Shadow AI Discovery
Shadow AI Discovery finds the AI tools your employees use that IT never approved — ChatGPT, Grammarly, Notion AI, Copilot, and hundreds more — classifies each one, and lets you enforce policy on unsanctioned tools directly in the browser. It appears under Security → Shadow AI.
The problem
Employees adopt AI tools faster than security teams can review them. A marketing hire pastes customer data into a free ChatGPT account; an engineer connects a browser AI assistant with broad OAuth scopes. None of it flows through a sanctioned gateway, so it never appears in your audit log. Shadow AI Discovery closes that visibility gap.
The AI tool catalogue
Everything keys off a single catalogue of known AI tools maintained by Nirmata. Each entry records the tool's domains, vendor, category, a risk profile, and a default classification:
| Classification | Meaning |
|---|---|
| Approved | Sanctioned for use; no action needed |
| Under review | Detected but not yet decided |
| Restricted | Allowed with conditions (e.g. no sensitive data) |
| Blocked | Not permitted |
On the Catalogue tab, admins add custom tools, override a Nirmata classification, and change the classification of any tool. Your custom entries merge on top of the bundled catalogue.
Changing a classification affects how already-detected usage is labelled and whether enforcement applies. It does not by itself widen what gets detected — detecting a brand-new domain requires a catalogue update.
How tools are detected
AIControls detects usage through three complementary layers. You can run any combination:
- Layer 1 — SSO grants. Reads which AI apps employees have connected through your identity provider, including the OAuth scopes they were granted.
- Layer 2 — Managed browser extension. A Chrome (Manifest V3) extension, distributed via MDM, that reports usage metadata — never page content when an employee visits a catalogue AI domain. This is the highest-fidelity signal for browser-based tools. See Set Up the Browser Extension.
- Layer 3 — Network logs. Correlates AI-tool traffic seen by the HTTPS intercept proxy.
Detected usage rolls up on the Usage tab: per tool, the number of distinct users, event count, first/last seen, and approximate active time. Tools seen but absent from the catalogue are flagged as unknown so you can classify them.
Privacy
The browser extension is deliberately narrow. It records the hostname, timing, and employee identity of visits to known AI domains — never page text, form input, or prompt/response content. Its host permissions are scoped to the catalogue; the one content script it injects renders the enforcement banner only and reads nothing from the page. Employee disclosure in your acceptable-use policy is expected before deployment.
Enforcement
Each tool has an enforcement mode, set on the Usage tab and applied by the browser extension:
| Mode | Behaviour in the browser |
|---|---|
| Notify | Report only (default) — no user-visible change |
| Warn | A dismissible banner informs the employee the tool is unsanctioned |
| Redirect | The tab is sent to an approved alternative or SSO URL |
| Block | The tab is sent to a block page |
The extension pulls the current policy periodically, so a mode change reaches employees within minutes. The mode applied to each visit is recorded alongside the usage event.