AgentSkillsCN

llm-redteaming

当您测试LLM应用或代理型系统时，用于检测提示注入、越狱攻击、工具滥用、数据泄露，以及评估回归问题。该工具依托garak/promptfoo与配置好的MCP服务器，将所有产出物统一存储于/pentest/。

SKILL.md
--- frontmatter
name: llm-redteaming
description: Use when testing LLM applications or agentic systems for prompt injection, jailbreaks, tool abuse, data leakage, and eval regressions. Uses garak/promptfoo and the configured MCP servers; writes all artifacts under /pentest/.
license: MIT
metadata:
  owner: mayflowergmbh/kali-ai-redteam

LLM Red Teaming Runbook

Inputs To Collect

  • •Target base URL / API endpoint(s)
  • •Auth method (key, cookie, OAuth), rate limits, and scope constraints
  • •Model/provider (if known) and tool/plugin surface (if any)

Baseline

  1. •Run /recon <target-url> and save output under /pentest/recon/.
  2. •Record request/response schema (including streaming) and any moderation behavior.

Automated Coverage

  • •garak for probe-driven scanning; save to /pentest/scans/garak-*.
  • •promptfoo for eval suites/redteam runs; save to /pentest/scans/promptfoo-*.

Manual Exploitation

  • •Use /probe for quick injections and /extract for system prompt extraction attempts.
  • •Use Playwright MCP for headless flows (login, multi-step prompt injection, indirect injection via UI content).
  • •Persist key artifacts in /pentest/prompts/ and /pentest/exploits/ (full payload + full response).

Reporting

  • •Map findings to OWASP LLM Top 10 categories.
  • •Include exact repro steps and expected/actual behavior.