Tech BlogNovember 18, 2026Eunji Han1 views

What is Shadow AI? Risks, Detection Methods, and an AI Governance Management Checklist

Shadow AI refers to the use of AI tools in business operations that are not approved or managed by the company. This document compiles the differences from Shadow IT, reasons for its emergence, information leakage and regulatory risks, a comparison of detection methods, usage policies, guardrail management, and a 30-60-90 day AI governance implementation plan.

#shadow ai#what is shadow ai#shadow ai#what is shadow ai#shadow ai detection#shadow ai detection#shadow ai management#shadow ai risks#shadow it#shadow it differences#ai governance#what is ai governance#ai governance meaning#ai governance guidelines#ai governance framework#generative ai security#generative ai internal usage policy#chatgpt internal usage#local llm security#ai dlp#ai guardrail
What is Shadow AI? Risks, Detection Methods, and an AI Governance Management Checklist
Eunji Han

November 18, 2026

Original Content:

Key Answer: Shadow AI refers to employees using AI tools, models, or features for work purposes that have not been approved or managed by the company. This includes chatbots in personal accounts, browser AI extensions, local LLMs downloaded to PCs, and newly activated AI features within SaaS applications. The primary risk is the leakage of confidential and personal information through prompts and uploaded files. Rather than outright blocking, it is more practical to manage it by "making it visible, providing approved alternatives, and obscuring and logging activity in the path."

  1. Discover: Identify actual usage status through network/DNS logs, SaaS detection, endpoint inspections, browser extension lists, and corporate card/subscription histories.
  2. Define: Document allowed, conditionally allowed, and prohibited lists, along with criteria for information prohibited from input, as a usage policy.
  3. Provide Alternatives: Offering approved in-house AI channels is essential to reduce circumvention.
  4. Control: Mask sensitive information using gateways/guardrails and maintain prompt-level logs.
  5. Integrate into Governance: Incorporate responsible persons, AI asset lists, risk assessments, and regular re-audits into the AI governance framework.

Last Updated: November 2026

What is Shadow AI? Meaning and Differences from Shadow IT

Shadow AI refers to AI used in business operations without the knowledge of security or IT departments. While it is a branch of Shadow IT, the nature of its risks differs because "the entered content itself is transmitted externally."

Shadow IT refers to the use of unauthorized software or services such as cloud storage, messengers, and SaaS. Shadow AI shares the characteristic of being used without approval, but its use involves users directly pasting business data for tasks like document summarization, code review, or email drafting, making each request a data transfer. Furthermore, if AI features are added as an update within an already approved SaaS, the tool remains approved, but its AI functionalities may not have been reviewed.

CategoryShadow ITShadow AI
ScopeUnauthorized SaaS, cloud storage, messengers, personal devicesUnauthorized chatbots, AI extensions, local LLMs, in-SaaS AI features, coding assistants
Data FlowFile storage/sharing-centricData transfer with each request via prompts, uploads, and conversation history
Additional RisksLack of access control/backupPotential for input data to be used for training, incorrect responses (hallucinations), copyright/licensing issues
VisibilityRelatively identifiable at the domain/app levelInternal features of approved apps, locally executed models are difficult to detect solely via network
Managing AuthorityIT/Security departmentIT/Security + Personal Information/Legal/Business units involved together

Considering scenarios where AI agents call tools on behalf of users, the scope expands further. The structure and risks of agents are detailed separately in What are AI Agents? Meaning, Types, Use Cases, and Security Risks When Enterprises Adopt Them.

Why Does Shadow AI Emerge?

Shadow AI typically emerges not out of malicious intent, but due to the desire to "complete work quickly when approved means are unavailable or inconvenient."

  • Lack of Approved Alternatives: If companies do not provide available AI tools, employees use public services with personal accounts.
  • Low Barrier to Entry: Browser extensions and free tiers can be used immediately without installation or payment approval processes.
  • Addition of AI Features to Existing Tools: When AI features are activated by default in document, collaboration, and email tools, users may not perceive themselves as using new tools.
  • Proliferation of Local LLMs: Running open-source models directly on a PC can easily lead to the misunderstanding that it is safe due to no external data transmission, but the model source, license, and output quality remain unreviewed.
  • Ambiguity of Policies: If only vague notices like "Caution when using AI" are provided, employees may not know what is allowed and make their own judgments.

Risks of Shadow AI: What Are the Problems?

The greatest risk is information leakage through prompts and uploads, compounded by copyright/licensing issues, regulatory violations, and reliance on incorrect responses.

OWASP addresses LLM02:2025 Sensitive Information Disclosure in its Top 10 LLM Application Risks 2025, explaining that personal information, financial information, health information, confidential business data, and security credentials can be exposed through LLMs and their applications. Countermeasures proposed include data sanitization, least-privilege access control, user education to avoid sensitive inputs, and offering opt-out options for training data usage. Reports covering the National Intelligence Service's "Security Guidelines for Using Generative AI such as ChatGPT" also advised against entering sensitive information, including personal data, within an organization and emphasized proper use and management of plugins and extensions.

Risk TypeExampleCountermeasure
Confidential/Personal Information LeakagePasting customer lists, contracts, source code, or error logs into a personal chatbot or uploading them as filesCriteria for prohibited input information, masking before transmission, upload inspection, unified approved channels
Unclear Training/Retention PolicyNot verifying if conversation history in free tiers is used for service improvementReviewing training usage/retention periods for each approved service, confirming opt-out settings
Copyright/LicenseIncluding AI-generated code/images in products without source verification, using local models with unknown license conditionsGenerated content review procedures, model/tool license verification, usage restrictions
Regulatory/Contract ViolationProcessing personal information beyond the scope of data subject consent, violating client non-disclosure agreementsPersonal information/legal review, confirmation of processing purpose/delegation relationship, log retention
Reliance on Incorrect ResponsesUsing AI responses for reports, legal, or security judgments without verificationMandatory human review, prohibited list for high-risk uses
Local LLMs/Extensions/CopilotsExecuting model files from unknown sources, browser AI extensions reading entire pages, coding assistants accessing entire repositoriesEndpoint detection, extension allowlist, repository access scope limitation

Shadow AI Detection Methods: Where and How to Find It?

No single method provides complete visibility, therefore overlapping network, SaaS, endpoint, browser, and cost data is necessary to reveal the full scope of Shadow AI.

Network-based detection is effective for identifying known AI service domains, but it often misses AI features within approved SaaS applications or local models running on PCs. Conversely, endpoint inspection can detect local models and extensions, but it cannot monitor personal devices that are not managed.

Detection MethodScope of VisibilityLimitations
Network/Proxy/DNS LogsAI service domain access, user/frequency of access, upload volumeNew services are missed until domain lists are updated, encrypted traffic content requires separate inspection, excludes external networks/personal devices
CASB/SaaS DetectionList of SaaS in use, risk ratings, corporate account-linked appsRequires separate verification for activation status of internal AI features within approved SaaS
Endpoint DetectionInstalled AI apps, local LLM execution, browser AI usage, clipboard/upload activitiesLimited to managed devices with agents installed, operational burden
Browser Extension ListInstalled AI extensions and requested permissions (e.g., page reading)Only collectible from devices with browser management policies applied
Cost/Subscription HistoryAI subscriptions paid with corporate cards/expenses, departmental usage scaleFree tiers and personal payments are not visible
Surveys/InterviewsPurpose of use, required features, discomfortsProne to omissions as self-reported, supplementary to other methods

Taking the endpoint area as an example, KYRA AI PC provides functions for discovering and classifying AI tools used per PC (local LLM detection, browser AI monitoring, Shadow AI identification, usage history recording) and a usage status dashboard. When used in conjunction with network/SaaS detection, it helps consolidate "which departments are using which AI in what manner" into a single list.

Shadow AI Management Methods: Block or Allow?

Effective Shadow AI management approaches combine "providing approved alternatives + progressively blocking only risky uses + masking and logging in the path" rather than implementing a complete ban.

1. Usage Policies and Approved Lists

Tools should be categorized as allowed, conditionally allowed (e.g., mandatory company account and non-usage for training), or prohibited. Information that should not be entered (e.g., personal information, customer secrets, source code, authentication credentials, undisclosed financial information) should be specified with concrete examples. The structure of policy document items is detailed as a checklist in Generative AI Usage Guidelines and Security.

2. Approved In-house AI as an Alternative

If only blocking is implemented, users will migrate to personal devices or external networks, making detection even more challenging. Providing summary, translation, and draft generation functionalities through approved channels, and streamlining the application and approval process, is the most direct way to reduce circumvention.

3. Masking and Logging with Gateways/Guardrails

Approved channels should incorporate control points between users and LLMs. KYRA AI Guardrail acts as a gateway positioned between users/agents and LLMs, masking information such as resident registration numbers, account numbers, credit cards, contact details, and emails before transmission. It also implements prompt injection defense, response scope limitation, and tool call control, and can send prompt-level audit logs to SIEM. For methods on defining masking criteria, please refer to Personal Data Masking Criteria and Methods: LLM/Generative AI Application Guide.

At the endpoint, it may be necessary to apply policies on PCs even without a network connection. KYRA AI PC implements context-aware DLP (sensitive information blocking, real-time masking, clipboard control, upload inspection) directly on the device.

4. Training

The "Personal Information Protection Guide for Generative AI Service Users" published by the Personal Information Protection Commission in May 2026, provides guidance from a user perspective on issues such as whether input content is used for training, precautions when entering business data, opt-out options for training usage, and settings for conversation history storage and deletion. When creating internal training materials, it is advisable to use such official documents as a starting point and supplement them with examples of allowed/prohibited practices based on company policy.

5. Phased Blocking and Allowing

  • Observation Phase: Collect logs without blocking to understand actual usage patterns.
  • Warning Phase: Provide warnings and guidance to approved channels upon detection of unauthorized service access or sensitive information input.
  • Control Phase: Begin blocking high-risk activities (e.g., large-volume uploads, inputs containing personal information, extensions from unknown sources).
  • Review: Regularly update the approved list based on exception requests and blocking logs.

How to Address Shadow AI within AI Governance?

Shadow AI management constitutes the initial phase of AI governance, specifically "knowing what is being used," and must be linked to responsible persons, asset lists, risk assessments, and continuous monitoring to be sustainable.

The U.S. NIST released the AI Risk Management Framework (AI RMF 1.0) on January 26, 2023, comprising four functions: Govern, Map, Measure, and Manage. On July 26, 2024, it also published a profile for applying it to generative AI (NIST AI 600-1). In South Korea, the Personal Information Protection Commission's "Personal Information Processing Guidelines for Generative Artificial Intelligence (AI) Development and Utilization," released on August 6, 2025, targets businesses utilizing commercial LLM services and proposes a CPO-centric governance framework. The National Intelligence Service distributed its revised "AI Security Guidebook" (from 2023 guidelines) to public institutions and others in December 2025.

Governance ElementActions from a Shadow AI PerspectivePrimary Responsible Party
Roles/ResponsibilitiesDesignate AI usage approvers, exception approval procedures, incident reporting channelsManagement, CISO, CPO
AI Asset ListManage detection results in lists categorized by tool, model, department, and data typeIT/Security
Risk AssessmentReview data processing location per tool, training usage, retention period, licenseSecurity, Personal Information, Legal
Control ImplementationMasking, upload inspection, extension allowlist, access permissionsSecurity
Monitoring/AuditPrompt logs, blocking/warning history, regular re-auditsSecurity Operations, Audit

The relationship with business obligations under the AI Framework Act can be found in AI Framework Act Implementation Details and Corporate Obligations Checklist, and the operational governance framework in AI Governance and LLMOps Checklist (AX Part 5).

Shadow AI Management 30-60-90 Day Implementation Plan

The first 30 days focus on status assessment, up to 60 days on policy and alternative provision, and up to 90 days on control and governance establishment.

Days 1-30: Achieving Visibility

  • Designate responsible parties (Security, Personal Information, IT, Business Unit representatives).
  • Extract AI service access status from proxy/DNS logs.
  • Mark apps with AI features in CASB/SaaS lists.
  • Collect status of AI apps, local LLMs, and browser AI extensions on managed PCs.
  • Identify AI subscriptions from corporate card/expense records.
  • Confirm departmental usage purposes via a brief survey.

Days 31-60: Policy and Alternatives

  • Finalize allowed, conditionally allowed, and prohibited lists, along with criteria for information prohibited from input.
  • Provide approved in-house AI channels, starting with frequently used purposes.
  • Review training usage/retention settings for approved services.
  • Establish an exception request procedure and define processing deadlines.
  • Conduct training for all employees.

Days 61-90: Control and Establishment

  • Implement masking and prompt logging in approved channels and integrate with SIEM.
  • Start upload/clipboard controls on endpoints in warning mode and gradually transition to blocking.
  • Begin blocking unauthorized extensions and high-risk services.
  • Incorporating AI asset lists and risk assessment results into governance documents.
  • Set a schedule for quarterly re-audits.

Summary at a Glance

  • What is Shadow AI: It refers to using AI tools/features for work that are not approved or managed by the company.
  • Differences from Shadow IT: Each request constitutes data transfer, and many cases, like AI features within approved apps or local models, are not easily visible.
  • Key Risks: Leakage via prompts/uploads, copyright/licensing, regulatory/contract violations, reliance on incorrect responses.
  • Detection: Overlap network, SaaS, endpoint, browser, and cost data.
  • Management: Use a combination of policies, approved lists, approved alternatives, masking and logging, training, and phased blocking.
  • Governance: Sustain through responsible parties, AI asset lists, risk assessments, and monitoring.

Frequently Asked Questions

Should Shadow AI be completely blocked?

Complete blocking often does not resolve the issue. If business demand remains, users will migrate to personal devices or external networks, making it even harder for the company to monitor usage. It is more practical to first provide approved alternatives and then progressively block high-risk activities, such as sensitive information input or extensions from unknown sources.

What is AI Governance?

AI governance is a framework that defines who makes decisions, how risks are assessed, and what controls and audits are applied when an organization adopts, uses, and operates AI. The NIST AI RMF describes this with four functions: Govern, Map, Measure, and Manage. Shadow AI detection serves as the starting point for creating an AI asset list within this framework.

Are Shadow AI and Shadow IT the same?

Shadow AI can be considered a branch of Shadow IT, but its usage inherently involves inputting business data into external models, making the risk of information leakage more direct. Additionally, cases like AI features within approved SaaS or local models are often not easily detected by existing Shadow IT detection methods.

Are local LLMs safe because data is not transmitted externally?

While the risk of external transmission may be reduced, the source and integrity of the model file, license conditions, and response quality must be reviewed separately. Furthermore, if a local model is configured to access company documents, the access controls on that PC determine the data protection level. It is advisable to manage them by including them in an approved list.

What information should not be entered into generative AI?

Generally, personal information, customer/client secrets, source code, authentication credentials (passwords, API keys), and undisclosed financial/HR information are typical examples of prohibited input. Reports covering the National Intelligence Service's guidelines also advise against entering sensitive information, including personal data, within an organization. The actual criteria should be clearly specified in company policy with examples.

Where should Shadow AI detection begin?

Start with existing data. Proxy/DNS logs and corporate card subscription histories can quickly reveal the usage status of major AI services. This can then be supplemented by endpoint and browser extension inspections to cover usage that is not easily visible on the network, such as local models and extensions.

References

Stay Updated

Get the latest security insights delivered to your inbox.

Tags

#shadow ai#what is shadow ai#shadow ai#what is shadow ai#shadow ai detection#shadow ai detection#shadow ai management#shadow ai risks#shadow it#shadow it differences#ai governance#what is ai governance#ai governance meaning#ai governance guidelines#ai governance framework#generative ai security#generative ai internal usage policy#chatgpt internal usage#local llm security#ai dlp#ai guardrail