hanshuang-codex: jailbreak tool injects 'Hanshuang' prompt sets into OpenAI Codex CLI config
一个名为 Hanshuang-Codex 的 Python + PySide6 桌面应用出现在 GitHub 上，针对 GPT 5.6 和 V4flash 进行 '破甲'（防护绕过）实验。它通过将 '寒霜' 指令集注入 OpenAI Codex CLI 的 config.toml（通过 model_instructions_file），在对话管道中启用竞赛沙箱模式。该工具结合了两种指令集变体（flash 和 变体B），据称针对不同层级的模型。
- GitHub tool (48 stars) injects 'Hanshuang' jailbreak prompt sets into Codex CLI's config.toml via model_instructions_file, targeting GPT-5.6 and v4-flash
- Offers two prompt variants: Hanshuang-flash-v2 for mid-tier models and variant B-v3 (English) for top models like GPT-5.6/v4-pro with anti-refusal overrides
- Includes M3 protection layer with anti-debugging, anti-VM, and integrity checks; uninstall restores original config with no residue
- Activation requires typing '寒霜' in Codex, which replies '已成功破甲'; effect resets after system reboot
Related
BossConsole: open-source JVM-based harness for AI agents released on GitHub
BossConsole, an open-source, multi-platform operator's console for AI agents, has been released on GitHub. Built on the JVM (not Electron), it supports running Claude Code, Codex, Gemini, or OpenCode with a real browser, terminal, editor, secrets, and over 100 MCP tools. The project targets enterprises, science, and research.
BossConsole, an open-source, multi-platform operator's console for AI agents, has been released on GitHub. Built on the JVM (not Electron), it supports running Claude Code, Codex, Gemini, or OpenCode with a real browser, terminal, editor, secrets, and over 100 MCP tools. The project targets enterprises, science, and research.
OpenClaude improved coding agent released on GitHub
An improved version of OpenClaude, an open-source coding agent for the CLI, has been released on GitHub. It supports cloud APIs, gateways, and local models with the same tools and workflow, and runs on Windows, macOS, and Linux.
An improved version of OpenClaude, an open-source coding agent for the CLI, has been released on GitHub. It supports cloud APIs, gateways, and local models with the same tools and workflow, and runs on Windows, macOS, and Linux.
CORTEX // MODEL OBSERVATORY: unified mechanistic interpretability workspace
A unified workspace that integrates Python scripts, Jupyter notebooks, model hooks, exported tensors, and visualization tools into a single interface for mechanistic interpretability. It uses Codex with GPT-5.6 Sol Ultra to automate and streamline workflows, making interpretability accessible to non-coders.
A unified workspace that integrates Python scripts, Jupyter notebooks, model hooks, exported tensors, and visualization tools into a single interface for mechanistic interpretability. It uses Codex with GPT-5.6 Sol Ultra to automate and streamline workflows, making interpretability accessible to non-coders.
Developer builds Codex plugin to orchestrate Claude Code with Opus, Haiku, Sonnet
A developer created an open-source Codex plugin called codex-claude-orchestrator that lets Codex orchestrate a persistent Claude Code session, routing work to different Claude models: Opus maintains context, while Haiku and Sonnet handle bounded tasks. The tool aims to optimize model usage by assigning the strongest model only to work that truly requires it.
A developer created an open-source Codex plugin called codex-claude-orchestrator that lets Codex orchestrate a persistent Claude Code session, routing work to different Claude models: Opus maintains context, while Haiku and Sonnet handle bounded tasks. The tool aims to optimize model usage by assigning the strongest model only to work that truly requires it.
Codex 0.143.0 adds Amazon Bedrock GPT-5.6 Sol, Terra, Luna models with max reasoning effort
OpenAI's Codex release 0.143.0 enables remote plugins by default and adds first-class support for Amazon Bedrock's GPT-5.6 Sol, Terra, and Luna models, including a `max` reasoning effort parameter. The update also improves proxy routing for authentication and Responses API traffic on macOS and Windows.
OpenAI's Codex release 0.143.0 enables remote plugins by default and adds first-class support for Amazon Bedrock's GPT-5.6 Sol, Terra, and Luna models, including a `max` reasoning effort parameter. The update also improves proxy routing for authentication and Responses API traffic on macOS and Windows.