It started like any other working session.
Our consultant joined a Microsoft Teams call with a customer to walk through their security posture. A few minutes in, another “participant” quietly joined the meeting: an AI assistant. It began recording, transcribing, and emailing out notes.
The customer paused.
“Is that yours? Did you guys set this AI up?”
We hadn’t. In fact, we only use Copilot in our own environment—not third-party AI bots.
That simple question turned into an “aha!” moment for the customer and a textbook example of how shadow IT and unsanctioned AI tools can slip into even highly regulated environments—and how Microsoft Defender for Cloud Apps can give you the visibility (and control) you thought you already had.
The problem: Shadow AI hiding in plain sight
Behind the scenes, a user at the customer had previously clicked a vendor link that “helpfully” offered to activate an AI meeting assistant trial. It was marketed as a productivity booster: record your calls, produce summaries, capture action items. Sound familiar?
From that moment on, the assistant followed that user to every Teams meeting, recording and sending back detailed notes—whether anyone realized it or not.
For many organizations, that’s a minor convenience issue. For this customer, it was a big problem:
- They work with sensitive IP and confidential projects.
- They have an internal policy that explicitly disallows unsanctioned third-party AI tools.
- They had no idea this trial had been activated or what data it was collecting.
This is shadow IT in 2025: not just rogue SaaS apps, but AI agents quietly recording and exporting conversations.
The twist: The tooling was already there (but no one was using it)
Although the security team knew Defender solutions were deployed and providing protection, the detailed configuration and day-to-day usage of Microsoft Defender for Cloud Apps hadn’t fully carried over in the transition between teams. As a result, valuable insights—like AI assistant traffic and cloud app risk scoring—were being generated but not consistently reviewed, translated into policy, or used to drive action.
So the AI assistant only came to light because it literally showed up with a name tag in a meeting.
Once we suspected something was off, we opened the Defender for Cloud Apps portal and quickly:
- Located the specific AI app and its risk profile.
- Identified which user had activated it and when.
- Saw how much data had been uploaded and from where.
From there, the customer could:
- Contact the impacted user.
- Revoke access to the unsanctioned app.
- Update their policies to block that AI assistant going forward.
What changed their mind about fully licensing and adopting Defender for Cloud Apps? Not a feature list. A real-world incident that showed them exactly what they were the challenge was.
What is Microsoft Defender for Cloud Apps, really?
If you’ve heard of it only as “that CASB thing,” it’s worth reframing:
Microsoft Defender for Cloud Apps is a SaaS security platform that helps you:
- Discover and control shadow IT using Cloud Discovery and a catalog of 31,000+ cloud apps, each scored against ~90 risk factors.
- Govern app-to-app interactions, including how SaaS apps and AI tools connect to each other and to your data.
- Detect threats and anomalies across connected applications, like unusual data exfiltration.
And, increasingly important, it also helps secure AI applications and agents—including runtime protection for Copilot Studio AI agents that can block suspicious behavior in real time and alert security teams.
Used correctly, it becomes your control tower for SaaS and AI usage.
Why unsanctioned AI assistants are uniquely risky
AI meeting assistants (think Read AI, Fireflies, etc.) aren’t inherently evil. They deliver real value. But from an enterprise security perspective, they’re… tricky.
Common risks include:
- Uncontrolled data capture. Every call they join can include customer data, IP, strategy, or even regulated information.
- Opaque data handling. Where is the data stored? Who has access? How long is it retained? Is it used to train other models?
- App-to-app sprawl. These tools often connect to email, calendars, CRM systems, and more—multiplying your blast radius if they’re compromised.
For organizations with strict governance—like the customer in our story—an unvetted AI assistant is essentially an uninvited observer with a perfect memory.
The real lesson: Tech alone doesn’t fix shadow IT
The key takeaway from this incident wasn’t “you need another security product.”
It was this:
You may already have the tools—but if you’re not using them, you’re not protected.
The customer had:
- Defender for Cloud Apps deployed.
- Cloud Discovery running.
- Logs and insights flowing in for over a year.
What they didn’t have was:
- Clear ownership of the tool after staff turnover.
- Policies and alerts configured around AI and high-risk SaaS categories.
- Operational processes to regularly review and act on the data.
The aha moment wasn’t just spotting the AI bot; it was realizing how much more proactive they could have been if Defender for Cloud Apps had been properly operationalized.
5 practical steps to get ahead of shadow AI and unsanctioned apps
If this story feels uncomfortably familiar, here’s how to get ahead of it:
- Inventory what you already own.
Confirm whether Defender for Cloud Apps (or equivalent) is already licensed in your tenant and who “owns” it today.
- Turn on (and tune) Cloud Discovery.
- Ingest traffic logs from firewalls, proxies, or endpoint agents.
- Use built-in risk scores to identify AI tools, meeting assistants, and other high-risk categories.
- Define a clear AI governance stance.
- Which AI tools are sanctioned (e.g., Microsoft Copilot)?
- Which are explicitly banned?
- Communicate that policy to users in plain language—not just in a PDF.
- Create policies and alerts for AI & high-risk SaaS.
- Alert on new apps in “AI” or “productivity assistant” categories.
- Auto-tag or block apps that exceed your defined risk threshold.
- Operationalize review and response.
- Assign a security owner to review Cloud Discovery reports on a regular cadence.
- Document what happens when a new unsanctioned app is found (who contacts the user, who updates policies, etc.).
From incident to advantage
For our customer, the uninvited AI guest became a turning point.
- It validated the value of Defender for Cloud Apps in a real-world scenario.
- It sparked a renewed focus on governance over AI tools, not just generic SaaS.
- It became a story they could share internally to justify investment in licenses, policies, and security ownership.
If you’re somewhere between “Wild West” and “North Korea” on your governance spectrum (as we joked in the meeting), the goal isn’t to block everything or trust everything. It’s to give your admins line of sight into what’s happening—then empower them to act quickly when an unfamiliar “participant” pops into a call.
Because in 2025, if an AI is going to join your meetings, it should be by design—not by surprise.
