Your Biggest AI Risk Is the AI You Do Not Know About

In November 2025, cybersecurity firm UpGuard released a study that should have been a wake-up call for every CISO and risk manager in the country.

More than 80% of workers, including nearly 90% of security professionals, were using AI tools their employer had not approved. Half said they used these tools regularly. Less than 20% used only company-sanctioned AI. The people hired to protect enterprise data were among the worst offenders.

A month later, BlackFog surveyed 2,000 workers at companies with over 500 employees and found 49% had adopted AI tools without employer approval. Many used free consumer versions that offer no data protection guarantees. Perhaps the most alarming finding: 69% of C-suite executives were comfortable with this, prioritizing productivity over security.

Welcome to the shadow AI problem. And unlike shadow IT, which took years to become a widespread enterprise concern, shadow AI reached crisis scale in months. The tools are free, browser-based, and devastatingly effective at making people more productive. That combination makes them nearly impossible to resist and even harder to govern.

This guide is for risk managers, CISOs, and compliance leaders who need to move from awareness to action. It covers what shadow AI actually is (and how it differs from shadow IT), the real financial and regulatory exposure it creates, practical detection methods that work in 2026, a governance framework you can implement in 90 days, and KRIs to track whether your controls are working. Everything maps to your existing enterprise risk management framework because shadow AI is not a separate program. It is a risk category that belongs in your ERM.

What Shadow AI Is and Why It Is Different from Shadow IT

Shadow AI is the unauthorized use of AI tools and platforms by employees without IT approval, security review, or governance oversight. It includes employees using ChatGPT, Claude, Gemini, Copilot, Perplexity, and dozens of smaller tools through personal accounts, free tiers, and browser extensions to do their jobs faster.

If you managed shadow IT a decade ago, you might think this is familiar territory. It is not. Shadow AI is qualitatively different in four ways:

The Komprise COO put it plainly: shadow AI is a “much greater problem than shadow IT, which primarily focuses on departmental power users purchasing cloud instances or SaaS tools without IT approval. Now we have got an unlimited number of employees using tools like ChatGPT or Claude AI to get work done, but not understanding the potential risk they are putting their organizations at.”

The Scale of the Problem: Shadow AI by the Numbers

The research data from 2025 and early 2026 paints a consistent and sobering picture. These are not projections. These are measurements of what is happening right now in US enterprises:

Read those numbers together and the picture is clear: the vast majority of enterprises have widespread, uncontrolled AI usage that has already caused measurable financial harm. The gap between AI adoption speed and AI governance capability is where shadow AI thrives and where the costs accumulate.

Shadow AI Risk Taxonomy: Six Categories of Exposure

Shadow AI creates risk across multiple dimensions. Mapping these into a structured taxonomy helps you prioritize controls and integrate shadow AI into your existing risk control self-assessment process:

Detection Methods: Finding Shadow AI in Your Organization

You cannot govern what you cannot see. Detection is the first step. The challenge is real: most AI tools operate over encrypted HTTPS connections, and employees increasingly use personal devices and off-network access that bypasses corporate gateways entirely. A 2025 F5 report highlighted that organizations need detection and prevention specifically for AI workloads over encrypted channels, an emerging capability gap.

Effective detection combines five approaches. No single method is sufficient. You need layered visibility:

Layer 1: Network and Cloud Monitoring

Cloud Access Security Brokers (CASBs) are your primary detection tool. Modern CASBs can flag connections to known AI service endpoints (OpenAI, Anthropic, Google AI, Hugging Face, and hundreds of smaller providers). Cato Networks, Netskope, Zscaler, Palo Alto Networks, and Microsoft Defender for Cloud Apps all offer AI-specific CASB capabilities as of 2025. Deploy your CASB to discover and classify AI tools in use across your network, categorize discovered tools by risk level, monitor data volumes flowing to AI endpoints, and alert on new AI services that appear in your environment.

Layer 2: Data Loss Prevention at the Prompt Level

Standard DLP was not built for AI prompts. You need AI-aware DLP that can inspect content at the prompt level, including typed inputs, file uploads, and clipboard pastes. As Proofpoint describes it, modern DLP tools scan for sensitive data going to major chatbots including ChatGPT, Copilot, Claude, Gemini, and Perplexity.

Context-aware DLP reduces false positives by understanding the data flow, not just pattern matching. Configure DLP to block Tier 3 and Tier 4 data (using your data classification scheme) from reaching any external AI endpoint. Log all blocks and alert the user with a clear explanation of why the submission was stopped.

Layer 3: Endpoint and Browser Visibility

AI has moved into browser extensions, desktop apps, and embedded features within existing SaaS tools. IntelligenceX’s 2025 analysis identified several detection signals that go beyond network traffic: browser extension scanning to identify installed AI tools, monitoring for unusual GPU usage or LLM telemetry signals, detecting model API tokens in system configurations, and tracking AI-specific executables or command-line activity. Implement allowlisting for browser extensions through group policy or endpoint management. Create alerts when new AI-related extensions appear on managed devices.

Layer 4: Behavioral Analytics

Sometimes shadow AI reveals itself through behavioral patterns rather than direct detection. CIO.com’s coverage of the shadow AI phenomenon noted that auditors and analysts can identify patterns deviating from established baselines: a marketing account suddenly transmitting structured data to an external domain, a finance user issuing repeated calls to a generative API, large data exports followed by external connections. User behavior analytics (UBA) tools can detect these anomalous patterns when trained to recognize AI-specific indicators.

Layer 5: Human Intelligence (the Most Underrated Method)

CIO.com made a crucial observation: “Employees are often willing to disclose AI use if disclosure is treated as learning, not punishment. A transparent declaration process built into compliance training or self-assessment can reveal far more than any algorithmic scan.

Run a confidential AI usage survey. Ask teams what tools they use, what data they access, and what problems they are solving with AI. Frame it as discovery, not enforcement. Combine survey results with your technical detection findings to build a complete shadow AI inventory.

The Komprise survey found 75% of IT leaders plan to use data management technologies for shadow AI risk, and 74% are investing in AI discovery and monitoring tools. Do both.

Shadow AI Detection Technology Comparison

A critical caveat from BlackFog: traditional DLP and CASB tools are often blind to shadow AI because generative AI services operate via HTTPS web traffic. Without comprehensive SSL inspection (which many organizations do not implement end to end), network-based tools cannot see inside encrypted sessions.

Employees on personal devices or off-network compound this problem. This is why layered detection, combining technical, behavioral, and human intelligence methods, is essential.

Shadow AI Governance Framework: From Discovery to Control

Detection tells you what is happening. Governance tells you what to do about it. ISACA, the World Economic Forum, and Gartner all converge on a similar governance model. Here is a five-phase framework that integrates with your existing ERM governance structure:

Phase 1: Discovery and Inventory

Deploy the detection methods described above. Build a living AI tool inventory that maps each discovered tool to: business unit using it, use cases and data types being processed, risk classification (Unacceptable, High, Moderate, Low), data residency (where is data processed and stored), and vendor terms of service (does the vendor use data for training?).

The World Economic Forum’s 2025 report emphasized that this technology audit is the foundation for transparency, accountability, and resilience in AI governance.

Phase 2: Risk Assessment and Classification

Not all shadow AI poses the same risk. Classify discovered tools using your existing risk assessment methodology. Proofpoint recommends creating a risk heat map considering data sensitivity, regulatory exposure, and business impact.

A developer using GitHub Copilot for non-proprietary code is a different risk profile than a finance analyst pasting quarterly earnings data into free ChatGPT.

For tools in each risk category, determine your response: sanction (approve and add to the approved tools list), replace (provide an approved alternative that meets the same need), restrict (allow with additional controls and monitoring), or block (prohibit entirely and enforce through technical controls).

The ISACA framework recommends integrating this classification directly into your enterprise risk register.

Phase 3: Policy and Approved Alternatives

Draft an acceptable use policy that covers all generative AI tools (see our companion guide: How to Build a Generative AI Acceptable Use Policy).

But here is the critical insight that most organizations miss: if you do not provide approved alternatives, employees will keep using shadow AI. The UpGuard report found employees use unapproved tools because they believe they can manage the risks themselves.

That confidence is misplaced, but the underlying need is real. Build an “AI App Store” or approved tools catalog. Enterprise versions of ChatGPT, Claude, and Copilot cost money, but they contractually guarantee data is not used for model training, provide audit logging, and support SSO integration.

The cost of enterprise AI licenses is a fraction of one shadow AI data breach ($670,000 premium per IBM).

Phase 4: Technical Controls and Enforcement

Deploy the detection and prevention stack from the technology comparison above. Key controls include CASB with AI-specific policy enforcement (block unsanctioned AI endpoints, allow approved tools with granular conditions).

DLP with prompt-level inspection (block sensitive data before it reaches any AI tool), browser extension controls (allowlist approved extensions, alert on unauthorized installations), SSO-only access to approved AI tools (no personal accounts for work use), and network-level blocking of consumer AI endpoints for managed devices (optional depending on your enforcement posture). For integration with your broader cybersecurity and ERM program, route shadow AI alerts into your existing SIEM and incident management workflows.

Phase 5: Training, Culture, and Continuous Monitoring

CybSafe and NCA’s survey found 52% of employed participants had never received training on safe AI use. That gap drives shadow AI. ISACA recommends regular training sessions covering ethical considerations, data privacy, bias and fairness, regulatory requirements, transparency and accountability, and the practical implications of AI use.

But training alone is not enough. You need a culture shift. Komprise’s research found that blocking AI tools is “largely ineffective.” Instead, foster a culture where employees understand why governance matters, know how to request new AI tools through a formal process, feel safe disclosing past shadow AI usage without punitive consequences (amnesty for disclosure), and see IT and security as enablers, not blockers.

Shadow AI Risk Register Template

Integrate shadow AI risks into your operational risk register. Here is a starter template with populated examples. For guidance on comprehensive risk registers, see our RCSA guide:

KRIs for Shadow AI Monitoring

Your shadow AI governance program needs measurable indicators. These KRIs should feed into your existing risk dashboard and board reporting framework:

The approved tool adoption rate is your single most important KRI. If you provide enterprise AI tools with proper governance but employees still use consumer alternatives at high rates, your program is failing.

Either the approved tools do not meet their needs (fix the tools) or your enforcement is ineffective (fix the controls). Track this monthly and investigate any decline.

Integrating Shadow AI into Your ERM and BCM Programs

Shadow AI risk management does not exist in a silo. Here is how to connect it to your existing governance frameworks:

• Enterprise Risk Register: Add shadow AI scenarios from the risk register template above to your operational risk register. Map to existing risk appetite thresholds and escalation criteria. Shadow AI creates risk across data security, regulatory compliance, and operational effectiveness, so it touches multiple risk categories in your existing taxonomy.
• Business Impact Analysis: If departments have become dependent on AI tools for critical processes (and they have), those tools are dependencies in your business impact analysis.
• Ask: what processes would be disrupted if all AI tools became unavailable? What is the RTO? What manual workarounds exist? Most organizations will discover they have undocumented AI dependencies in critical workflows.
• Business Continuity Planning: Build AI tool outage scenarios into your BCP exercise program. Test what happens when approved AI tools go down.
• Test what happens when you need to block all AI access during an active data breach. Your disaster recovery plans need to address scenarios where AI-generated content was compromised and needs to be recalled.
• Incident Response: Define shadow AI-specific incident types: sensitive data submitted to unauthorized AI, AI-generated content published without review causing harm, vendor ToS change requiring immediate response, AI tool compromise creating backdoor access. Route these through your existing incident management framework.
• Three Lines Model: First line (business units) owns approved tool usage and data classification compliance. Second line (risk, compliance, IT security) owns the governance framework, detection tooling, KRI monitoring, and policy.
• Third line (internal audit) tests control effectiveness through shadow AI simulation exercises and compliance audits.
• Board Reporting: Include shadow AI risk in your quarterly risk dashboard. Gartner predicts AI governance will become a requirement of all sovereign AI laws by 2027. Boards need visibility into shadow AI exposure, control maturity, and regulatory readiness. Use the KRIs above for a traffic-light dashboard that senior leadership can act on.

90-Day Implementation Roadmap

Five Mistakes That Guarantee Shadow AI Governance Failure

1. Banning AI instead of governing it. Samsung banned ChatGPT in 2023. UpGuard found in 2025 that even security professionals use unauthorized tools at rates exceeding 80%. A ban without approved alternatives is a signal to employees that you do not understand the problem. They will route around it. Every time.

2. Treating shadow AI as an IT problem. Shadow AI is a business risk problem. It touches data privacy (Legal), information security (CISO), human resources (HR), regulatory compliance (Compliance), vendor management (Procurement), and operational effectiveness (business units). ISACA recommends cross-functional governance teams spanning IT, legal, compliance, HR, and cybersecurity. If your CISO is running this alone, your governance will fail.

3. Buying detection tools without a policy framework. Detection without governance is expensive monitoring of a problem you are not solving. You will generate thousands of alerts you cannot act on because you have not defined what is acceptable, what is prohibited, and what the consequences are. Build the policy framework first, then deploy tools to enforce it.

4. One-time training instead of continuous awareness. AI tools, risks, and regulations change quarterly. Annual compliance training is not sufficient. Build AI governance into your regular communications cadence. Share incidents (anonymized), update the approved tools list publicly, and run tabletop exercises that simulate shadow AI scenarios.

5. Ignoring the “why” behind shadow AI. Employees use unauthorized AI because it makes them significantly more productive and their employer has not provided a governed alternative.

The Komprise COO was right: shadow AI is a problem because employees want to work faster and do not understand the risk. Address both. Provide tools that meet their needs AND education that builds risk awareness. One without the other fails.

The Bottom Line

Shadow AI is the fastest-growing risk blind spot in enterprise security. The statistics are unambiguous: 80% of your employees are using AI tools you did not approve, nearly half your IT leaders have already seen negative outcomes, and the average shadow AI breach costs $670,000 more than a standard incident. Gartner predicts 40% of AI-related data breaches by 2027 will come from uncontrolled GenAI misuse.

The organizations that will manage this successfully are those that treat shadow AI as an enterprise risk management problem, not a technology problem.

That means cross-functional governance, layered detection, risk-based controls, approved alternatives that employees actually want to use, and continuous monitoring with clear KRIs tied to board reporting.

The 90-day roadmap in this guide gives you a practical path from discovery to governance. Start with the shadow AI audit. You will not like what you find. But you will have something infinitely more valuable than blissful ignorance: a baseline you can manage.

Strengthen your AI governance program with riskpublishing.com: Enterprise Risk Management Frameworks | Key Risk Indicators | RCSA Risk Management | ERM Cybersecurity Integration | Business Continuity Elements | BCP Auditing | ISO 22301 BCM | ERM Technology | Cloud ERM | Risk Management Integration

Sources and References

• UpGuard: Shadow AI in the Enterprise (November 2025): 80%+ of workers use unapproved AI tools, 90% of security professionals — https://www.cybersecuritydive.com/news/shadow-ai-employee-trust-upguard/805280/
• BlackFog: Shadow AI and the Expanding Enterprise Attack Surface (January 2026): 49% employee unauthorized AI adoption; 97% lack AI usage controls — https://www.blackfog.com/shadow-ai-and-expanding-enterprise-attack-surface/
• Komprise: IT Survey: AI, Data and Enterprise Risk (June 2025): 90% IT leaders concerned; 79% experienced negative AI incidents; 13% with financial/reputational damage — https://www.komprise.com/resource/ai-survey/
• IBM: 2025 Cost of Data Breach Report: Shadow AI = 20% of all breaches; $670,000 cost premium per AI-associated breach — https://www.ibm.com/think/topics/shadow-ai
• Cisco: 2025 Data Privacy Benchmark: 46% of organizations experienced data leakage through generative AI prompts — https://www.cisco.com
• Menlo Security / Proofpoint: 2025 Shadow AI Analysis: 68% personal accounts; 57% sensitive data; 50% traffic growth to AI sites; 10.53B monthly visits — https://www.proofpoint.com/us/threat-reference/shadow-ai
• Varonis: State of Data Security Report 2025: 99% of enterprises had sensitive data exposed to AI tools — https://www.varonis.com
• Gartner: Predicts 2025: 40%+ of AI data breaches from cross-border GenAI misuse by 2027; AI governance mandatory in all sovereign AI laws by 2027; AI TRiSM reduces inaccurate info by 50% by 2026 — https://www.gartner.com/en/newsroom/press-releases/2025-02-17-gartner-predicts-forty-percent-of-ai-data-breaches-will-arise-from-cross-border-genai-misuse-by-2027
• ISACA: The Rise of Shadow AI: Auditing Unauthorized AI Tools in the Enterprise (2025) — https://www.isaca.org/resources/news-and-trends/industry-news/2025/the-rise-of-shadow-ai-auditing-unauthorized-ai-tools-in-the-enterprise
• Knostic / Vectra AI: AI Governance Tools: Selection and Security Guide (2026): 65% of AI tools operate without IT approval — https://www.vectra.ai/topics/ai-governance-tools
• Mindgard / IntelligenceX: 2025 Shadow AI Security Research: 76% estimate security teams use ChatGPT/Copilot without approval; 12% do not know what data is being input — https://blog.intelligencex.org/shadow-ai-enterprise-risk-governance-2025
• CIO.com: Shadow AI: The Hidden Agents Beyond Traditional Governance (November 2025) — https://www.cio.com/article/4083473/shadow-ai-the-hidden-agents-beyond-traditional-governance.html
• World Economic Forum: 2025 AI Governance Report: Technology audit foundation for transparency and accountability in AI governance — https://www.weforum.org

Proofpoint / Cato Networks: CASB and DLP GenAI Security Controls for shadow AI detection and enforcement (2025) — https://www.prnewswire.com/news-releases/cato-networks-introduces-generative-ai-security-controls-for-cato-casb-to-mitigate-shadow-ai-risk-302428212.html

Chris Ekai

Chris Ekai is a Risk Management expert with over 10 years of experience in the field. He has a Master’s(MSc) degree in Risk Management from University of Portsmouth and is a CPA and Finance professional. He currently works as a Content Manager at Risk Publishing, writing about Enterprise Risk Management, Business Continuity Management and Project Management.