Key takeaways
- Shadow IT is business technology outside the organization’s managed processes; shadow AI applies that concern to AI tools and integrations.
- Review the account, data, connections, and allowed actions—not only whether a familiar application appears on an approved list.
- Give employees a usable approved route, limit agent permissions, and assign ownership for monitoring, exceptions, and response.
Original source
WIRED reportingWhat should a business understand about shadow AI and shadow IT?
Shadow AI and shadow IT create a management gap: technology is being used for company work without the ownership and controls the organization expects. At Datapath, we recommend finding those workflows, understanding the business need, and giving employees an approved way to complete the work. Useful AI adoption needs clear data boundaries and permissions as well as capable tools.
The report behind this briefing
On October 29, 2026, WIRED’s Vittoria Elliott reported allegations that an Accenture contractor connected a personal Claude Code agent to a DOD environment handling controlled unclassified information. Accenture told WIRED an independent investigation found the allegations without merit; DOD declined to comment on its systems. The report describes NIPRNet, an unclassified network, and says there is no indication attackers entered classified systems. These are disputed reported allegations, not an independently confirmed breach finding by Datapath.
Read WIRED’s original reporting and the supplied International Cyber Digest X post. The discussion post is not independent corroboration.
The sections below explain general business controls. The illustrative situations are not additional findings about the organizations in the report.
What is shadow IT?
The UK National Cyber Security Centre defines shadow IT around business-use assets that are unknown to, and therefore unmanaged by, the organization. It can include services and accounts as well as devices. The concern is the missing management process: who owns the asset, what information it holds, and how access or recovery is handled.
For an illustrative business example, a department might depend on a file-transfer account controlled by a former employee. Even if the service itself works correctly, the company may have no dependable process for removing access, preserving records, or restoring the workflow. Documenting the dependency is the first step toward an accountable decision about it.
What is shadow AI?
Microsoft’s shadow AI guidance describes AI use outside the knowledge, approval, or governance of an organization’s IT or security teams. NCSC also identifies unapproved AI use as a form of shadow IT.
AI adds questions about what information enters a prompt, which files are retrieved, what tools are connected, and whether the system can act. An approved product name is not enough to answer them. An employee’s personal account, an enterprise workspace, and an agent connected to production services can have different owners, terms, and access boundaries.
The distinction matters when a pilot expands. A tool initially used to draft public text may later receive internal documents or credentials for another system. That change deserves a fresh review even if the application icon and subscription name remain the same.
| Review question | Shadow IT concern | Additional AI concern |
|---|---|---|
| Who owns it? | Unmanaged account or service | Who owns prompts, agents, and connections? |
| What can it access? | Unrecorded business data | Which retrieval sources and tool permissions are available? |
| What can it do? | Uncontrolled business workflow | Can it send, change, deploy, or delete? |
| How is it reviewed? | Missing lifecycle process | Are model, tool, and permission changes evaluated? |
Why does unclassified not mean public?
The National Archives definition of controlled unclassified information explains that CUI requires safeguarding or dissemination controls under applicable authorities even though it is not classified. Classification and permission to share are different questions.
Businesses should apply the same careful distinction to their own information categories. A document can be ordinary company information and still contain confidential commercial details, employee records, or customer information. These are illustrative business categories, not a statement that every such document is CUI. Use the organization’s classification and handling rules to decide whether a particular AI workflow may process it.
A convenient upload button is not evidence of approval. The responsible owner should know what information is permitted, which service receives it, what retention terms apply, and how employees can obtain help when a document falls outside the approved boundary.
Why do AI agents need a separate permissions review?
An agent can combine model outputs with tools that read information or perform actions. OWASP’s excessive-agency guidance recommends limiting functions and permissions, enforcing authorization in connected systems, and requiring human approval for high-impact actions. These controls address what the system is allowed to do, rather than relying only on instructions telling it to behave.
For a proposed business assistant, separate reading records, preparing a draft, and executing a change. Give the integration only the authority required for its approved task. A human who may deploy software does not automatically need an assistant that can deploy with the same broad credentials. Evaluate that delegation explicitly.
Test the boundary with realistic scenarios before expansion: another user’s records, an unexpected destination, a request outside the approved workflow, and a failed or interrupted action. Preserve useful audit evidence while avoiding unnecessary copies of secrets or confidential content. Monitoring helps find problems; it does not replace authorization.
How can organizations discover shadow AI without driving it further underground?
Ask teams which tools they use and which tasks the approved environment makes difficult. NCSC recommends a positive, no-blame approach and practical alternatives so people will disclose workarounds. A slow or confusing request process can make it harder to learn where company work is actually happening.
Combine those conversations with the visibility your existing environment lawfully provides, such as application inventories, identity grants, endpoint records, and supplier reviews. Explain the purpose and scope of the review to employees. Absence from one dashboard is not proof that a tool is unused, and observing a connection does not by itself prove misuse.
Record a named owner, business purpose, approved information, account type, connections, and review date for each material workflow. Decide whether to approve it, improve it, replace it, or stop it through the normal change process. Preserve relevant evidence and coordinate access changes when investigating a possible incident.
What should an approved AI rollout include?
Microsoft’s deployment guidance separates discovering AI use, restricting unapproved apps, protecting sensitive information even in approved apps, and governing interactions. That is a useful reminder to evaluate both the tool and the information flowing through it. Product-specific features depend on the organization’s configuration and licensing.
A practical review should produce decisions that employees can follow:
- Name the business owner and the specific workflow being approved.
- Identify allowed information and prohibited uploads or retrieval sources.
- Review supplier terms, retention, account administration, and exit options.
- Define permitted actions and the point at which a person must approve a consequential change.
- Establish support, monitoring, incident handoffs, and a review when permissions or integrations change.
NIST’s Generative AI Profile provides voluntary risk-management context. It can help structure governance work, but a framework is not automatic certification, legal compliance, or a guarantee that an implementation cannot fail.
How can Datapath help bring shadow AI into a managed process?
Datapath helps organizations connect technology use with clear owners and achievable implementation work. Our AI governance consulting covers readiness, data handling, approved tools, vendor questions, policy ownership, and rollout priorities. Our managed cybersecurity services connect monitoring, alert triage, response coordination, and remediation.
Start with one important workflow and the services, accounts, and information it depends on. We can help organize the review and responsibilities needed to give employees a usable approved path. Our incident response runbook guidance also explains how a lean team can prepare handoffs before an urgent issue arises.
Need a clearer path for company AI use?
Talk with Datapath about approved tools, data boundaries, agent access, and practical governance ownership.
Frequently asked questions
Is shadow AI different from shadow IT?
Shadow AI applies the unmanaged-technology concern to AI tools and workflows. Its review includes the information supplied to models, connected services, and actions an agent can perform, alongside ordinary account and service ownership.
Is every AI tool on an approved list safe for all company data?
No. Approval should specify the account, workflow, information, connections, and permitted actions. A change in those boundaries may require another review even when the product name stays the same.
Should an AI agent receive an employee’s full access automatically?
No. Review the delegation and give the integration the authority needed for its approved task. Enforce permissions in connected systems and use appropriate approval for consequential actions rather than relying only on the model’s instructions.
Disclaimer: This news summary is intended for informational and marketing purposes only, and nothing presented here is contractually binding or necessarily the final opinion of the authors.