Shadow AI
Find the AI agents people run on their devices without the gateway, spot unrecognised ones and agents with prompts off
If you don't see this feature in your console, ask your SecureAuth contact to
enable the device_sensor feature flag for your organization.
People install coding agents and assistants that call models and tools with no rule deciding what they do: your shadow AI. A device sensor finds them by their processes, not their traffic, whether or not anyone connected them to Agent Authority.
Where to review it
You review them on the Shadow Agents tab of Inventory › Agents & NHIs.
The tab appears when Endpoint Control is on for your organization and you hold view_device_observations.
Recognised and unrecognised agents
Every agent process a sensor reports is sorted by whether Agent Authority knows what it is:
- Recognised – a catalog signature matched it, so the row names the agent, such as Claude Code or Cursor. Recognised by says how: from Its own process, or from the Endpoint it reached while running inside another app, such as an editor.
- Unrecognised – no catalog signature matched. It may be a new or home-grown agent, or a tool you haven't seen before, so it deserves a closer look.
The In catalog and Not in catalog filters split the list the same way, and the Unrecognised card counts the second group.
Agents with their approval prompts off
Most coding agents ask before they run a command or change a file. Prompts off flags an agent that a sensor saw launched with those prompts turned off on at least one device, such as with a flag that skips permission checks. An agent in that state acts on its own, so these are the first rows to follow up. Expand the row to see the exact flag on each device.
What a sensor reports for each device
Select an agent to see each device that ran it, with its owner and when it was first and last seen. For each device, the sensor also reports:
- Launched by – the app that started the agent, such as a terminal or an editor.
- Prompts off – the launch flag that turned off its approval prompts, when there was one.
- How it signed in – the name of the API key variable it used, such as
ANTHROPIC_API_KEY, never its value. Stored login means no key variable was set, so the agent most likely used a saved sign-in, and Credential unknown means the sensor couldn't tell. - Other hosts it reached – the other domains the agent connected to on that device.
What to do with what you find
- A recognised agent people rely on: connect it through the gateway, so your rules, data protection and Agent Trail cover it. See Agent Clients.
- An unrecognised agent: ask its owner what it is before you decide whether to allow it.
- Prompts off: talk to the owner, and connect the agent through the gateway so your rules decide each tool call.
- An API key on a laptop: move the agent to the AI gateway, so model calls use an agent token and your model access rules.
Next steps
You can now see which AI agents run on your devices, which ones Agent Authority recognises, and which run with their prompts off. Next, bring the ones people rely on under your rules.
- Agent Clients – connect an agent through the gateway
- Install Endpoint Control – enroll more devices so fewer agents go unseen