Skip to content

Document a model-free preflight for the actual Deep Scan worker sandbox #2575

Open

Description

@9keysGroup

Request

Please document the supported, no-authentication/no-inference method to qualify the actual generated Deep Scan command-worker sandbox before permitting a real scan.

Target package is @openai/codex-security@0.8.3 (bundled plugin 0.1.95, Codex dependency 0.166.0). This is a setup/documentation question, not a demonstrated sandbox bypass.

Intended synthetic qualification environment

A disposable Ubuntu Server 24.04 LTS VM with:

  • A dedicated non-root scanner user and an independently enforced read-only synthetic source mount at /mnt/synthetic-source.
  • Fresh private scanner state outside source, initially containing a synthetic credential canary rather than real credentials.
  • An external default-deny gateway permitting only the parent's necessary transport.
  • No inherited host files, agents, production access or unrelated MCP connections.

The required boundary is that the scanner's command execution cannot write source, read protected credential state, or reach either the gateway proxy or a controlled TCP/UDP canary, while the parent can eventually authenticate and perform inference.

What existing checks establish

The documented CLI scan --mode deep --dry-run and SDK preflight() check local inputs without starting Codex or loading credentials. They therefore do not demonstrate enforcement under the actual generated Deep worker policy.

The generic native sandbox/tool-version startup check is also insufficient to identify and exercise that actual worker policy. A separately constructed command/exec policy would only test a lookalike configuration.

Your SECURITY.md correctly says read-only Deep workers do not themselves establish an offline environment or disable inherited MCP servers. We are not assuming otherwise.

Requested supported contract

  1. Is there a supported CLI/SDK operation that materializes or exports the exact generated Deep worker permission policy without login, a model request, or starting a scan?
  2. Can a bounded synthetic command be executed under that same policy, with verifiable policy identity, to test source write denial, credential-canary denial and command-network denial (including the parent's approved proxy)?
  3. How are parent authentication/inference networking and command-worker networking independently configured in this pinned release? Are required settings inherited by every Deep worker?
  4. What are the supported Ubuntu 24.04 AppArmor/unprivileged-user-namespace prerequisites? No host security-policy relaxation is acceptable.
  5. If this qualification cannot currently be performed through supported interfaces, please confirm that limitation and the supported escalation/version path.

Current evidence limits

No VM has been provisioned, no package installed in a guest, no login/model request/Deep Scan executed, and no actual-worker enforcement result is claimed. There is no scan ID. No private source, credential values or runtime logs are included. We will keep execution blocked rather than patch validators, invoke private helpers, mock authentication, or claim a dry run proves containment.

References:

Activity

  1. mldangelo-oai commented on Nov 15, 2026

    @mldangelo-oai
    Collaborator

    I checked the public CLI and SDK on 0.8.3 and on current main. scan --dry-run and SDK preflight() can both succeed without launching Codex. Both returned a scan plan before any worker started, so there was no generated sandbox to probe.

    I couldn’t find a supported command in either version that runs a probe under that exact policy before login or inference. The Ubuntu setup guide covers the Bubblewrap and user-namespace prerequisites. I’m keeping this issue open to track the missing worker-level check you described. Ubuntu setup guide · Preflight documentation change

  2. added
    documentationImprovements or additions to documentation
    area:sandboxExecution permissions, isolation, trusted paths, and filesystem access controls.
    on Nov 15, 2026
  3. mldangelo-oai commented on Nov 15, 2026

    @mldangelo-oai
    Collaborator

    For the networking question in 0.8.3, new and resumed native Codex workers receive the selected provider and authentication context. Their inference traffic can originate in those processes. A gateway rule that permits only the top-level scanner process can therefore prevent a real scan from working. For commands launched by a worker, the plugin extends :read-only, carries forward applicable parent filesystem denies, and sets network.enabled = false. Verifying credential-read and proxy denial still requires a test under those effective permissions. Worker launch · Worker permissions

    In the bundled Codex 0.166.0, codex sandbox --permission-profile selects a profile from the active configuration. Selecting codex_security_deep_scan_worker in a fresh Codex home exited 1 during configuration loading. The other relevant option, --sandbox-state-json, requires sandbox state the caller already has. Neither obtains the effective permissions of a future Deep Scan worker. If your deployment requires that proof before the first scan, this issue tracks the missing capability. Upstream CLI reference

  4. TheOutlandsDev commented on Nov 19, 2026

    @TheOutlandsDev
    Contributor

    I traced the Deep Scan worker launch and the existing preflight tests. The worker receives the generated codex_security_deep_scan_worker profile, and its internal preflight checks native config/read and permissionProfile/list. Existing tests cover the profile and matching preflight/worker config overrides. I found no supported path that runs a synthetic command under that profile before a scan, consistent with your comments and the merged documentation update.

    I'd like to contribute a focused test and implementation for that remaining qualification gap. Which public surface would you prefer: a separate SDK/CLI qualification operation, or a native Codex sandbox capability that Codex Security calls? Is creating a temporary native thread to obtain the effective parent permission profile acceptable under the no-auth/no-inference, no-scan contract, and can the native command probe return a verifiable policy identity? My test would use only disposable files and a local canary, with a positive command-start control. It would not treat matching config text alone as enforcement proof.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:sandboxExecution permissions, isolation, trusted paths, and filesystem access controls.documentationImprovements or additions to documentation

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions