AI 智能体
skill-marketplace
更新于 2026-08-04
Skill Marketplace 工程规则
Marketplace rules: signature verification and least-privilege installation
- 技术栈
- Skill Marketplace
- 主题
- 工程基线
- 条目
- 9 条
- 版本
- v0.2.0
✓ 应当这么做
- ✓ Treat model and tool output as untrusted input; isolate it with explicit delimiters
- ✓ Authorise every tool call server-side against the authenticated principal
- ✓ Keep system instructions out of user-controlled prompt regions
✕ 常见反模式
- ✕ Concatenating raw user input into instruction text
- ✕ Executing shell strings generated by a model
- ✕ Trusting client-side role claims for privilege decisions
◎ 交付前验证
- ◎ Injection cases added to the adversarial test corpus
- ◎ Tool allow-lists reviewed; new tools default to deny
- ◎ Secrets scanned in output before responding
同分类条目
42 篇AI 智能体
Vector Search 工程规则
Vector search rules: chunking strategy, recall evaluation and index rebuild path
AI 智能体
Feature Store 工程规则
Feature store rules: point-in-time correctness and train/serve parity
AI 智能体
Agent Skill Design 工程规则
Skill design rules: narrow triggers, explicit boundaries and verifiable outputs
AI 智能体
Agent Tool Calls 工程规则
Tool-call rules: server-side authorisation and deny-by-default allow-lists
AI 智能体
Agent Memory 工程规则
Memory rules: scoped retrieval, expiry and user-visible deletion
AI 智能体
RAG Retrieval 工程规则
RAG rules: citation-grade retrieval and evaluated recall before shipping