Codex Red-Team Opt-In Mode

中文

Current version: v2.1.0

Normal mode remains the default. The durable red-team operation runtime starts only after explicit activation, while the system-layer base instructions and active model profile remain loaded in every mode.

An opt-in, durable red-team runtime for Codex App and Codex CLI. It compiles each objective into a criteria-bearing GoalContract, executes typed WorkflowSpec graphs through live MCP tools, preserves verified evidence, and reaches successful terminal state only after TerminalJudge proves every criterion.

Usage Instructions

Before using this project, you should read the README file completely to prevent any issues during use. If you're a beginner, please read the GUIDE file completely to avoid any unnecessary trouble while using it.

Models and System Prompts

The installer composes the user's prior instructions, instruction.ctf.md, and the matching Jailbreak.gpt-5.x.md through the system-layer model_instructions_file. Codex App tasks select the current profile from Hook model metadata; the CLI wrapper builds a single-profile system file before process startup and pins the model family.

Asking Questions Wisely

1. If you have questions about the content in the README, you should submit an ISSUE, including the corresponding content from the README to seek help, or ask in the QQ group in the same way.

   Do:

   [Paste the text from the README], I have a question here, how/why/what should I do
   

   Don't:

   Why is it different when I open it compared to the video?
   

2. If you have suggestions or ideas for maintaining the project, you can submit an ISSUE, submit a Pull Request, or suggest in the QQ group.

   Do:

   I have a suggestion regarding xxx [image], here we can modify/add/delete xxx
   

   Don't:

   Can we add a plugin/feature for cracking keys?
   

QQ Group

805865662

Operating Boundary

The runtime treats task targets as operation assets whose scope was resolved outside the model, or as local sandbox fixtures. It does not repeat scope-document requests. Irreversible file replacement and external publication still retain explicit engineering confirmation to prevent accidental side effects.

Why This Project

Motivation

Codex can use many security tools, but long-running workflows are vulnerable to session interruption, tool variance, manual result relay, and false completion based on a tool success flag. This project unifies objectives, actions, evidence, recovery, and terminal judgment in one durable execution path:

GoalContract
  -> WorkflowSpec
  -> Durable Scheduler
  -> ToolBroker
  -> SemanticVerifier
  -> EvidenceGraph
  -> TerminalJudge

Version 2.1.0 uses the unified generic-adaptive runtime without the former phase -> router -> pack -> leaf runtime, regex domain routers, Markdown exit gates, or a second Automation state machine. Domain names act only as asset and technique metadata rather than control-plane branches.

Core Features

• Explicit activation: normal is always the default, and red-team operations start only after a mode command.
• Unified adaptive execution: one versioned generic-adaptive workflow handles every domain; scenes, assets, and techniques remain metadata and dynamic hypotheses instead of duplicated control planes.
• Two-layer Prompt Rewrite: the default Hook builds a local lossless Research Brief; an optional loopback Proxy performs a real pre-model rewrite through the current model or a custom relay.
• Live tool collaboration: tools are discovered through stdio or Streamable HTTP MCP; Host-only capabilities execute against an output contract and submit observations without user relay.
• Durability and batch autonomy: SQLite WAL, events, leases, idempotency keys, and independent run IDs support recovery, concurrency control, multi-target batches, and cancellation cleanup.
• Evidence-driven completion: only semantically verified results with target binding, parent lineage, reproduction, impact, coverage, and rollback proof can satisfy TerminalJudge.

Installation

Install Commands

From the repository root, install dependencies and deploy to the current user Codex Home:

python -m pip install -r requirements.txt
python scripts/install.py

The installer detects the model from the environment, target config, or existing manifest; --model is normally unnecessary.

Project-local installation:

python scripts/install.py --project-home PATH

Custom Skill and durable operation roots:

python scripts/install.py --project-home PATH --agents-home AGENTS_PATH --enable-custom-skill-dirs --log-root OPERATION_PATH

Options

# Preview a project-local installation
python scripts/install.py --project-home PATH --dry-run

# Uninstall a user-level deployment
python scripts/install.py --uninstall

# Uninstall a project-local deployment
python scripts/install.py --project-home PATH --uninstall

Runtime State Locations

The default operation root is:

$CODEX_HOME/redteam-mode/operations/
├── runtime.sqlite3
└── artifacts/<run-id>/*.json

Hook session state is stored under $CODEX_HOME/redteam-mode/state/sessions. Use --log-root OPERATION_PATH to relocate operation data. Status responses inline no more than 64 KiB of evidence; larger verified payloads are fetched with redteam_evidence, and each evidence payload is capped at 4 MiB.

Upgrade and uninstall preserve runtime session, memory, and operation data for recovery, auditing, or manual cleanup.

What the Installer Does

• Preserves existing user configuration and manages only installer-owned fields and files.
• Merges config.toml, Hooks, the system-layer model profile catalog, Runtime MCP server, workflows, and the single boundary Skill.
• Deploys rewrite_proxy.py and cross-platform launchers; model_provider changes only when Proxy mode is explicitly enabled.
• Deploys a candidate through a pending transaction and commits the manifest only after operational validation passes.
• Preflights invalid manifests, TOML, Hook config, and out-of-scope managed paths before upgrade cleanup.
• Removes only unchanged managed content during uninstall while preserving user-modified config, prompts, Hooks, and AGENTS.md content.

Quick Start

1. Complete installation and restart Codex App or Codex CLI.
2. Start a new task and submit redteam on, redteam light, or redteam full as a standalone prompt.
3. Submit the complete objective. The Hook compiles it into a GoalContract, with independent operations for single or multiple targets.
4. redteam_run starts or resumes execution and uses ToolBroker to invoke available MCP tools.
5. Host-only tools receive a next_action_spec, gate, exit condition, and output contract, then automatically submit their observation to the Runtime.
6. TerminalJudge emits the final result after proving every target criterion, evidence lineage, and cleanup state.
7. Submit redteam off or disable red team mode to return to normal mode.

Prompt Rewrite

Default context mode leaves the Provider unchanged. Only after explicit red-team activation, the Hook resolves the requested action before the technical object, keeps ordinary development out of the red-team dispatcher, inherits short follow-up context, and emits Clause hashes, taxonomy, risk level, evidence requirements, and a reproducible validation plan. It does not force remediation output unless the user requested it.

Enable real pre-request rewriting with:

python scripts/install.py --enable-rewrite-proxy

The request path is:

Codex App/CLI
  -> 127.0.0.1 Rewrite Proxy
  -> Rewrite Provider (rewrite rules + current raw Prompt only)
  -> four locally injected research-history messages (only when Hook red-team context is present)
  -> original Codex Provider (original request envelope with the rewritten user Prompt, system instructions, and tool schemas)

provider = "inherit" and model = "inherit" reuse the original Provider and current Codex model. For a separate relay, add a model_providers entry and set [redteam.prompt_rewrite] provider, model, and api_key_env. Normal-mode requests pass through unchanged. If the relay drops a URL, IP, CVE, path, CLI flag, Clause, high-risk marker, context bundle, or semantic-fidelity flag, the Proxy uses the deterministic local rewrite. Uninstall restores the previous model_provider.

The optional CLI wrapper pins one model-family profile for the current process:

codex-redteam --model gpt-5.6-sol

Changing model families requires a new App task or wrapper process so that model_instructions_file is reloaded.

Modes

Validation

# Install test dependencies
python -m pip install -r requirements-dev.txt

# Run the complete test suite
python -m pytest -q

# Validate an installation rooted at the current directory
python scripts/validate.py --codex-home .

Installer and test coverage includes:

• Starting the Runtime MCP server and validating the single generic-adaptive workflow.
• Responses/Chat Completions Proxy routing, explicit-mode gating, ordinary-development pass-through, relay isolation, action/risk fidelity, anchor fidelity, and deterministic fallback.
• Config merge, transactional upgrade, App/CLI Hooks, model profiles, and uninstall preservation.
• Concurrent start/recovery, leases, idempotent results, batch cancellation, and cleanup status.
• Rejection of false terminal states, false evidence, incorrect target binding, and derived evidence without trusted parents.
• Live stdio/Streamable HTTP MCP, Host Agent handoff, secret redaction, and workflow drift protection.

GitHub Actions runs the full suite on Windows, Ubuntu, and macOS with Python 3.11.

Contributions & Acknowledgements

Known Limitations

• Execution capability depends on the MCP or Host Agent tools currently available; missing capabilities remain as a pending handoff or deferred action.
• MCP tools whose names, descriptions, and input schemas do not expose their capabilities require explicit automation.tool_capabilities entries.
• redteam-light and redteam-full share the same engine and terminal rules in v2.1.0; the mode label does not select a separate workflow policy.
• Rewrite Proxy is an explicitly enabled Provider-level feature; default context mode remains red-team gated and starts no resident service.
• Each evidence payload is capped at 4 MiB, while normal status output inlines only 64 KiB; larger evidence must be fetched through redteam_evidence.
• New tasks reset to normal mode, and changing model families requires a new task or wrapper process.

Contributions and Acknowledgements

Individual Contributors

• Mingxi / 洺熙 — suggested semantic judgment as a phase-detection fallback and proposed removing the methodology layer while subdividing Skills to improve Agent behavior.
• Nirvana — proposed workflow optimization and overlay installation support.
• PINGS — contributed jailbreak text enhancements and prompt-chain robustness improvements.

Reference Projects

The earlier routing/Skill design and current runtime research drew from:

• qiushi-skill
• yaklang/hack-skills
• Anthropic-Cybersecurity-Skills
• gpt-5.6-instruct

Star History

Contributing

See CONTRIBUTING.md for the contribution workflow, coding conventions, and submission requirements.

License

MIT