hoahan191/agent-guard
GitHub: hoahan191/agent-guard
一款集成到 CI/CD 流水线中的 DevSecOps CLI 工具，利用 Gemini 驱动的对抗性智能体框架自动对 AI Agent 执行越狱攻击测试并评估其安全性。
Stars: 0 | Forks: 0
# AgentGuard：DevSecOps 信任层
```
flowchart TB
subgraph CI["GitHub Actions (DevSecOps Pipeline)"]
PR[Lập trình viên mở Pull Request]
GHA[GitHub Action Runner]
PR_Comment[Bot Comment cảnh báo trên PR]
end
subgraph GCP["Google Cloud Platform (Zero-Trust Security)"]
WIF{Workload Identity
Federation - OIDC} Vertex[Vertex AI
Gemini API] end subgraph AgentGuard["AgentGuard CLI Tool"] Orchestrator((Antigravity
Manager)) subgraph Agents["The Adversarial Triad"] Attacker[Attacker Agent
Red Team: Gemini 1.5 Flash] Judge[Judge Agent
Referee: Gemini 1.5 Pro] end MCP[(Jailbreak Arsenal
MCP Server)] end subgraph Target["Môi trường Staging"] TargetAPI[Target Agent API
Blue Team] end %% Flow Logic PR -->|1. Kích hoạt Workflow| GHA GHA -->|2. Trao đổi OIDC Token| WIF WIF -->|3. Trả về Token ngắn hạn| GHA GHA -->|4. Lệnh: agent-guard scan| Orchestrator Orchestrator -->|5. Khởi tạo State & Cấp quyền| Attacker Attacker <-->|6. Lấy Attack Vector| MCP Attacker <-->|Inference| Vertex Attacker -->|7. Gửi Payload Độc hại| TargetAPI TargetAPI -->|8. Trả về kết quả| Orchestrator Orchestrator -->|9. Chuyển giao Context Window| Judge Judge <-->|Inference & Suy luận sâu| Vertex Judge -->|10. Trả về Báo cáo JSON| Orchestrator Orchestrator -->|11. Quyết định Exit Code 0/1| GHA GHA -->|12. Hiển thị BLOCKED hay PASSED| PR_Comment %% Styling cho sự trực quan classDef secure fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px,color:#000; classDef danger fill:#ffebee,stroke:#c62828,stroke-width:2px,color:#000; classDef core fill:#e3f2fd,stroke:#1565c0,stroke-width:2px,color:#000; classDef gcp fill:#fff8e1,stroke:#f57f17,stroke-width:2px,color:#000; classDef neutral fill:#f5f5f5,stroke:#616161,stroke-width:2px,color:#000; class WIF,Vertex gcp; class Attacker,MCP danger; class TargetAPI secure; class Orchestrator,Judge core; class PR,GHA,PR_Comment neutral; ```
Federation - OIDC} Vertex[Vertex AI
Gemini API] end subgraph AgentGuard["AgentGuard CLI Tool"] Orchestrator((Antigravity
Manager)) subgraph Agents["The Adversarial Triad"] Attacker[Attacker Agent
Red Team: Gemini 1.5 Flash] Judge[Judge Agent
Referee: Gemini 1.5 Pro] end MCP[(Jailbreak Arsenal
MCP Server)] end subgraph Target["Môi trường Staging"] TargetAPI[Target Agent API
Blue Team] end %% Flow Logic PR -->|1. Kích hoạt Workflow| GHA GHA -->|2. Trao đổi OIDC Token| WIF WIF -->|3. Trả về Token ngắn hạn| GHA GHA -->|4. Lệnh: agent-guard scan| Orchestrator Orchestrator -->|5. Khởi tạo State & Cấp quyền| Attacker Attacker <-->|6. Lấy Attack Vector| MCP Attacker <-->|Inference| Vertex Attacker -->|7. Gửi Payload Độc hại| TargetAPI TargetAPI -->|8. Trả về kết quả| Orchestrator Orchestrator -->|9. Chuyển giao Context Window| Judge Judge <-->|Inference & Suy luận sâu| Vertex Judge -->|10. Trả về Báo cáo JSON| Orchestrator Orchestrator -->|11. Quyết định Exit Code 0/1| GHA GHA -->|12. Hiển thị BLOCKED hay PASSED| PR_Comment %% Styling cho sự trực quan classDef secure fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px,color:#000; classDef danger fill:#ffebee,stroke:#c62828,stroke-width:2px,color:#000; classDef core fill:#e3f2fd,stroke:#1565c0,stroke-width:2px,color:#000; classDef gcp fill:#fff8e1,stroke:#f57f17,stroke-width:2px,color:#000; classDef neutral fill:#f5f5f5,stroke:#616161,stroke-width:2px,color:#000; class WIF,Vertex gcp; class Attacker,MCP danger; class TargetAPI secure; class Orchestrator,Judge core; class PR,GHA,PR_Comment neutral; ```
标签：AI智能体, DevSecOps, DLL 劫持, LNA, 上游代理, 人工智能安全, 合规性, 大语言模型