Agentbeamdocs

Start here

MDM / Getting startedDevice managementOne place to understand your Apple and Windows fleet, deliver configuration, and request device actions.MDM / Getting startedSet up your workspaceRequest MDM access, check platform readiness, and get your first managed device into inventory.Agent CLI / Getting startedAgent CLI overviewUnderstand agent activity locally, scan the files your assistants use, and connect shared workspace controls when needed.Agent CLI / Getting startedCLI quickstartInstall Beam, capture a first event, and open your local activity dashboard.Agent CLI / Getting startedConnect a workspacePair your machine through the browser or an enrollment code, then verify identity and policy synchronization.
↑ ↓ to navigate · Enter to openEsc to close
Website Open dashboard
Agent CLI/Browser extension
Browser extension

Review, redact & allow once

Handle a paused submission, inspect edited content, and understand what the local action history records.

On this page

Beam Sentinel checks supported composer submissions for heuristic matches. Treat a pause as a request to inspect the content and destination. The extension cannot decide whether every confidential statement belongs in a particular AI conversation.

Review a paused action

When a high or critical match is found, the extension presents Action paused with a redacted preview and the matching finding context. Read the message before choosing:

  • Keep blocked to leave the submission paused and return to editing.
  • Allow once to allow the matching content through the immediate retry flow.

The allow-once exception is tied to a content fingerprint and a short five-second window. It is not a permanent domain allowlist or an instruction to disable protection. Changing the content can cause another check.

Illustration of Beam Sentinel pausing an AI prompt and offering Keep blocked or Allow once.
Existing product illustration of the review flow; always inspect your actual composer and destination.

Inspect automatic redaction

Where the editor can be safely changed, recognized secret text is removed and the send remains held. Review the modified composer before sending again. If the editor cannot be updated safely, the extension falls back to the pause dialog.

Illustration of a recognized secret being redacted from an AI prompt before submission.
Redaction is pattern-based and depends on the editable surface. It does not remove every kind of private information.

Email addresses and labeled private information can trigger a high-severity pause. Do not substitute real credentials into a test prompt to demonstrate detection; use clearly synthetic test material and inspect the resulting behavior locally.

Understand warnings and history

Medium matches are allowed with a short warning. Ordinary messages continue. The popup shows checked, blocked, and warning counts from retained local decisions; redaction decisions count as blocked because submission was held.

Only the latest 100 decisions are retained. Each stores redacted preview context, not the entire raw editor buffer. Clear resets that stored history; it does not erase the assistant provider's conversation or undo any action already taken.

Diagnose a missing pause

Confirm protection is on, the site matches the manifest, and Chrome grants site access. Reload the page after extension installation or update. Check the actual input surface: unsupported iframes, custom editors, programmatic sends, and site changes may bypass the recognized hooks.

If a message was already submitted, handle it through the destination provider and your incident process. Browser interception cannot retract data that has already left the composer.

Keep exploring

Explore the docs