Does discovery decrypt our traffic?

No. Identifying a tool needs the process and the destination, both visible without opening the session. Only endpoints in the deep-capture set are decrypted; every other catalog domain is logged by hostname and tunneled through encrypted.

Will it find AI tools outside the browser?

Yes. The sensor sits at the socket layer, so a desktop app, an IDE assistant and a command-line agent are all visible. That is the gap a browser extension cannot close, and it is where the highest-value context now travels.

What happens with a tool nobody has heard of?

If its destination is on the catalog it appears straight away. If it is not, it still shows as traffic from a process to an uncategorised destination, which is itself a signal, and the catalog is updated centrally rather than by you.

Shadow AI /

Find the AI nobody told you about

Asking people which AI tools they use returns an undercount every time. Discovery on the device returns the list, per machine, without reading a single prompt.

The problem

Shadow IT needed a purchase order and an admin account. Shadow AI needs a browser tab and a personal email address, which is why the gap between what an organisation believes it uses and what it actually uses is measured in dozens of tools rather than a handful.

It stays invisible for structural reasons rather than clever ones. Every AI tool is an ordinary process opening an ordinary TLS session to a reputable domain. Your EDR sees a browser. Your firewall sees an allowed destination. Neither has a reason to care, and neither is wrong.

So the inventory ends up being built by asking, which produces the tools people remember, the tools they are comfortable admitting to, and none of the ones that arrived last month.

How discovery works

01
One sensor, deployed by MDM
The agent installs from a signed package pushed through your MDM, or from a one-line install command. It runs as a transparent proxy at the socket layer, so it sees outbound connections whichever application makes them, and unsetting a proxy variable does not route around it.
02
Matched against the AI catalog
Connections are checked against a maintained catalog of AI destinations: chat apps, assistants, IDE plugins, agent endpoints, MCP servers. A hit is recorded as device, tool, timestamp and count.
03
No decryption required
Naming a tool needs the process and the destination, both visible without opening the session. Catalog domains outside the deep-capture set are tunneled through untouched: recorded, not read.
04
The list becomes a view
The dashboard answers which AI is in use, on which devices, and how often. For most organisations that view is the first honest inventory they have had.

What you get

A per-device inventory

Not a fleet-wide guess. Which laptop, which tool, when it was last seen. That is the shape of the answer you need when the question is about one team, one contractor, or one person who left on Friday.

Coverage of tools you have not heard of

The catalog ships with the agent and is maintained centrally, so a tool released after your last policy review still shows up without anyone adding it to a list.

An honest uninspected share

Certificate-pinned applications cannot be decrypted by anyone. They still appear in discovery at the domain level, and the dashboard reports how much traffic sits in that bucket rather than leaving a silent gap.

A basis for policy

An inventory is what turns a blanket ban into a decision per tool: onboard it for full inspection, leave it recorded, or block it outright.

What it does not do today

Discovery names a tool, it does not read it. Prompt-level detail requires the tool to be in the deep-capture set, which today means Anthropic endpoints plus OTLP sources such as Claude Code.
A catalog is a list, and lists lag. A brand new destination is identified once the catalog learns it, which is a release rather than a magic trick.

Frequently asked questions

Does discovery decrypt our traffic?: No. Identifying a tool needs the process and the destination, both visible without opening the session. Only endpoints in the deep-capture set are decrypted; every other catalog domain is logged by hostname and tunneled through encrypted.
Will it find AI tools outside the browser?: Yes. The sensor sits at the socket layer, so a desktop app, an IDE assistant and a command-line agent are all visible. That is the gap a browser extension cannot close, and it is where the highest-value context now travels.
What happens with a tool nobody has heard of?: If its destination is on the catalog it appears straight away. If it is not, it still shows as traffic from a process to an uncategorised destination, which is itself a signal, and the catalog is updated centrally rather than by you.

Keep reading

Stop the paste that should not have happened What to do once a tool is on the list and people are pasting into it.
AI App Risk Directory What leaks through each named tool, and what its vendor does with it.
Sighted, recorded, inspected The three fidelities, and why they are not interchangeable.

See it on your own fleet.

One sensor, deployed through your MDM, showing every AI tool in use and what is leaving the device.