Shadow AI: A New Insider Risk for Cybersecurity Teams to Tackle Now

Disclaimer: Don't get me wrong, I love using generative AI daily for research and writing. This is about how other users could be using it when they don't know what they don't know and are accidental in their actions to hurt the organization where they work.

Shadow IT has always lived in the background of organizations' environments with unapproved apps, rogue cloud services, and forgotten BYOD systems. Like all technology, the Shadow IT ecology is evolving. It's evolved into something more challenging to detect and even more complex to control, and that's Shadow AI. As employees lean on AI to get their work done faster, they may introduce risk without realizing it.

From marketing teams using Claude to research and write content to developers pasting proprietary code into Gemini, the line between productivity and exposure is thin. These tools promise speed and convenience but can become a growing liability without governance.

Nearly 74% of ChatGPT usage in corporate environments happens through personal accounts. That means enterprise controls like data loss prevention (DLP), encryption, or logging are nowhere in sight. Combine that with the 38% of employees who admit to inputting sensitive work data into AI tools without permission, and you've got a significant insider threat. While accidental, it's no less dangerous than a user clicking on a link in a phishing email.

Source: Infosecurity Magazine

The Security Risks Lurking in Plain Sight
The risks tied to Shadow AI extend beyond accidental data leaks. Developers using AI coding assistants may inject insecure code into applications, especially when no review or validation process is in place as part of the software development life cycle. 

Customer support teams might lean on chatbots to handle inquiries, which introduces privacy risks when sensitive customer data flows through third-party tools. Even browser plugins with AI functionality can quietly siphon everything from form data to clipboard content to recordings of confidential meetings.

And then there's the network side. Employees who use AI-powered proxies or VPNs to get around access controls aren't just sidestepping policies. They are opening doors that attackers can exploit. AI-enhanced meeting tools like transcription services can store confidential conversations offsite, outside IT's control and purview.

We are no longer dealing with isolated risks as we increase our attack surface, all created by convenience and productivity.  

Strategies to Tackle Shadow AI
Now, we don't need to start blocking all aspects of generative AI platforms in our firewalls because that's like putting a finger into a crack in the dam to prevent water from spilling out. It's futile, and like how water will always find a way, so will your users. 

To get ahead of Shadow AI, we start with transparency. Organizations must create clear acceptable use policies (AUP) that address what is acceptable regarding AI usage. It is a considerable undertaking to communicate with the organization what AI practices are permissible. Teams must know which tools are approved, what kinds of data can be input, and where the line is drawn.

Education must go deeper than awareness and focus on managing the human risk element. People aren't misusing AI tools because they want to cause harm, they are looking to solve a problem. Education should focus on the consequences of these tools to build an understanding and not as a scare tactic. When users see how a prompt can lead to a data leak or a compliance breach, it allows them to see the dots connected between their action and the impact.

Visibility is also critical. Monitoring systems should be in place to detect when unapproved AI tools are used, whether through browser telemetry, endpoint detection, or network traffic analysis. Rather than blocking everything outright, IT and security teams should focus on understanding their users' needs. If that's access to a GenAI platform, consider a GenAI portal, where users can interact with various platforms through API but with a filter to ensure no sensitive organizational data is exfiltrated. 

Finally, reviewing AI tools before approving them for use must become a formal part of your software procurement process. When a user wants to utilize an AI tool or platform, having a process to address the need and business case ensures that people are not just jumping on the latest GenAI bandwagon but allows for your organization's legal, communications, IT, and cybersecurity teams to review and ensure that data is protected.

The process should include vetting how data is stored, processed, and shared and verifying whether the tool offers enterprise features like encryption, single sign-on (SSO), and audit logs. If a tool can't meet those standards, it shouldn't be in your infrastructure. 

Real-World Example: Samsung's Shadow AI Incident
A recent case that brought Shadow AI risks into sharp focus happened at Samsung in 2023. Several engineers reportedly used ChatGPT to help debug code and optimize workloads. But in doing so, they inadvertently submitted sensitive internal data, including proprietary source code, into ChatGPT. The incident prompted Samsung to get their legal team to contact OpenAI to request they remove the source code upload to prevent it being used in their training models. Within the organization, a ban on generative AI tools was implemented across the company.

This event wasn't an advanced persistent threat. There was no malware or phishing campaign. Just users trying to do their jobs better. And in the process, they exposed critical intellectual property.

Case Study: Building a Proactive AI Governance Program
A Fortune 500 financial firm saw early signs of Shadow AI use across marketing, legal, and IT teams. Employees were using GenAI tools to summarize documents, create internal reports, and generate content for social media. Leadership recognized the risk and launched a six-month initiative to bring it under control.

The firm started with a survey to understand which AI tools were used and why. They discovered over 20 unique AI tools used without approval. Most of these were routing data through unsecured APIs. Next, they created an AI AUP that clearly defined approved tools, banned use cases, and outlined employee responsibilities.

With the policy in place, governance was needed. They developed an allowlist of vetted AI tools for use by the users. They also deployed browser telemetry tools to flag unauthorized tools and added AI usage reviews to their internal audit checklist. Most importantly, to address the human risk element, they rolled out quarterly training sessions focused on GenAI and AI risk to keep their users updated on the latest AI trends, threats, and attack vectors.

By focusing on enablement instead of enforcement, they saw a 60% drop in Shadow AI usage within four months. More importantly, employee satisfaction remained high because the tools they needed were available, now with much needed guardrails.

A Cultural Shift Is Required
Shadow AI isn't a technology problem. It's a human one. And like any insider risk, it stems from people doing what they think is right without realizing the implications. Organizations that treat this as just another enforcement issue will fall behind. Those who work to cultivate a secure culture and ensure users feel empowered and supported will continue to be ahead of the curve.

Ask yourself: 

1. Can you see how AI is used within your organization? 

2. Do your employees understand where the boundaries are? 

3. Do your systems and policies reflect the reality of how work gets done?

If not, now's the time to fix that. Shadow AI is here. It's not going away. It's your move.