Skip to main content
The browser extension (version 1.2.0, shipped with desktop agent 0.10.0 and later) checks a prompt on the page, at the moment it is sent, before the browser does anything with it. That closes the gaps the network path has in a browser: text a web app sends over a WebSocket, and a browser that uses QUIC to go around the agent. It runs in Chrome, Edge (and other Chromium browsers) and Firefox 121 or later. There is no Safari version. It is an addition to the desktop agent, not a replacement: install both. Which tools each one covers, per operating system, is on Supported AI tools.

What it does

On an AI site the agent monitors (chatgpt.com, claude.ai, Gemini, Copilot, Perplexity and the rest of the agent’s list), the extension holds the prompt when it is sent and asks the agent on the same machine. The agent judges it with the same policy, masking and engine as everything else it sees, and the page then:
  • sends it, if it is fine;
  • sends a masked copy, with what was found replaced by a label, if your policy masks it;
  • keeps it on the page and says why, if your policy blocks it, or if your organization holds prompts while the engine can’t be reached.
The prompt appears in Activity once: when the network path sees the same prompt a moment later, it recognises it and doesn’t check or record it again. On every other page, and on AI sites when no agent is running, the extension uses its own check: a credential (an AWS key, an API token, a GitHub token, a private key, a Slack token, a high-entropy secret) is kept on the page and the send cancelled.

What counts as sending

The extension catches the three ways a page sends a prompt: pressing Enter (without Shift), submitting a form, and clicking a send button. It does not act on paste or on file upload: a file attached to a prompt is checked by the desktop agent on the network path, as described in Known limitations. It reads nothing while you type.

Sending an uncertain prompt anyway

When the only reason to stop a prompt is a finding from one detection model, the page shows a PromptGuard warning with the reason instead of a hard stop. The person can choose Send anyway and say why, and that one prompt is sent; the reason is recorded in Activity and reported to your organization like any other decision. Or they choose Take it back and keep editing. Secrets, your organization’s own rules and prompts that several detectors agree on are still stopped outright. An organization that wants every such prompt blocked sets the managed key single_model_findings: block, and the page then shows a plain block. The same choice exists outside the browser; see Coaching and overrides.

Pairing with the agent

The agent registers a native-messaging host for each user and each browser installed when it starts, and the extension pairs through it, so no website can use the exchange. If a browser was installed after the agent started, restart the agent. The extension’s popup has an Open PromptGuard button that brings the desktop app to the front.

Rolling it out

Force-install it with your device management, beside the desktop agent, so it can’t be quietly removed:
  • Each agent release carries the Chrome and Edge package (promptguard-shadow-<version>.crx) and its update.xml. You host them and point the browser policy at your update.xml.
  • Policies for Chrome, Edge and Firefox, on macOS (configuration profiles) and Windows (a .reg file, or an Intune Remediation pair), come with the agent’s MDM templates. They also turn off Encrypted Client Hello in each browser, which the agent needs to tell AI connections apart. Ask support@promptguard.co for the templates and for the signed Firefox package.
  • On Windows, Chrome takes the package only on a PC joined to Microsoft Entra ID or Active Directory; Edge also takes it on a PC enrolled with Enroll only in device management. See Enrolling a Windows PC in Intune.

Install the desktop agent

The extension asks the agent for its verdicts. Without one, it only stops credentials.

Next steps

Fleet enrollment

Roll Shadow AI out to every device

Customise policies

Tune checks, warn instead of block, add your own rules

Known limitations

What Shadow AI does not cover
Last modified on November 16, 2026