Shadow-AI discovery

Agents arrive the way shadow SaaS did: a framework wired into a service, an MCP server in an editor config, a local model on a workstation. None of it waits for review, so assume your inventory is wrong.

Conventional asset inventories list installed software and devices. They do not say which code projects are building agents, which MCP servers are configured, or which projects already carry a governance harness, and that last item is the one a rollout plan actually needs.

The visiq-discovery sensor detects agent frameworks through installed npm and pip packages, LangChain, LlamaIndex, AutoGPT, and CrewAI among them, plus MCP servers, local model runtimes such as Ollama, and tool-routing services like OpenClaw. Anything not registered with the platform appears as a finding.

The scan is passive and read-only: it observes installed packages and running services, modifies nothing, and captures no traffic. Secret-shaped findings are fingerprinted, never collected in plaintext, and a scan cut short by its own budgets is reported as incomplete, never folded into a clean result.

Each code project that declares an agent framework carries a three-state badge: ungoverned, harnessed, or governed. A harness that is merely declared in a manifest still counts as ungoverned, because a dependency line is not protection.

An ungoverned agentic project opens a finding, and it auto-resolves only on a clean scan that verifies the project is harnessed. A truncated walk never closes it, since could not see is never the same as safe.

A Deployed Sensor installs persistently as a managed application through your device-management platform's app catalog, self-heals, and scans on the cadence you set. An On-Demand Scan runs the same binary once through the platform's own execution channel, Intune Remediations for example, and leaves nothing installed.

Both paths run the identical binary, so findings are identical everywhere. The delivery script verifies the sensor's SHA-256 against the published digest and refuses to run a binary that does not match.

Each finding can be registered, acknowledged as intentional, or flagged for investigation. Registering does not start a second deployment: an instrumented agent self-registers on first contact, begins in monitor mode, and its policy coverage grows from its own observed traffic.

The home dashboard's Open Shadow Cases KPI counts findings still unresolved over the last 90 days, so progress is a number rather than an impression.