TL;DR: Unapproved AI use is being driven by workflow pressure and fear of falling behind, and more than half of employees are already using shadow AI while over half of those users feed sensitive company data into tools their organisations do not manage, monitor, or even know about, according to Aikido. The control problem is not only tool approval, it is visibility, workflow fit, and safe alternatives that keep sensitive data out of unmanaged models.
At a glance
What this is: Shadow AI is presented as a governance and behaviour problem, with employees using unapproved tools because approved paths lag behind real work.
Why it matters: IAM, NHI, and AI security teams need to treat unsanctioned AI use as an access and data-governance issue, not only an acceptable-use policy problem.
Context
Shadow AI emerges when employees adopt AI tools outside approved governance because the sanctioned path is too slow, too narrow, or too disconnected from daily work. In practice, that creates an access-control problem as much as a policy problem, because sensitive information can move into systems security teams do not inventory, monitor, or bound.
The identity angle is real even when the subject is not traditional IAM. Unmanaged AI tools behave like ungoverned third-party services, often reached through personal accounts, browser sessions, or ad hoc sharing of documents and prompts. That means the control gap sits at the boundary between user behaviour, data handling, and sanctioned access paths.
Key questions
Q: What breaks when organisations ban shadow AI instead of governing it?
A: Bans often push AI use into personal accounts, unmanaged devices, and hidden workflows, which removes visibility from security and makes data exposure harder to detect. The control failure is not usage itself, but concealment. A better model is to approve fast, workable alternatives and enforce policy on identity, data handling, and logging.
Q: Why do shadow AI tools create identity governance risk?
A: Shadow AI is risky because users often reach those tools through identities, browser sessions, or tokens that were never assessed for data handling or access scope. The issue is not just policy compliance. It is whether the identity path into the tool is authorised, reviewable, and reversible.
Q: What are the signs that shadow AI is being driven underground?
A: Look for AI use that appears in personal accounts, browser extensions, or devices outside the approved stack, especially when teams keep asking for exceptions or quietly switching tools. Rising workflow frustration and repeated attempts to bypass policy are strong signals that the sanctioned path does not meet business needs.
Q: How should security teams respond when employees need AI tools but approved options lag?
A: Security teams should shorten intake, define acceptable use for sensitive data, and provide sanctioned options that solve the same job-to-be-done. If the approved route is slower or less useful, employees will continue to hide the tools they use, and the organisation will lose both visibility and control.
Technical breakdown
Why shadow AI becomes a governance failure
Shadow AI is not simply careless use of new tools. It appears when employees conclude that the approved path does not help them finish work quickly enough, so they move to unmanaged services that solve the immediate problem. That creates an unsanctioned decision layer outside procurement, security review, and data policy. The result is not just tool sprawl but invisible data movement. Once users are comfortable putting work into an unmanaged model, the organisation loses control over retention, secondary use, and incident visibility.
Practical implication: security teams need to govern the use case and the data path, not just the application catalogue.
Why banning tools pushes usage underground
A ban does not remove demand when the underlying work still requires AI assistance. It usually shifts usage onto personal accounts, unmanaged devices, or shadow workflows where monitoring is weaker and reporting disappears. That is a familiar pattern from shadow IT and BYOD, but AI raises the stakes because the content entered into prompts can be highly sensitive and instantly copied into external systems. The policy failure is assuming that prohibition reduces risk when it often reduces visibility first.
Practical implication: replace blanket bans with enforceable guardrails that keep approved usage visible and auditable.
Why AI note-takers and chat tools create disclosure risk
AI note-takers and chat assistants sit directly in the path of confidential material, so they can transform normal collaboration into exposure if the output is shared with the wrong party or retained beyond the intended context. The article's litigation example shows the issue clearly: a tool summarised a confidential conversation and the result became part of an adversarial process. That is a classic data-governance failure, but it also reflects a trust boundary problem, because the organisation never formally accepted that tool into its control model.
Practical implication: classify meeting assistants and chat tools as data-handling systems that require explicit governance before use.
Threat narrative
Attacker objective: The end result is unauthorised access to sensitive corporate information through a workflow the organisation did not approve or monitor.
- Entry begins when an employee uses an unapproved AI tool because it appears faster or more useful than the sanctioned option.
- Credential or data exposure follows when the employee pastes confidential documents, prompts, or meeting content into a system the organisation does not manage.
- Impact occurs when that content is retained, reused, disclosed, or surfaced in a legal or security context outside the organisation's control.
Breaches seen in the wild
- Vercel Context.ai OAuth Supply Chain Breach: Shadow AI app Context.ai OAuth integration exposes Vercel customer data via unmanaged third-party token.
- Samsung ChatGPT leak 2023: Samsung staff pasted chip source code and meeting notes into ChatGPT weeks after it was allowed, leading Samsung to restrict generative AI tools.
Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.
NHI Mgmt Group analysis
Shadow AI is a governance gap before it is a tooling gap: the central failure is not that employees have access to AI, but that organisations have not made approved AI good enough to compete with the work people actually do. When the sanctioned path is slower than the business need, users route around it. That turns policy into theatre and makes data governance depend on luck rather than control.
The named concept here is approval-lag leakage: the longer the gap between employee need and approved capability, the more sensitive material escapes into unmanaged systems. This is not just shadow IT with a new label. AI collapses the time between intent and disclosure, so the governance model has to move from waiting for approval to setting safe defaults at the point of use.
Shadow AI changes the boundary of identity governance: the real question is no longer only who is allowed to use a tool, but which accounts, prompts, documents, and sessions are permitted to cross into an external model. That means identity teams, data teams, and security teams have to define where sanctioned access ends and uncontrolled delegation begins.
Banning is a weak control when the control objective is visibility: prohibition may satisfy policy language while increasing hidden usage. Mature programmes will treat shadow AI as a discoverability problem first, then a workflow and access problem, because controls cannot govern what they cannot see. The practical lesson is to reduce concealment, not just usage.
AI note-takers belong in the same governance conversation as third-party access: they can capture confidential exchanges, transform them into reusable artefacts, and move them into legal or operational workflows the organisation did not intend. That makes them a data-exposure and accountability issue, not merely a productivity choice. Practitioner conclusion: govern the path into the model, the output from the model, and the sharing of both.
From our research library:
- One in five organisations reported a breach due to shadow AI, and 97% of those breached through an AI model or application lacked proper AI access controls, according to IBM's 2025 Cost of a Data Breach Report.
- Read next: Shadow AI and AI Agent Discovery Guide
What this signals
Approval-lag leakage: when sanctioned AI tooling moves slower than employee demand, people route sensitive work into unmanaged systems. That shifts the problem from acceptable use policy to invisible data movement, so the first control objective is visibility into where prompts, transcripts, and uploads are actually going.
Shadow AI is converging with long-standing shadow IT lessons, but the governance bar is higher because the content itself can be transformed, retained, or disclosed by the model. Security teams should focus on safe alternatives, low-friction reporting, and controls that make approved use easier than concealment.
For practitioners
- Build a shadow AI inventory Map the AI tools employees already use across browsers, devices, and accounts before tightening controls. Include note-takers, chat assistants, and browser-based copilots, then classify where sensitive data is being entered.
- Design approved AI for real workflows Identify the top workflow gaps that drive unsanctioned use and replace them with sanctioned tools that fit how teams actually work. If the approved path is slower or less useful, the shadow path will persist.
- Treat AI note-takers as data-handling systems Require explicit review for meeting assistants that transcribe, summarise, or share conversations. Focus on retention, sharing, and downstream use of outputs, not just whether the tool is marketed as productivity software.
- Create safe disclosure channels Give staff a low-friction way to say which AI tools they are using without fear of immediate punishment. The goal is to surface hidden usage early enough for security to govern it.
- Shorten the approval cycle for high-demand tools Move away from month-long POCs for AI tools that change weekly. Establish a lighter intake process for low-risk use cases so the business does not route around security simply to keep pace.
Key takeaways
- Shadow AI is best understood as a governance and workflow failure, not a simple policy violation, because employees often adopt unapproved tools to keep up with work demands.
- The article ties hidden AI use to real exposure risk when confidential material enters tools the organisation does not manage or monitor.
- Security teams reduce risk fastest by improving visibility, fitting approved tools to real workflows, and making disclosure of AI use safe.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Non-Human Identity Top 10, OWASP Agentic AI Top 10 and MITRE ATT&CK address the attack and risk surface, while NIST CSF 2.0 sets the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Non-Human Identity Top 10 | NHI-10 — Human Use of NHI | Employees using unmanaged AI tools mirrors human use of non-human systems outside governance. |
| Recommendation — Map unsanctioned AI use to NHI-10 and govern human access to machine tools through approved, auditable paths. | ||
| OWASP Agentic AI Top 10 | ASI09 — Human-Agent Trust Exploitation | Employees are lured into unsafe AI use by trust and convenience pressures. |
| Recommendation — Apply ASI09 thinking to reduce trust abuse by making sanctioned AI safer and easier to use than shadow alternatives. | ||
| NIST CSF 2.0 | PR.AA-05 — Access Permissions, Entitlements and Authorizations | Shadow AI exposes a gap in controlling who can send data to external services. |
| PR.DS-01 — Data-at-rest security | Unmanaged AI use can place sensitive data into third-party environments beyond internal protection. | |
| Recommendation — Use PR.AA-05 to define which users and data types may cross into AI services and under what approval. Apply PR.DS-01 to keep sensitive content out of unapproved AI systems and enforce handling rules at the source. | ||
| MITRE ATT&CK | TA0001;TA0010 — Initial Access; Exfiltration | Shadow AI usage creates a path for sensitive data to leave the environment through user-driven workflows. |
| Recommendation — Track shadow AI exposure as a data-exfiltration path and prioritise discovery where users can upload sensitive content. | ||
Key terms
- Shadow AI: AI agents, copilots, or connected tools operating without full visibility or governance from security teams. Shadow AI becomes an identity problem when those systems authenticate with unmanaged tokens, service accounts, or OAuth apps that can reach production resources.
- Submission Lag: The time gap between a valid finding being discovered and that finding being accepted, credited, or routed into remediation. In agentic offensive work, submission lag can matter as much as discovery skill because speed determines whether the result is still actionable.
- Sanctioned AI: Sanctioned AI is an AI system that has gone through procurement, legal, and security review and is governed by defined controls. The term matters because approved status should reflect real access scoping, ownership, and data handling rules, not just a business decision to use the tool.
- AI Note-Taker Risk: AI note-taker risk is the possibility that meeting transcription or summarisation tools expose confidential conversation content through retention, sharing, training use, or unintended distribution. These tools can convert a private discussion into a durable artefact with broader legal and operational reach.
Deepen your knowledge
NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, workload identity, and secrets management in the context of modern access control. It helps security practitioners connect identity governance to the broader control problems created by unmanaged AI and machine-use patterns.
Published by the NHIMG editorial team on August 2, 2026.
Updated on October 12, 2026.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org