AI Discovery

AI Discovery is the review and proof surface for AI found across workstations, browsers, runtime hooks, MCP, inventories, and MDM preseed evidence.

Page Model

AI Discovery has four top-level jobs:

Area	Use it for	What belongs there
Review	Reviewing discovered AI surfaces.	Review state, risk, source, owner, evidence, affected workstations, and decisions such as approved, unsanctioned, accepted risk, or not AI.
Catalog	Explaining recognized AI surfaces.	Known providers, catalog labels, source types, and detection fingerprints.
Discovered	Proving fleet coverage.	AI surface rows, provider proof type, provider/platform proof, catalog version, freshness, covered counts, present-not-covered counts, stale check-ins, old detector catalogs, unsupported paths, and affected workstations.
Usage	Measuring activity.	Cost, token volume, sessions, providers, and tool-call volume.

Keep these concepts separate:

Review answers: what is this AI surface and what decision did we make?
Discovered answers: can AgentKeeper prove this surface is covered across the fleet?
Usage answers: how much activity, cost, token volume, or tool-call volume happened?

MDM Preseed Discovery

The universal macOS PKG reports managed discovery through its per-device credential. Package reconciliation can report a selected surface as present, absent, or preseeded before the developer launches it; only a real runtime event can make that surface active.

POST /api/v1/runtime/managed-coverage

The report can arrive before any developer launches the AI tool. This lets AgentKeeper show workstation-local Shadow AI that runtime hooks have not observed yet.

Use the provider deployment guides to enable the scheduled check-in:

Provider path	Where to enable Shadow AI Discovery
Microsoft Intune	Microsoft Intune deployment: deploy the metadata profile, universal PKG, and root Post-install token script. Use package-matched status for inventory.
Jamf	Jamf deployment: deploy the metadata profile, universal PKG, and After token script. Use the Extension Attribute and package-matched status helper for inventory.
Iru/Kandji	Iru/Kandji deployment: deploy the metadata profile and Custom App with its Post-Install token script. An Audit Script may report package-matched status but must not reinstall AgentKeeper.

Native Linux RPM discovery

The Linux Universal RPM uses the same managed-coverage API with provider Native Linux RPM. Its scanner runs only for explicitly enrolled target users and checks allowlisted binaries, package-manager records, desktop entries, and known configuration paths for Claude Code, Cursor, Windsurf, GitHub Copilot CLI, Codex, Gemini CLI, Antigravity, Cowork, and MCP Gateway.

The scanner reports surface presence/version, AgentKeeper-owned hook or plugin state, Gateway routing state, adapter/catalog version, freshness, limitation, and bounded skill/plugin inventory where the product contract supports it. It may send content or path hashes for drift. It does not upload source files, prompts, tool arguments, credentials, raw configuration, MCP commands/arguments/environment/headers/URLs, or arbitrary user paths.

Linux evidence keeps installation, configuration, and runtime activity separate:

Evidence	Meaning
`not_seen`	The current allowlisted scan did not find the optional product. This does not fail package health.
`preseeded`	AgentKeeper wrote or verified valid owned configuration. It is not proof that the vendor loaded it.
`active`	A real vendor hook or routed Gateway event reached AgentKeeper within the freshness window.
`unsupported`	The product/distribution combination has no certified control path, such as Cowork on RHEL/Rocky.
`repair_blocked_malformed_config`	AgentKeeper preserved malformed customer configuration and reported a local repair path without uploading its contents.

EDR Discovery Proof

CrowdStrike and SentinelOne Endpoint Discovery evidence appears in Discovered as provider proof. It does not need a separate Sources or Coverage sub-navigation. An EDR-discovered AI source should show as a discrete proof type alongside MDM preseed evidence, with provider, platform, freshness, affected workstation, and proof-state context.

SaaS Compliance Proof

Compliance connectors such as OpenAI Compliance API can show ChatGPT Enterprise or Codex activity as observed SaaS evidence. These rows prove that the provider exported activity for the workspace and that AgentKeeper normalized it into Activity and AI Discovery. They do not prove workstation hook coverage and they do not imply pre-execution blocking.

Proof States

Label	Meaning	Operator action
Covered	The current detector saw the surface and AgentKeeper coverage is preseeded or active.	Keep monitoring.
Present, not covered	The AI surface is present but the expected hook or managed coverage is missing.	Repair the MDM package or investigate local drift.
Limited	The surface is covered only in a scoped or workspace-limited way.	Confirm the limitation is acceptable.
Not seen	A fresh, current detector did not see the surface.	Treat as negative evidence for that detector version and check-in window.
Unknown: stale	The device has not checked in recently enough.	Refresh MDM execution or investigate device health.
Unknown: old detector	The device checked in, but the script catalog is older than the current managed IDE catalog.	Roll out the newer package.
Unsupported	The provider/platform cannot safely detect or preseed this surface.	Treat as a documented limitation, not a failure.

Where To Drill Down

Use Review to review MDM-discovered candidates such as Present, not covered.
Use Discovered to prove coverage by AI surface, provider proof type, platform, freshness, and catalog version.
Use Workstations to inspect one device and see its per-surface MDM preseed evidence.
Use the provider deployment context when many devices show stale check-ins or old detector catalogs.

Validation Checklist

Review still shows reviewable candidates only.
Discovered shows covered, present not covered, limited, not seen, stale, old detector, and unsupported proof states.
EDR-discovered AI appears as provider proof in Discovered, not as a separate sub-navigation.
Usage remains separate from coverage proof.
MDM package docs explain scheduled check-ins for Intune Windows, Intune macOS, Jamf, and Iru/Kandji.
Native Linux RPM rows show distribution, workstation, explicit target user, service/package state, catalog version, exact limitation, and provider Native Linux RPM.
Synthetic managed coverage reports appear in AI Discovery without exposing API keys, raw secrets, prompt text, or sensitive full paths.

Was this page helpful?