AI安全检查表
用于全面评估的结构化测试协议
✅
清单指南
使用这些清单来确保彻底且合法的安全测试
Pre Engagement
3 项目在任何测试开始之前完成
Testing
3 项目在主动测试阶段应用
Reporting
3 项目Prompt Injection
10 项目Mcp Security
10 项目LLM渗透测试
10 项目Agentic Ai
10 项目清单进度
总体完成
0%
自定义清单生成器
清单提示
- 按照依赖关系存在的顺序完成项目
- 记录例外情况并说明理由
- 定期与利益相关者一起审查
- 随着标准的发展更新清单
Align each item with the AI security standards it maps to, so the checklist stays auditable.
AH
Frequently Asked Questions
What should be included in an AI pentesting checklist?
An AI pentesting checklist should cover pre-engagement planning, reconnaissance of LLM endpoints and MCP servers, prompt injection testing (direct and indirect), model extraction attempts, data poisoning verification, output handling review, API security assessment, and reporting guidelines.
How do you test prompt injection vulnerabilities?
Testing prompt injection involves crafting inputs that attempt to override system prompts, bypass content filters, or extract sensitive instructions. Testers use techniques like delimiter injection, role-playing scenarios, encoding bypasses, multi-language attacks, and indirect injection through documents or web content the LLM processes.
What are the key checks for MCP server security?
Key MCP server security checks include verifying authentication mechanisms, testing tool parameter validation, checking for resource exhaustion vulnerabilities, auditing MCP resource permissions, testing for prompt injection through tool outputs, and validating the MCP server's rate limiting and access controls.
What should an LLM pentesting checklist include?
An LLM pentesting checklist should include mapping the application's attack surface (endpoints, tools, data sources), direct and indirect prompt injection tests, system prompt extraction attempts, output sanitization verification, excessive agency checks, training data extraction probes, model denial of service tests, API authentication/authorization review, and logging/leakage inspection.
How should you document AI security assessment findings?
AI security findings should be documented with: a clear description of the vulnerability, affected AI components (LLM model, MCP tool, API endpoint), proof of concept showing impact, severity rating using CVSS adapted for AI systems, and actionable remediation steps prioritized by risk.