Shadow AI risks refers to the use of unauthorized AI tools by employees without IT knowledge or approval, and it is already one of the fastest-growing security threats in the workplace. According to a PagerDuty workplace survey, two-thirds of office professionals have used unauthorized AI tools at work, and 88% of them have shared work-related information with public AI tools. IBM’s research found that a high level of shadow AI adds an extra $670,000 to the average data breach cost. This is not a fringe problem. It is happening right now, across every department, in organizations of every size.

That number floored me when I first read it. Six hundred and seventy thousand dollars tacked onto a breach, not because someone got phished, but because an employee opened a browser tab and pasted client data into a free chatbot. Painful. And completely preventable. This guide cuts through the noise on shadow AI risks, what causes them, how to spot them, and what to do before they cost you.

What Is Shadow AI?

Shadow AI is the use of AI tools, applications, or services inside an organization without the knowledge, approval, or oversight of IT or security teams. This covers everything from an employee pasting a legal contract into ChatGPT to a developer using an AI coding assistant that IT never vetted. If the tool was not formally approved, it is shadow AI.

The term extends the older concept of shadow IT, but with a sharper edge. Shadow IT typically meant an employee downloading Dropbox or signing up for a project management app. Shadow AI involves large language models (LLMs) that can ingest, process, and potentially store sensitive data at scale. The risk profile is fundamentally different.

Shadow AI tools span a wide range. Free-tier generative AI chatbots accessed through personal accounts. Browser extensions powered by AI. Third-party plugins connected to internal systems. AI-assisted writing, coding, and analytics platforms that employees find on their own and start using immediately.

The common thread: no security review, no data processing agreement, no visibility for your IT team. The data goes in. You have no idea where it goes after that.

Shadow AI vs. Shadow IT: Key Differences

Shadow AI and shadow IT both describe unauthorized tool use, but shadow AI carries significantly greater data exposure risk because AI tools actively process, learn from, and sometimes retain the content users feed them.

Traditional shadow IT created blind spots. An employee using an unapproved cloud storage service was a governance problem. Shadow AI is a data sovereignty problem. When someone feeds a large language model proprietary code, financial projections, or patient records, that content can become training data for the model, get stored on third-party servers in unknown jurisdictions, or be accessed by the vendor’s staff under their own terms.

Shadow IT grew up alongside SaaS. Shadow AI has grown up alongside generative AI tools that are free, powerful, and instantly available without any procurement process. That accessibility is exactly what makes it harder to govern.

What Causes Shadow AI? Why Employees Adopt Unauthorized AI Tools

Employees adopt unauthorized AI tools primarily because those tools make their jobs faster and easier, and most organizations have given them no approved alternative to use instead. Speed beats governance when no policy exists to say otherwise.

According to Microsoft’s Work Trend Index, 78% of AI users bring their own tools to work. That is not a small rogue minority. That is the majority of your AI-using workforce operating outside your control.

Several factors drive this behavior. First, the tools are free and require no approval process. An employee can open ChatGPT in a browser, paste in a document, and get a summary in thirty seconds. No IT ticket. No waiting. Second, many employees simply do not know the risk. They see AI as a productivity tool, not a data handling channel. Third, when organizations do restrict AI access, they often do so without providing a sanctioned alternative, which pushes employees toward personal accounts and free-tier tools.

The pressure employees feel is real. 77% of office professionals believe AI restrictions are limiting their career mobility.People who feel left behind by AI adoption will find ways around restrictions. That is human nature, not malice.

Organizations that ban AI without building a sanctioned path forward are not solving the shadow AI problem. They are just making it invisible.

Shadow AI Examples by Department

Shadow AI shows up differently depending on the department, but the underlying pattern is the same: an employee finds an AI tool that solves a real problem, uses it without approval, and feeds it sensitive data without realizing the exposure.

Legal and Finance Teams

Lawyers and finance professionals routinely handle the most sensitive data in any organization. Contracts, M&A documents, financial models, audit files. When someone on the legal team pastes a draft contract into a generative AI tool to clean up the language, that contract, including party names, deal terms, and confidential obligations, goes to a third-party server. If the tool’s terms of service permit using inputs to improve the model, that data may never come back.

Finance analysts uploading revenue forecasts or customer pricing data into AI-driven analytics tools create the same exposure. The data leaves the building, often without any data processing agreement in place.

Developers and Engineering Teams

Developers are among the heaviest shadow AI users. AI coding assistants are extraordinarily useful. They are also a direct pipeline for source code, API keys, database schemas, and authentication credentials to leave your environment. A developer pasting a code block into an LLM to debug it may not notice that the block contains hardcoded credentials or proprietary logic.

This is not theoretical. Credential leakage through AI coding assistants is an active and documented attack surface.

Marketing and Sales

Marketing teams use generative AI tools to draft content, analyze campaign data, and summarize customer research. Sales teams use them to write proposals and analyze prospect information. Both functions regularly handle customer data, competitive intelligence, and personally identifiable information that falls squarely within GDPR and HIPAA jurisdiction. Most employees using free-tier ChatGPT for these tasks have no idea their inputs may be stored or used for model training.

HR and Recruitment

HR professionals using AI to screen CVs, draft job descriptions, or summarize performance reviews are processing personal employee data. Under GDPR, that data requires explicit legal grounds for processing. Feeding it into an unauthorized AI tool with unknown data retention policies creates a compliance violation before anyone has even noticed shadow AI is in play.

The Biggest Shadow AI Risks Organizations Face

Shadow AI creates a set of security and operational risks that go well beyond the usual concerns about unsanctioned software, and the data on those risks should get any CISO’s attention.

Data Leakage and Sensitive Data Exposure

Data leakage is the core risk of shadow AI. Cyberhaven’s research found that 66% of AI tools are high or critical risk, with 39.5% inadvertently exposing user interaction and training data. Nearly four in ten tools your employees might be using right now are quietly exposing what they type into them.

The exposure mechanism matters. When an employee uses a personal account on a free-tier AI platform, the service’s terms often permit using that input for model improvement. Sensitive data including client names, financial figures, legal arguments, and medical information can become part of a model’s training corpus. Once it is in there, you cannot get it out.

Netskope’s 2026 cloud and threat report found that 47% of people using generative AI platforms do so through personal accounts, where enterprise data controls have zero reach. Your DLP tools, your endpoint controls, your access policies: none of them touch a personal browser tab.

Expanded Attack Surface

Every unauthorized AI tool an employee connects to internal systems or feeds with business data is a new entry point for attackers. Shadow AI tools are not vetted for security. They may have poor authentication, weak encryption, or supply chain vulnerabilities in their own underlying models.

Prompt injection attacks, where malicious instructions hidden in content manipulate an LLM into taking harmful actions, are a particular concern for agentic AI tools that can take actions on behalf of users. If your employee is using an unapproved AI agent that has access to their email or file system, a prompt injection attack can turn that tool into an insider threat.

Lack of Visibility and IT Oversight

You cannot protect what you cannot see. Only 17% of companies have technical controls to prevent data uploads to public AI tools. That means 83% of organizations have employees freely uploading sensitive data to AI platforms and have no technical mechanism to stop it.

Shadow AI expands the visibility gap that shadow IT created. But where shadow IT tools were often discoverable through network traffic analysis, many shadow AI tools operate through encrypted browser sessions, personal accounts, and browser extensions that are nearly invisible to standard monitoring.

Poisoned Outputs and Unreliable AI

When employees use unauthorized AI tools, there is no guarantee those tools are producing accurate, unbiased, or safe outputs. Unsanctioned AI models may be trained on low-quality data, may have been tampered with through supply chain attacks, or may produce outputs that embed subtle biases or errors into business decisions. Employees acting on those outputs without knowing the model’s provenance are making decisions on a foundation nobody audited.

Regulatory and Compliance Risks of Shadow AI

Shadow AI creates direct regulatory exposure under multiple frameworks, and those frameworks are now fully in force. Organizations that have not built AI governance policies face real compliance liability, not just theoretical risk.

The EU AI Act came into full enforcement in 2025. It imposes specific obligations on organizations deploying AI systems, including requirements for transparency, human oversight, and risk assessment. Using unauthorized AI tools in regulated contexts, without documentation, risk assessment, or oversight, is a straightforward violation of those obligations.

GDPR is the more immediate risk for most organizations. Any unauthorized AI tool processing personal data of EU residents triggers GDPR obligations: a lawful basis for processing, a data processing agreement with the vendor, data subject rights compliance, and breach notification procedures. Shadow AI tools, by definition, have none of that documentation in place.

HIPAA creates an equivalent exposure in healthcare. An employee feeding patient information into an unauthorized AI tool may have just committed a HIPAA violation regardless of intent. There is no “I didn’t know it was unauthorized” defense under federal healthcare privacy law.

PCI DSS adds another layer for any organization handling payment card data. AI tools processing cardholder data must be part of your PCI scope, which requires formal vendor assessment and contractual controls. Shadow AI tools operating outside your knowledge cannot be in your PCI scope, which means they are out of compliance by definition.

And yet, 63% of organizations currently have no AI governance policies in place. That is not a gap. That is an open door.

How to Detect Shadow AI in Your Organization

Detecting shadow AI requires a combination of network visibility, endpoint monitoring, and employee engagement, because shadow AI hides in places that traditional security tools were not built to find.

Verizon’s DBIR research found that employee use of unapproved shadow AI tripled to 45% of the workforce. If nearly half your employees are potentially using unauthorized AI tools, a passive approach to detection is not a strategy. It is negligence.

Network and SaaS Discovery

Start with your network. DNS logs, proxy logs, and firewall traffic will reveal connections to known AI platforms. Tools like Netskope or Zscaler function as Cloud Access Security Brokers (CASBs) and can identify AI tool usage across your environment, including which tools are being used, by whom, and what categories of data are being uploaded.

Run a SaaS discovery scan. Most organizations are surprised by how many AI applications appear. Your approved list likely includes two or three tools. Your discovery scan will likely surface dozens.

Endpoint and Browser Monitoring

Browser extensions are a major shadow AI vector. An employee can install an AI writing assistant as a browser plugin and give it access to every page they visit, every document they open, and every email they write. Endpoint Detection and Response (EDR) tools and browser management platforms can audit installed extensions and flag unknown or high-risk ones.

Review personal account usage. Menlo Security’s research found that 68% of employees use free-tier AI tools like ChatGPT through personal accounts. Corporate data controls do not follow employees into personal accounts. If you see authenticated sessions to AI platforms using non-corporate email addresses, you have found your shadow AI.

Employee Surveys and Open Conversation

Do not underestimate the value of just asking. Anonymous employee surveys about AI tool usage consistently reveal far more shadow AI than technical monitoring alone. Employees who do not know they are doing something wrong will tell you what they are using if you ask in a non-threatening way.

That conversation also gives you the information you need to build a better approved alternatives list.

How to Mitigate Shadow AI Risks: 5 Actionable Steps

Mitigating shadow AI is not about locking everything down. That approach fails, as the data on restriction-driven workarounds makes clear. It is about building a framework that gives employees what they need while keeping your data where it belongs.

Organizations with expanded AI access had a 43% breach rate, compared to 11% for those without AI change, which tells you that ungoverned AI access is a material security risk. The goal is governed access, not zero access.

1. Build an AI governance policy now. It does not need to be perfect. It needs to exist. Define what AI tools are approved, what data categories can be processed by AI, and what approval process applies to new tools. A one-page policy is better than none. Publish it, communicate it, and revisit it quarterly because the AI tool market moves fast.
2. Create an approved AI tool list with enterprise agreements. Identify the AI tools your employees actually need for their roles. Negotiate enterprise agreements that include data processing terms, no-training clauses on your inputs, and geographic data residency controls. Give employees a sanctioned path. They will take it if it works as well as the free alternative.Provide a sanctioned AI toolkit with DPAs, no-training clauses, and data residency controls.
3. Deploy technical controls for data upload prevention. Only 17% of organizations currently have these controls in place, which means 83% are relying on policy alone. DLP tools configured to detect and block uploads of sensitive data categories to unapproved AI endpoints close the gap between what your policy says and what actually happens on a Tuesday afternoon when someone is under deadline pressure.
4. Train your people on shadow AI risks. Not a compliance checkbox training. A real conversation about what happens when customer data goes into a free chatbot, where that data goes, and why it matters. Employees who understand the risk make better decisions. Employees who only know there is a policy they are supposed to follow will work around it the moment the policy inconveniences them.
5. Monitor continuously and iterate. 90% of organizations now actively block at least one generative AI application, which shows the field is moving toward active governance. Set up regular SaaS discovery scans, review AI tool usage reports monthly, and build a cross-functional AI governance committee that includes IT, legal, compliance, and business unit leaders. Shadow AI is not a one-time problem you solve. It requires continuous attention as new tools emerge.

The right framework for AI governance does not start with “what do we ban?” It starts with “what do our people need, and how do we give it to them safely?” That shift in framing changes everything about how governance gets received by employees.

For a broader look at how these principles connect to your overall security posture, the RiskAware guide to cybersecurity frameworks covers how to build layered governance that actually holds up under real-world pressure.

Frequently Asked Questions About Shadow AI

Is shadow AI illegal?

Shadow AI itself is not illegal, but using unauthorized AI tools to process regulated data often is. Feeding personal data into an unapproved AI tool can violate GDPR. Processing patient health information in an unauthorized system may violate HIPAA. Handling payment card data outside your PCI-scoped environment breaches PCI DSS. The AI tool being “free” or “just a productivity shortcut” is not a defense under any of these frameworks.

How is shadow AI different from shadow IT?

Shadow IT covers any unauthorized technology, apps, or services used without IT approval. Shadow AI is a subset of shadow IT, but it carries a distinct risk profile because AI tools actively process and may retain the content fed into them. A shadow IT file-sharing app stores your data somewhere unauthorized. A shadow AI large language model may train on it, making exposure permanent and irreversible.

What data is most at risk from shadow AI?

The highest-risk data categories are those with regulatory protection: personally identifiable information under GDPR, protected health information under HIPAA, and payment card data under PCI DSS. Beyond regulated data, intellectual property including source code, financial models, legal strategy, and product roadmaps is at serious risk. These are exactly the categories employees are most likely to feed into AI tools when trying to solve a complex work problem quickly.

Can I block shadow AI entirely?

Technically, you can block known AI platforms at the network level. Practically, complete elimination is neither achievable nor advisable. Employees using personal devices or mobile connections bypass network controls entirely. And a blanket ban without a sanctioned alternative creates the exact conditions that drive shadow AI underground. Focus on detection, sanctioned alternatives, and governance rather than trying to build a wall that employees will find ways around.

What should my first step be?

Audit what is already in use. Run a SaaS discovery scan, review DNS logs for connections to known AI platforms, and survey employees anonymously about what tools they are using. You cannot govern what you do not know exists. Most organizations are surprised by what they find. Use that discovery as the input to your AI governance policy and approved tool list.

Where to Go from Here

Shadow AI is not going away. If anything, the pressure employees feel to use AI will increase, not decrease. The organizations that get ahead of this problem are the ones that treat it as a governance challenge, not a ban-and-block challenge.

The numbers tell a clear story. Two-thirds of your employees have probably already used an unauthorized AI tool. Three-quarters of those who use unapproved tools have shared potentially sensitive information through them. And a high level of shadow AI exposure adds hundreds of thousands of dollars to your breach cost if something goes wrong.

Do this before anything else: find out what AI tools are already operating in your environment. Then build the governance framework to make sanctioned AI use the easier path. Train your people. Deploy technical controls. And revisit your policy regularly, because this market changes faster than any annual review cycle can keep up with.

Your employees are not the enemy here. They are trying to do their jobs well. Give them the tools to do that safely, and shadow AI stops being a threat and starts being a managed capability. For practical next steps on building the underlying security controls that support AI governance, the RiskAware resource on data security best practices is a solid place to continue.

Secure your systems. Train your people. And deal with shadow AI before it deals with you.