Why shadow AI discovery has to reach browsers and endpoints

Shadow AI is AI that people use without review. A coding assistant installed on a laptop, a browser extension that summarizes pages, and a model running on a workstation can each help someone finish a task, and each can stay invisible to the tools a security team already runs.

Many programs begin shadow AI discovery by watching the network or routing traffic through an AI gateway. That gives solid visibility into the traffic that reaches the checkpoint. A good deal of the AI that people adopt on their own never goes there.

What network and gateway views can see

A gateway sees the requests that pass through it. Proxy logs and DNS records can show that a device reached a well-known AI domain. Those are useful signals, and they depend on traffic crossing a wire that someone watches.

Several common AI surfaces produce little or no traffic of that kind. The list below shows where a network-only view runs out.

• Local runtimes. A model running on a laptop through an Ollama-style runtime can answer questions with no outbound request at all.
• IDE and desktop tools. Plug-ins and developer AI tools often call providers directly, sometimes from home networks or hotspots that corporate monitoring does not reach.
• Browser extensions. An extension reads and changes pages inside the browser, so its activity looks like ordinary browsing to the network.
• Personal accounts on approved sites. The domain may be allowed even when the signed-in account is a personal one.
• Files at rest. Startup items, add-ins, and AI-related configuration files sit on disk long before they generate any traffic.

Discovery that runs where the AI runs can see what a gateway never receives.

A short example

For example, an engineer might install a local runtime and an IDE plug-in to try a code assistant. The model runs on the laptop, and the plug-in connects to a provider from a home network. Nothing crosses the company gateway.

Endpoint discovery on that laptop can record the runtime, the plug-in, and the evidence for each, including paths, hashes, and permissions. The security team then has something concrete to review, even though no network log mentions either one.

Put discovery where the AI runs

Discovery on the device and in the browser looks at the source. It can see what is installed, what is running, and which AI sites and extensions people use, whether or not any traffic reaches a gateway.

Halo splits that work across two components.

• Halo Browser. An extension for Chrome and Edge, plus Safari on macOS. It discovers use of supported AI websites, AI browser extensions, and account posture signals.
• Halo Endpoint. A component for macOS and Windows. It discovers AI applications, browsers, IDE plug-ins and developer AI tools, local AI runtimes, model artifact locations, startup items, add-ins, and AI-related configuration files.

Each one adds a view that a gateway lacks, and the gateway keeps its own view of the traffic that reaches it.

Start with metadata

Halo can record what is on a device without reading what people do with it. Halo Endpoint is metadata-first. It records paths, hashes, and permissions as evidence that a tool exists and what it can reach.

Standard collection does not include prompts, responses, documents, clipboard content, or credentials.

Browser discovery follows the same boundary. In the standard configuration, Halo Browser records which supported AI sites and extensions are in use. Content inspection requires expanded permissions that your organization approves, and it is off in the standard configuration.

These boundaries help adoption. People accept a tool more readily when it says what it collects and keeps content out by default. Privacy reviewers can read Halo's browser privacy and endpoint privacy pages before a rollout starts.

Turn findings into an inventory

A list of detections helps only when each item gets a status. Halo's AI inventory classifies items as sanctioned, monitored, blocked, shadow, or unknown, and records the source and confidence behind each one.

Expect a long list of unknown items after the first scan. Work through it in this order.

1. Assign an owner to the items that appear on many devices or touch sensitive teams.
2. Decide a classification. Approve the tools people need, monitor the ones that need a closer look, and block the ones your policy rules out.
3. Write down why. A short reason on each decision saves a second review later.
4. Export the result. The inventory exports as an AI bill of materials in CycloneDX, JSON, and CSV, so security, procurement, and audit can work from one list.

Say what the coverage is

Discovery has boundaries, and stating them builds trust with the teams that rely on the results.

Halo Browser works in Chrome and Edge, plus Safari on macOS. Halo Endpoint works on macOS and Windows. A device without the endpoint component, or a browser outside that list, sits outside what Halo can see.

Track coverage next to the counts. Record which devices and browsers report, how recently each one reported, and which surfaces remain out of scope. Endpoint discovery depends on known locations and signatures, so some apps, model files, or configurations may go unfound.

The post on inventory coverage and freshness goes deeper on keeping those numbers honest.

Begin with one group

Start with a team where local runtimes and developer tools are most likely, such as engineering. Run discovery without blocking anything. Review what turns up, classify it, and share the results with the team.

Browser and endpoint Discover rollouts are guided, so the Halo team works with you to plan the pilot.

Policy comes after you know what exists. Once the inventory looks trustworthy, you can set browser policy in audit, warn, or block mode and apply gateway controls to the traffic that does route through Halo. For the browser side of that work, read Browser AI needs its own governance model.

How Halo helps

Halo Browser and Halo Endpoint feed one AI inventory with classifications, source, and confidence, and the inventory exports as an AI BOM in CycloneDX, JSON, and CSV. Rollout is guided, so the Halo team works with you to plan the pilot. To see how discovery would work for your teams, book a walkthrough.