CodexLoginManager handles authentication metadata on the local Mac, so reports about credential exposure, unsafe command execution, or misleading login state are treated as security-sensitive.

Security fixes target the latest revision on the default branch and the latest tagged beta or stable release. Older snapshots may receive fixes only when the change can be applied safely.

Do not open a public issue for a vulnerability that could expose credentials or private data. Submit a private report through the repository's GitHub Security Advisories page instead.

For non-sensitive security hardening suggestions, open a GitHub Issue. If you are uncertain whether a report is sensitive, use a Security Advisory.

Include only the minimum information needed to reproduce the problem:

• The affected commit, tag, or source revision.
• The macOS and Codex CLI versions.
• The Codex CLI installation path or installation method.
• Reproduction steps using dummy credentials and sanitized paths.
• The security impact and any known workaround.

The maintainer will review the report, request clarification if needed, and coordinate disclosure through the advisory. No response-time or release-time guarantee is currently offered.

Never upload, paste, attach, or commit any of the following:

• ~/.codex/auth.json, in whole or in part.
• API keys, including values beginning with sk-.
• OAuth access, refresh, or ID tokens.
• JWTs or session cookies.
• Unredacted Codex CLI output, application logs, screenshots, or crash reports.
• Private usernames, home-directory paths, account identifiers, or browser data.

Security warning: do not rely on the application's redactor as a guarantee. Review every report manually and replace sensitive values with obvious dummy text such as <redacted>.

安全警告：请勿上传 auth.json、API key、OAuth token、JWT 或未经人工检查的日志。

If a secret was exposed, revoke or rotate it through the appropriate official service before continuing the report. Removing a secret from a comment does not invalidate a credential that has already been disclosed.

The current-source Dictation Patch tab uses a built-in native engine and a structural parser, replacing beta.3's external-script interface. Unlike login diagnostics, enable/disable/recover explicitly modify the installed Codex application. The tool requires a supported build and a stopped Codex UI, validates staged code signatures, retains account/workspace checks, and supports rollback. Enabling uses local signing rather than OpenAI's release identity; permissions and some system integrations may be affected. The engine checks the official signing identity and exact ASAR/framework hashes before preparing local copies. Original backups and patched Codex bundles are not distributed. Each build has an independently derived plan and verified original backup. Unknown/ambiguous code structures are refused. The official-update path is a separate confirmed operation: it validates HTTPS endpoints, full archive length and pinned Ed25519 signature before extraction, then the OpenAI signature and advertised app version/build/architecture. It uses full archives, not deltas, normal termination only, fingerprint-pinned replacement, and local snapshots. If reapplication fails, the verified official update is retained with a warning. The embedded engine is not yet included in a published release.

CodexLoginManager:

• Reads the local ~/.codex/auth.json file to determine the authentication mode and whether an access token is present.
• Calls the official codex login status command to verify login state.
• Calls the official codex login command to start the browser-based login flow.
• Sanitizes known API key and token patterns before showing CLI output.
• Does not directly write ~/.codex/config.toml.
• Does not operate a project-owned server, analytics service, or telemetry endpoint.

The official Codex CLI and browser-based authentication flow are separate components with their own security and privacy behavior. A vulnerability in those components should also be reported through the relevant official channel.

Please allow time for investigation and remediation before publishing details. Do not test against accounts, devices, or credentials that you do not own or have explicit permission to use.