跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
安全

using-perseus

Use when starting a security conversation to understand the Perseus methodology

kaivyy · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add kaivyy/perseus --skill using-perseus

技能说明


name: using-perseus description: Use when starting a security conversation to understand the Perseus methodology

Using Perseus

Overview

Perseus is a rigorous, automated security assessment framework for Claude Code. It transforms the AI into a structured penetration tester.

Core Principle: Methodological rigor over ad-hoc guessing. We do not "look around"; we execute specific phases.

The Workflow

The assessment follows a structured sequence. If subagents (Task, Agent, invoke_subagent) are available in your runtime, spawn specialists in parallel. If running in a single-agent CLI or session, execute each phase directly in the current session without failing.

Phase 1: Reconnaissance (/scan)

Goal: Map the attack surface.

  • Action: Run Skill: perseus-scan (or execute scan commands directly).
  • Output: deliverables/code_analysis_deliverable.md (Target Knowledge Graph).
  • Stop Condition: Do not proceed until you know what to attack.

Phase 2: Vulnerability Analysis (Audit)

Goal: Prove potential vulnerability.

  • Action: Run Skill: perseus-audit (or execute audit commands directly).
  • Logic: Launch 5 parallel specialists (or 5 sequential passes: Injection, XSS, Auth, Authz, SSRF).
  • Method: "Negative Analysis" (Source -> Flow -> Sink -> Defense -> Verdict).
  • Output: Specialized reports in deliverables/.

Phase 3: Exploitation & Verification (Exploit)

Goal: Verify impact (False Positive Filtering).

  • Action: Run Skill: perseus-exploit (or verify findings directly).
  • Safety: Use SAFE non-destructive proofs (code-level taint proofs, mock tests, harmless canaries).
  • Output: Verified proofs in deliverables/exploitation_report.md.

Phase 4: Reporting (/report)

Goal: Communicate risk and provide actionable fixes.

  • Action: Run Skill: perseus-report.
  • Output: Final SECURITY_REPORT.md with executive summary, CVSS scoring, unified diff patches (git diff), and optional SARIF export.

Optional: Specialists (/specialist)

Goal: Run all deep-dive specialists (in parallel if subagents available, or sequentially).

  • Action: Run Skill: perseus-specialist (or /specialist).

Engagement Modes

Always select engagement mode before Phase 1. If user does not specify, default to PRODUCTION_SAFE.

ModeIntended EnvironmentVerification Style
PRODUCTION_SAFELive productionPassive analysis + minimal non-disruptive verification
STAGING_ACTIVEStaging/pre-prodTargeted active verification with throttling
LAB_FULLIsolated labFull dynamic verification for hard-to-reproduce findings
LAB_RED_TEAMDedicated security labAdversarial chain simulation with strict legal scope

Mode selection rule:

  1. If environment is unclear, assume production and use PRODUCTION_SAFE.
  2. If user requests aggressive testing, require explicit confirmation that target is staging/lab and authorized.
  3. Record chosen mode in deliverables so report consumers understand test depth.

Critical Rules

  1. No Hallucinations: Only report vulnerabilities you have verified via Audit or Exploit.
  2. Safe Mode: Never execute destructive commands (e.g., rm, DROP TABLE).
  3. Evidence-Based: Every finding must cite a specific File:Line or HTTP Request/Response.
  4. Authorization Gate: Do not run active exploit validation on production without explicit approved test window.
  5. Stability First: Keep verification low-rate and stop immediately if service degradation appears.
  6. Lab Isolation for Red-Team Mode: LAB_RED_TEAM only on isolated test environment with non-production data.

How to Start

If the user asks for a security review, pentest, or audit, ALWAYS start with:

I will use the Perseus methodology to assess this codebase.
Starting Phase 1: Reconnaissance...
[Invoking /scan]