The definition of shadow AI is any AI tool, model, or service your employees use without IT or security teams knowing about or governing it.

Shadow AI vs. shadow IT

Shadow AI might like shadow IT with a new coat of paint, but it isn’t:

• Shadow IT involves data storage. If a confidential file has been uploaded to an unapproved cloud drive, and you can trace and delete it to contain the risk.
• Shadow AI involves data processing. Paste sensitive data into a public tool, and you can’t control where that information goes. You can’t delete a prompt from a third-party tool’s memory, making it a permanent security vulnerability.

Even Samsung couldn’t make AI unlearn their information. Realizing how dangerous shadow AI security risks could be, they banned ChatGPT outright and started building their own in-house version.

The risks of shadow AI go beyond simple policy violations. Compliance exposure under the General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA) is a major concern. Confidently wrong AI-generated output and runaway costs are others. But for a lean team, the following consequences hit security and compliance hardest.

Human error: Data exfiltration through everyday AI interactions

The chain behind most generative AI data leakage is short. An employee trying to be more productive copies a confidential file or sensitive customer data. Then, they paste it into an unapproved AI tool. In one click, your data goes into a public model to be learned (ingested) and potentially shared with other users. You can’t recall that information once the tool processes it.

In 2024, CybSafe and the National Cybersecurity Alliance found that 43% of employees admitted to sharing sensitive work information with AI platforms without permission. And a Harmonic Security study in 2025 found that 16.9% of sensitive data exposures happened on personal free-tier accounts IT teams can’t see. IBM puts the added cost of a breach involving shadow data at $670,000 over other incidents. When your team uses these tools in the shadows, customer personally identifiable information (PII) and intellectual property (IP) are often the first things exposed.

Technical identity risks: Rogue OAuth tokens & AI integrations

When an employee connects an AI-powered tool to your company systems using a one-click login (OAuth), they’re handing that tool a digital master key. This process frequently grants the AI more access to your data than any security team would approve.

Multiply that across a team, and you’ve got rogue OAuth apps scattered through your workspace.The danger is that this access can outlive your normal controls. OAuth tokens can remain valid even after password resets and multi-factor authentication (MFA) changes unless they’re explicitly revoked or expired, so an attacker holding a stolen token may keep access long after you think the account is “locked down.

In 2025, threat actor UNC6395 used compromised OAuth tokens from the Salesloft Drift third-party application to access and export data from numerous Salesforce customer instances. Because the access came through a trusted SaaS connection, the traffic looked like normal business activity—an example of how stolen OAuth grants can let attackers “walk through the front door” without noisy login failures.

So how do you catch shadow AI as a lean IT team? First, know that shadow AI doesn't look like an attack. It uses the same encrypted web traffic (HTTPS) as your legitimate apps, so standard security alarms won’t go off. To catch it, you must look at these three areas.

1. Outbound network traffic to AI domains

   Your network logs show you when someone connects to AI domains like api.openai.com or claude.ai. But they can’t tell you who’s doing it or what data they’re feeding the model. That’s where endpoint detection and response (EDR) comes in. It lives on the device itself and can tell you exactly which application or browser extension reached out.

Don’t aim for a blanket block of AI domains. Many modern SaaS tools now bake AI features into their core functions, and blocking these domains can break those functions. Instead, prioritize visibility. Use EDR tools to build a map of which machines are talking to these AI services. Once you have that list, you can identify which users are breaking policy and address the behavior directly.

1. OAuth app authorizations

   Audit your OAuth app authorizations regularly for two things: apps connected without IT sign-off, and apps with permissions well beyond their job. The token an employee granted an AI tool months ago is often still live and unreviewed. Revoke access to over-permissioned apps.

While periodic reviews help, they only show a snapshot in time. A rogue app could be authorized minutes after you finish your audit. This is where managed identity threat detection and response (ITDR) helps. ITDR provides continuous monitoring of identity activity and OAuth apps, watching for suspicious behavior like an authorized app accessing sensitive data from unusual locations or at unusual hours. ITDR can then end compromised sessions before an attacker can move through your network.

1. Behavioral signals from security awareness training data

   Your phishing simulation results and training data already point to your highest-risk employees. These are the ones who repeatedly click the wrong thing, or who’ve said they’re unsure what AI usage is allowed. They’re the same people most likely to paste something sensitive into a tool they shouldn’t.

Huntress Managed Security Awareness Training identifies higher-risk learners by looking at real-world compromise and phishing-simulation data, then automatically routes Phishing Defense Coaching and targeted training their way.

Most lean teams can’t afford to build their own in-house generative AI platform, and simple bans usually backfire. Studies show employees continue using AI tools even without training or formal approval, often on personal accounts and devices that sit completely outside IT’s visibility.

Here are a few guardrails to get you most of the way without a full enterprise AI governance program:

• Educate your team on shadow AI risks: Close the knowledge gap with a managed SAT awareness program. Most employees don’t realize AI tools use their data for model training. Share concrete, practical guidance to change their behavior.
• Provide safe AI alternatives before enforcing restrictions: Avoid taking away a tool employees rely on without offering a vetted alternative. They’ll simply find another unsanctioned tool you won’t be able to track.
• Define clear data classification policies: Draft an AI acceptable use policy that explicitly details what data can and can’t go into external tools. Ensure your employees understand the consequences of mishandling company information.
• Automate oversight of unauthorized AI integrations: Stop relying on users to self-report their software choices. Use Managed ITDR to catch rogue OAuth apps on the identity side. And use Managed EDR to show which apps reach out to AI services from the endpoint.

The days of monitoring shadow AI through standard IT policies are over. You can’t secure what you can’t see. When data leaks through everyday prompts or rogue OAuth token grants, you lose control of your company’s sensitive IP.

What works is using tools to see exactly who’s moving through your environment and stop them the moment their behavior becomes a risk. By combining continuous identity monitoring with security awareness training, you can transform your team from a liability into a strong line of defense.

If you want to see what controlling shadow AI looks like in practice, Huntress can show you.