What a shadow AI inventory should contain
A shadow AI inventory needs one row per tool with the kind, owner, users, data reached, account type, source that saw it, risk, decision and review date. An owner for every entry matters most, because an unowned row cannot be approved, restricted or removed.
Facts read on .
Which fields belong in the inventory?
Nine fields do the work: tool name and vendor, kind (web app, desktop app, coding agent, extension, IDE plugin, MCP server, embedded feature, OAuth app), number and names of users, data reached, account type (corporate or personal), the source that saw it, risk, disposition and review date.
Vendors publish similar sets. Cyberhaven says it scores each app on five risk dimensions and records who uses each tool and with what data. Zscaler describes a per-app risk index with hosting, posture and retention, and usage by user and department. Aurascape lists more than 25 risk attributes per app, including account type.
Sources: Cyberhaven: AI security; Zscaler: public AI inventory, GenAI; Aurascape: discover and monitor AI; Aurascape source (aurascape-28).
Who owns each entry?
Name one accountable owner per entry: the business owner who wants the tool, not the security team that found it. Security owns the review process, the business owner owns the decision and the data.
Where users are many and no owner exists, assign the department head with the most users. Nudge Security lets admins set an approval status of Approved, Allowed, In Review or Not Permitted for each discovered agent, a pattern worth copying in any spreadsheet.
Sources: Nudge Security: AI agent discovery.
How do you record the evidence behind an entry?
Record the source and the date each entry was first and last seen. An auditor will ask how you know, and the answer is a log, an API record or an endpoint scan with a date. Keep the query or export that produced it.
Microsoft's Purview audit retention is 180 days for Audit (Standard), and records older than that are gone unless exported. Plan to export inventory snapshots on a schedule so history survives.
Sources: Auditing solutions in Microsoft Purview.
How should kinds of AI be tracked separately?
Track kinds separately because the control differs. A web app is blocked or coached at the browser or proxy, a coding agent is governed on the device, an MCP server is approved by name and version, and an OAuth app is revoked in the identity provider.
Microsoft's Defender for Endpoint support matrix lists local agents by name, including Claude Code, Cursor, Codex CLI, Gemini CLI and GitHub Copilot, and it reports each agent's configured MCP servers once discovery is available.
Sources: AI agent support matrix, Microsoft Defender for Endpoint; Discover local AI agents, Microsoft Defender for Endpoint.
How often should it be refreshed?
Refresh continuously from automated sources and review the full inventory at least quarterly. New AI tools appear weekly, and a quarterly cycle misses a tool that appears and spreads between reviews. Discovery policies that alert on a new app turn the review into a standing process.
Export a dated snapshot each quarter for the audit file.
Sources: Create cloud discovery policies, Microsoft Defender for Cloud Apps.
What should be left out of the inventory?
Leave out the prompts themselves and any personal data about what an individual typed. The inventory records tools and metadata, not content. Where a product stores conversation data, restrict access: Nudge Security says conversation data is available by API for ingestion into customer platforms and incident response, so decide who may query it before enabling it.
Keep usage counts at team level in widely shared reports. Individual names belong in the working record, where access can be limited, and not in a slide.
Add a field for the tool's own claims, with the source and date, only when a decision depends on it. A vendor statement on training, hosting or retention should be stored with its URL so the reviewer can re-read it. That keeps the register honest when a vendor page later changes.
Sources
- Cyberhaven: AI security (cyberhaven.com), read
- Zscaler: public AI inventory, GenAI (help.zscaler.com), read
- Aurascape: discover and monitor AI (aurascape.ai), read
- Nudge Security: AI agent discovery (nudgesecurity.com), read
- Auditing solutions in Microsoft Purview (learn.microsoft.com), read
- AI agent support matrix, Microsoft Defender for Endpoint (learn.microsoft.com), read
- Discover local AI agents, Microsoft Defender for Endpoint (learn.microsoft.com), read
- Create cloud discovery policies, Microsoft Defender for Cloud Apps (learn.microsoft.com), read
- Nudge Security: AI conversation monitoring (nudgesecurity.com), read