| Framework Home | Examples Index | Cross-Pillar Governance | AISM | NEXUS | Dashboard |
Current framework context: AI SAFE² v3.1. This example may preserve historical component versions or earlier framework references where they describe when the implementation was created. For current conformance, use the v3.1 framework and applicable profile requirements.
Framework: AI SAFE2 / AISM Level 4 (Sovereign Runtime Governance)
Version: 2.0
Source Pattern: Adapted from the Claude Code sovereign runtime example in the AI SAFE2 framework and optimized using the OpenClaw v2.0 governance model.
Audience: Individual developers, SMBs, enterprise engineering teams using OpenAI Codex CLI or the Codex desktop app.
The Claude example proves the core AI SAFE2 point: the trust boundary cannot live only inside a vendor agent. For Codex, the same principle applies, but the control surface is different.
Codex currently exposes strong sandbox and approval controls, project and user config.toml, profiles, MCP server configuration, and notification hooks. It does not provide Claude-style per-tool pre/post hooks. That means a Codex sovereign runtime must move enforcement outward into:
config.toml baselinesAGENTS.md policy filesThis package implements that pattern.
The upstream claude-code-sovereign-runtime example is strong on architecture and clear in its trust-boundary design.
What carries over directly:
What does not carry over directly:
AGENTS.md and config.toml, not CLAUDE.md plus settings.jsonapproval_policy, sandbox_mode, profiles, wrapper-enforced launch flags, and MCP governanceNet: the Claude package is the correct control philosophy, but not the correct implementation surface for Codex. This package replaces hook-driven controls with Codex-native and OS-level controls.
This package was tested against a live Windows Codex installation. That validation changed the design:
~/.codex/config.toml mutation is now optional, not the defaultPATH aloneThis is the same design lesson the OpenClaw example encodes at a higher maturity level: governance files help, but the strongest controls are the ones enforced at the execution boundary.
Pillar 1: Sanitize & Isolate -> scripts/codex-jit-wrapper.ps1 + IDENTITY.md + TOOLS.md
Pillar 2: Audit & Inventory -> scripts/audit-codex-install.ps1 + scripts/scan-dangerous-config.ps1 + EVALUATION.md
Pillar 3: Fail-Safe & Recovery -> wrapper-scoped launch controls + ci-cd/github-actions-codex-safe.yml
Pillar 4: Engage & Monitor -> monitoring/codex-notify.ps1 + monitoring/summarize-session.ps1 + HEARTBEAT guidance in EVALUATION.md
Pillar 5: Evolve & Educate -> QUICKSTART.md + ARTICLE.md + SOUL.md + AGENTS.md
| Control | File | Risk Addressed |
|---|---|---|
| Deny dangerous bypass startup | scripts/codex-jit-wrapper.ps1 |
--dangerously-bypass-approvals-and-sandbox |
| Enforce approval + sandbox baseline | scripts/codex-jit-wrapper.ps1 |
Overly broad execution |
| Narrow MCP inventory | managed-settings/config.team.toml + TOOLS.md |
Adversarial or excess MCP exposure |
| Audit Codex installs and versions | scripts/audit-codex-install.ps1 |
Unknown runtime surface |
| Scan risky config settings | scripts/scan-dangerous-config.ps1 |
Unsafe local overrides |
| Notification-based external logging | monitoring/codex-notify.ps1 |
Invisible session outcomes |
| Session summary and alert rollup | monitoring/summarize-session.ps1 |
Missing audit evidence |
| Identity anchoring | IDENTITY.md |
Identity replacement and prompt role drift |
| User data classification | USER.md |
Context-insensitive handling of sensitive user data |
| Subagent governance | SUBAGENT-POLICY.md |
ACT-4 delegation drift and privilege sprawl |
| Hardened project policy | AGENTS.md |
Prompt injection and unsafe operator behavior |
| Operating constitution | SOUL.md |
Value drift and inconsistent decision-making |
| Fail-closed CI template | ci-cd/github-actions-codex-safe.yml |
Unreviewed autonomous CI execution |
| MCP review checklist | integrations/mcp-allowlist.md |
Unvalidated external tools |
| Runtime validation checklist | EVALUATION.md |
Unverified claims about what is actually enforced |
WITHOUT this package:
User Intent -> Codex internal policies -> Your workstation / repo / network
WITH this package:
User Intent -> [Wrapper + Core Governance Files + Audits] -> Codex -> [External Notifications + Evaluation] -> Your infrastructure
^ ^
| controls you own | logs Codex cannot retroactively rewrite
Codex’s internal controls are still useful, but AI SAFE2 requires the runtime governor to live outside the model-controlled reasoning loop. For Codex today, that means startup enforcement, config governance, and external monitoring rather than tool hooks.
| Deployment Type | Supported |
|---|---|
| Codex CLI on Windows | Yes |
| Codex CLI on macOS/Linux | Yes, via bash wrapper templates |
| Codex desktop app | Partially |
| Codex in CI/CD | Yes |
| Codex with MCP servers | Yes |
| Codex multi-agent workflows | Yes, with tighter governance |
Desktop app note: the app honors project governance files and Codex config behavior, but wrapper-enforced launch controls are strongest when you launch via CLI or CI.
This pattern assumes ACT-3 by default and ACT-4 if you enable multi-agent roles or delegation.
CP.10 HEAR is required for production deployments at ACT-3 or ACT-4.CP.9 is required when Codex is allowed to delegate or orchestrate sub-agents.CP.8 gating is required if you combine broad tool access with regulated data or high-impact production workflows.| I am… | Start here |
|---|---|
| Individual developer hardening a workstation | QUICKSTART.md |
| Security reviewer validating real enforcement | EVALUATION.md |
| Team lead standardizing Codex configs | managed-settings |
| Security engineer auditing developer setups | scripts + monitoring |
| CI owner protecting non-interactive runs | ci-cd |
| MCP-heavy team | integrations |
| Reviewer looking for the rationale | ARTICLE.md |
These are platform constraints, not design omissions. The mitigation is to tighten launch controls, minimize MCP scope, keep launch-time approval/sandbox flags explicit, and use CI for high-consequence actions.
Inspired by the OpenClaw v2.0 approach, this package now includes a minimal Codex core-file standard:
This is deliberately smaller than the OpenClaw 11-file standard because Codex does not expose the same runtime surfaces. The goal is not file count parity. The goal is a tighter governance surface with real operational effect.
Cyber Strategy Institute – AI SAFE2 Framework adaptation for Codex, v2
| Examples Index | Framework Home | Cross-Pillar Governance | NEXUS | Scanner | MCP Profile |
AI SAFE² v3.1 | Cyber Strategy Institute