCP connectors, OAuth grants, and SaaS logs
2. Rank by real exposure — the sensitive data actually flowing through each tool, not the finding count
3. Turn on enforcement first — protecting the data buys you time to make good decisions about the apps
4. Then handle each tool on its merits — sanction what's useful, replace what isn't, restrict the data rather than the tool where you can, revoke what shouldn't exist
5. Give people a fast way to ask — shadow IT is a symptom of a process people are routing around

📖 Going deeper: How to Discover and Manage Shadow IT walks through building the full program.

Conclusion

Shadow IT isn't going away, and a policy that pretends otherwise just moves it somewhere darker. The teams handling it well stopped trying to win the app-approval race and started protecting the data instead.

Discovery tells you what exists. DLP decides what leaves.

🌶️ Spicy FAQs on Shadow IT and DLP

What is shadow IT?

‍Any app, service, cloud account, or integration used without IT approval. Usually adopted for good reasons, risky anyway — because the data moving through it isn't monitored, logged, or contractually covered.

How does DLP help?

‍Discovery finds the tool; DLP stops the leak. It inspects data as it moves and redacts, blocks, warns, or revokes in real time — so an unsanctioned tool existing doesn't mean regulated data ends up inside it.

Isn't it easier to just block unapproved apps?

‍No. Blanket blocking pushes people toward tools you haven't heard of yet. Restricting the sensitive data instead keeps them productive and the exposure controlled.

Does shadow IT include AI tools?

‍Yes, and shadow AI is the highest-risk form. A prompt can move regulated data into a model with no BAA in seconds, leaving nothing behind to revoke.

What about AI agents and MCP connectors?

‍The newest form. An OAuth-connected agent holds standing access and returns real records in its tool responses, so control has to reach the response itself — not just the OAuth scope.

Can you do this without monitoring employees?

‍Yes. Tracking destinations and data-classification events answers the security question without keystroke logs or screen recording.