附录 A：安全速查表

Security Recap（安全回顾）

对于开发者实施这些概念，安全 vibe-coded 架构依赖放弃隐式信任和实施以下实践基线：

1. Sandbox the Vibe Loop（沙箱氛围循环）

Always execute dynamically generated scripts within：
• Kernel-level, network-isolated sandboxes
→ contain blast radius

Embed：
• Up-to-date Software Composition Analysis (SCA)
→ actively scan hallucinated or vulnerable dependencies
→ before code reaches production

2. Shift the Perimeter Left（左移边界）

Enforce：
• Trusted sources
• Verified internal registries

IDE level：
• Advisory first step（咨询第一步）

CI/CD pipeline：
• Strict deterministic checks at multiple points
→ intercept vulnerable or malicious agent logic
→ before deployment

3. Enforce Zero Ambient Authority（强制零环境权限）

Never grant agent a "Global Key"

Restrict access by mandating：
• Delegated user identities
• Just-In-Time (JIT) hyper-restricted tokens
→ expire moment task concludes

High-stakes actions：
• Replace blind approval buttons
• Mandatory "Vibe Diff"
→ ensure developers understand generated logic

4. Deploy Agentic SecOps（部署代理安全运维）

Continuously stress-test architecture：

Deploy：
• Virtual Red-Teaming Agents
→ inject "Adversarial Vibes"

Leverage：
• Agent Behavioural Analytics
→ monitor dynamic Runtime AgBOM

Empower：
• Green Team
→ auto-refactor vulnerabilities on the fly

5. Trace the Execution Trajectory（追踪执行轨迹）

Log：
• Agent's API calls
• Tool inputs
• Reasoning steps

Security teams must：
• Continuously monitor execution logs
→ detect unexpected behaviour

Utilize：
• Version control checkpoints
→ revert access if agent drifts from intended task

安全架构关键原则

原则	英文术语	核心要求
放弃隐式信任	Abandon Implicit Trust	原始 AI 模型不是 Agent；只有包裹在严格 harness 中才成为企业就绪 Agent
动态信任	Continuous Effective Trust	信任是连续指标，基于供应链、身份、运行时行为、上下文关联评估
上下文边界	Context-as-a-Perimeter	从静态身份边界转向动态上下文边界
零环境权限	Zero Ambient Authority	Agent 绝不继承开发者完整管理权限
JIT 降级	JIT Token Downscoping	任务专用临时凭证，任务结束立即过期
状态隔离	Stateful Quarantine	优雅撤销权限，保留记忆用于取证，不破坏连接 API

---