Engineering governance · 10 minute read

Codex Local moves part of the boundary onto your workstation.

OpenAI’s current guidance says eligible Codex Local clients can include the CLI, IDE extension, and desktop app. Your organization remains responsible for the local execution environment, repositories, local retention, tools, and third-party services.

Source review: Independent checklist

Control checklist

Assign an owner and a stop condition.

01 · Local environment

Endpoint and workstation

Managed device posture, disk encryption, patching, screen lock, local storage, clipboard, downloads, backups, and remote access.

Responsible owner
Endpoint engineering
Evidence to collect
Device policy, compliance report, local-retention settings, and exception process
Stop condition
Unmanaged device or unbounded local persistence

02 · Authentication

Identity and access

Named user, eligible workspace or API organization, least privilege, joiner-mover-leaver process, and periodic access review.

Responsible owner
Identity and workspace administration
Evidence to collect
Group mapping, authentication policy, access review, and approved role inventory
Stop condition
Shared account, unclear organization, or unreviewed privileged role

03 · Connectivity

Network and egress

Approved destinations, proxies, TLS inspection, remote connections, Browser Use, Computer Use, and data movement outside the managed boundary.

Responsible owner
Network and security engineering
Evidence to collect
Egress policy, destination inventory, blocked-path test, and exception approval
Stop condition
Unknown destination or uncontrolled remote execution path

04 · Working set

Repositories and local files

Approved repositories, branch controls, test fixtures, generated output, terminal history, temporary files, and exclusion of production PHI.

Responsible owner
Application engineering
Evidence to collect
Repository allowlist, sanitized fixture policy, ignore rules, and code-review evidence
Stop condition
Production export, patient record, or unapproved repository enters context

05 · Credentials

Secrets

API keys, tokens, environment variables, secret scanning, command history, output redaction, rotation, and incident handling.

Responsible owner
Platform security
Evidence to collect
Secret-store integration, scanning results, redaction test, and rotation runbook
Stop condition
Long-lived secret can be read, pasted, logged, or committed

06 · Oversight

Logging and audit

OpenAI audit records, local command and file activity, access events, retention, review cadence, and content-minimizing diagnostics.

Responsible owner
Security operations
Evidence to collect
Event inventory, retention map, review procedure, and detection test
Stop condition
Activity cannot be reconstructed or logs capture unnecessary PHI

07 · Extensions

MCP and third-party services

Local and remote MCP servers, GitHub, Google Drive, browser destinations, plugins, skills, and every service Codex can access by instruction.

Responsible owner
Application owner and vendor risk
Evidence to collect
Tool allowlist, data-flow review, vendor agreement, permission test, and revocation path
Stop condition
Unapproved tool receives PHI or can mutate production without control

Rollout sequence

Prove the boundary before broad access.

  1. Inventory.Authentication, devices, repositories, tools, services, data classes, and owners.
  2. Configure.Managed settings, permissions, egress, retention, logging, and stop conditions.
  3. Exercise.Use synthetic fixtures to test denied paths, secret handling, logs, review, and incident response.
  4. Pilot.Start with named users and one bounded workflow; review evidence before expansion.

Official sources

Use the current OpenAI configuration guidance.

Apply the checklist

Map the workstation and tool boundary.

Use sanitized facts. Do not send sensitive information.

Review the Architecture Sprint