[Agent Action Guard Challenge] AI Agent 要动你的电脑前，先过赛博华佗这一关

[JinNing6]

CyberHuaTuo v0.2.3 is live. This issue is the public entry for the first Guard design-partner round:

AI Agent 要动你的电脑前，先过赛博华佗这一关。

First round: 5 design partners, up to 3 real command strings each

We are not collecting stars or synthetic examples. We want a maximum of 15 real, redacted command strings proposed by Codex, Claude Code, MCP clients, or other coding Agents.

Do not execute a dangerous command for this test. The Guard reviews the command text only.

python -m pip install -U cyberhuatuo==0.2.3
cyberhuatuo guard --self-test --workspace-root .
cyberhuatuo guard "<exact proposed command>" --cwd . --workspace-root . --expected ASK --report guard-case.md

The report is generated locally. CyberHuaTuo prints the complete redacted preview before writing, performs no network request or public upload, and refuses implicit overwrite. Review every field and use --redact for private project or business identifiers.

Choose the correct route

• False positive: https://github.com/JinNing6/CyberHuaTuo-Plugin/issues/new?template=guard-false-positive.yml
• Ordinary false negative: https://github.com/JinNing6/CyberHuaTuo-Plugin/issues/new?template=guard-false-negative.yml
• Agent host / Hook integration gap: https://github.com/JinNing6/CyberHuaTuo-Plugin/issues/new?template=guard-integration-gap.yml
• Reliable Hook, parser, wrapper, protocol, enforcement, secret, or redaction bypass: report privately at https://github.com/JinNing6/CyberHuaTuo-Plugin/security/advisories/new

Never publish tokens, credentials, private URLs, usernames, hostnames, absolute paths, proprietary data, or private repository identifiers.

What happens after submission

A submitted report is not automatically counted as proof or promoted as a prescription.

1. We reproduce the classification safely without executing a destructive command.
2. We decide whether it is a false positive, ordinary false negative, integration gap, or expected behavior.
3. An accepted case becomes a regression test.
4. The fix and test are released with contributor credit.
5. Only then can the case become a signed prescription and Soul Ring contribution.

The weekly north-star metric is accepted external real cases entering regression tests, not views, self-tests, comments, or stars. Current accepted external case count is zero.

Decision boundary

• ALLOW: no covered destructive signature was detected.
• ASK: destructive but bounded; stop for exact human approval and verified recovery evidence.
• BLOCK: machine-wide, identity-wide, outside-workspace, dynamic, wildcarded, unresolved, disk, or database scope that cannot be bounded safely.

CyberHuaTuo Agent Action Guard is a deterministic guardrail, not an operating-system sandbox or complete reference monitor. It can enforce only on host paths that emit and honor the supported Hook protocol.

• PyPI: https://pypi.org/project/cyberhuatuo/0.2.3/
• Release: https://github.com/JinNing6/CyberHuaTuo-Plugin/releases/tag/v0.2.3
• Quickstart: https://github.com/JinNing6/CyberHuaTuo-Plugin/blob/v0.2.3/docs/agent-guard-quickstart.md
• Security policy: https://github.com/JinNing6/CyberHuaTuo-Plugin/blob/v0.2.3/SECURITY.md

[JinNing6]

22 秒零执行 Guard 演示

点击封面观看 22 秒完整视频： https://github.com/JinNing6/CyberHuaTuo-Plugin/releases/download/v0.2.2/cyberhuatuo-agent-action-guard-demo-v0.2.2.mp4

视频实际运行：

python -m pip install -U cyberhuatuo==0.2.2
cyberhuatuo guard --self-test --workspace-root .

自检会返回一组 ALLOW / ASK / BLOCK 示例。视频中的危险命令只是被检查的内置字符串，没有执行任何展示命令。

你正在使用 Codex、Claude Code、MCP 或其他 Coding Agent，可以提交一条脱敏的真实命令字符串，并附上：

• 使用的 Agent 与操作系统
• 你预期的 ALLOW / ASK / BLOCK
• 实际判定与规则 ID
• 你认为存在的误拦、漏拦或解释问题

请勿执行危险命令，也不要提交密钥、用户名、绝对路径或业务数据。

先问诊，再动刀。