SKILL: security-pentest-ai-redteam

AI-agent / LLM red-teaming — harnessed eval, not live attack. The security closed-loop for this harness's MAIN BATTLEFIELD (AI agent build + deploy + embed). Planning/Read-only: this skill designs the test matrix and explains how to run it; the project runs PyRIT / promptfoo (CLI) and appsec-sdk persists findings. Pairs with security-app-llm (defensive design) — that one says "build it safe", this one says "now adversarially test that it IS safe".

1. Mission

对自有 / lab / staging 的 LLM / agentic 系统做对抗式红队测试，把 security-app-llm 设计的防御实测一遍。核心坐标：

• OWASP Top 10 for Agentic Applications（ASI01-ASI10, 2025-12） —— agentic 系统的权威红队清单（独立于 LLM Top 10）
• OWASP LLM Top 10（2025, LLM01-LLM10） —— LLM 应用层
• 工具：PyRIT（微软, MIT —— 可编程多轮 Crescendo / TAP）+ promptfoo（MIT —— CI/CD 原生红队，映射 OWASP-LLM / NIST-AI-RMF / MITRE-ATLAS）