Eve library

Browser Profile Isolation

Guide

Revision 1

On this page

Implementation task: 180.C.11.09

Bellona browser automation uses Bellona-managed isolated Chrome/Chromium profiles only. Personal browser profiles remain blocked at this stage.

This document is the operator and agent-facing policy for the Browser CDP MVP. It summarizes the behavior implemented by the remote protocol schemas and the remote-host browser adapter through 180.C.11.08.

Current Policy#

  • Browser commands must use adapter:browser-local.
  • Managed browser sessions must use capability:browser-managed-session.
  • browser.launch_managed creates a new Bellona-managed profile directory.
  • browser.navigate, browser.snapshot, and browser.screenshot operate only against a launched managed browser session.
  • browser.close closes the managed browser context and removes the managed profile by default.
  • Chrome/Chromium remote debugging is bound to 127.0.0.1; CDP is not exposed directly to LAN or internet.
  • Real profile access is represented as blocked in adapter metadata, capability metadata, discovery output, and command metadata.

The allowed profile mode is isolated. Protocol browser command schemas reject real-profile assumptions and require localhost-only CDP metadata. Host capabilities advertise browser.profile.isolated; the adapter does not grant browser.profile.real.

Managed Profile Shape#

The host creates browser profiles under the Bellona-managed profile root, which defaults to the host runtime's browser profile directory. A caller may provide a different managed root for tests or controlled deployments, but the root must not be a recognized Chrome/Chromium user-data directory.

Every created profile:

  • lives under the configured managed profile root
  • uses a bellona-profile- directory prefix
  • receives a bellona-profile.json marker
  • records isolated: true
  • records personalProfileAccess: false
  • records cleanup strategy rm_recursive_managed_path

The host rejects roots and profile paths that look like real browser profile locations, including Chrome/Chromium user-data roots, Default, and Profile N directories. Cleanup also validates that the path remains inside the managed root before recursive removal.

Why Personal Profiles Stay Blocked#

Personal browser profiles are not just local files. They are an authority boundary for signed-in accounts and private browsing state. A profile can contain cookies, session tokens, local storage, autofill data, browser extensions, password-manager integration state, history, downloads, and cached page data. Exposing that profile to a remote agent would let malicious pages, prompt-injected content, or a compromised control path access state that the operator did not explicitly grant.

The current Phase 180 browser stack is intentionally read-only after launch except for managed-profile lifecycle and screenshot artifact creation. It does not yet have the full set of controls required for real profile access:

  • explicit per-profile grants and revocation
  • operator approval UX for sensitive browser state
  • cookie, storage, download, upload, extension, and password-manager policy boundaries
  • origin-scoped allowlists and prompt-injection containment
  • redaction rules for account pages and authenticated screenshots
  • durable evidence-retention and deletion policy for sensitive browser data
  • distributed locks for shared browser profile access
  • rollback and compensating-action rules for account mutations
  • manual hardware evidence on a real host profile

Until those controls exist and an ADR accepts the remaining risk, real personal profiles remain out of scope.

Audit And Timeline Expectations#

Browser command audit entries must preserve enough context for later replay and incident review without exposing personal profile contents:

  • command id, device id, adapter id, capability id, and command name
  • actor chain and participant id
  • profile mode isolated
  • localhost CDP metadata
  • current page URL after blocked-domain enforcement
  • screenshot artifact ids and evidence ids when screenshots are captured
  • redaction status and redaction hook metadata for screenshot artifacts

Gateway timeline events merge command-store, audit-log, and artifact-store events. After 180.C.11.08, browser screenshot command results expose artifact and evidence IDs in command.completed, artifact.registered, and command.result timeline entries.

Future Unlock Criteria For Real Profiles#

Real profile access requires a later Phase 180 task or ADR to explicitly change this policy. At minimum, that work must add:

  • protocol fields that distinguish real profile grants from isolated profile sessions
  • host policy preflight that denies real profile commands without an active grant
  • an approval request that names the exact profile path, profile kind, page origin, command risk, retention policy, and rollback limitations
  • host tests that prove real profile access cannot be reached through managed profile arguments, cleanup paths, command metadata spoofing, or gateway routing
  • manual hardware evidence on a controlled test profile, not an operator's personal daily-use profile

Until then, agents must launch isolated profiles, navigate within the blocked domain policy, capture artifacts with redaction metadata, and close the managed session when finished.