perseus - AI-powered security assessment SKILLS for your cod

项目说明

Perseus Security Skills for Claude Code

!Perseus Plugin

Perseus is a comprehensi of interacti assessment skills for Claude Code. It transforms Claude into an autonomous penetration testing partner for your own codebase, capable of performing e from initial reconnaissance to deep-di research and executi.

> Defensi Testing: Perseus analyzes your own code to find before attackers do. This is equi to running a security linter or static analyzer.

Features

Multi-Language Support (8 Languages)

| Language | Frameworks |
|----------|------------|
| Ja/TypeScript | Express, Fastify, Next.js, Nest.js, Hono, Bun |
| Go | Gin, Echo, Fiber, Chi |
| PHP | Laravel, Symfony, Slim, Lumen |
| Python | FastAPI, Django, Flask, Starlette |
| Rust | Actix-web, Axum, Rocket, Warp |
| Java | Spring Boot, Quarkus, Micronaut |
| Ruby | Rails, Sinatra, Grape |
| C# | ASP.NET Core, Minimal APIs |

Smart Auto-Detection

Perseus automatically detects your project's:

Language & Framework (Next.js, Django, Spring, etc.)

Database (PostgreSQL, MongoDB, Redis, etc.)

Infrastructure (Docker, Kubernetes, AWS/GCP/Azure)

CI/CD (GitHub Actions, GitLab CI, Jenkins)

AI/LLM (OpenAI, Anthropic, LangChain)

Extended Co

API Security: REST, GraphQL, WebSocket, gRPC, OAuth, Cache poisoning

Injection: SQL, NoSQL, Command, SSTI, LDAP, XPath, Log4j

Infrastructure: Docker, CI/CD, Cloud (AWS/GCP/Azure), Kubernetes

AI Security: Prompt injection, RAG security, tool use

Client-Side: React, Next.js SSR, Vue, Angular, Ser

---

Installation

Claude Code

That's it! E is automatic:

Skills and commands auto-discovered

Hooks auto-registered

Context injected on session start

Codex

OpenCode

Uninstall

---

Quick Start

---

Engagement Modes

Perseus uses explicit modes during assessment:

| Mode | En | Style |
|------|-------------|--------------------|
| PRODUCTION_SAFE | Li | Passi checks + minimal non-disrupti |
| STAGING_ACTIVE | Staging/pre-production | Acti with strict throttling |
| LAB_FULL | Isolated lab | Broad dynamic |
| LAB_RED_TEAM | Dedicated security lab | Controlled ad chain simulation with kill-switches |

Default mode is PRODUCTION_SAFE when en is unclear.

---

Core Assessment Phases

Perseus follows a structured 4-phase methodology:

Phase 1: Scan (Reconnaissance)

Maps architecture, entry points, dependencies, and attack surface.

| Command | Agents | Output |
|---------|--------|--------|
| /scan | 13 parallel agents | deli/code_analysis_deli.md |

Co:

Architecture & Tech Stack (auto-detect 8 languages)

Entry Points (API, GraphQL, WebSocket, gRPC)

Dependencies & CVEs

Hardcoded Secrets

Security Patterns (Auth, Authz)

Injection Sinks & XSS Sinks

SSRF & Data Flows

Crypto Usage

Security Headers & Config

Phase 2: Audit ( Analysis)

Deep white-box analysis using Negati Loop (Source → Flow → Sink → Defense → ).

Runs automatically after Scan during /start.

Wave 1: SQL Injection, Command Injection, XSS, Auth, Authz
Wave 2: SSRF, Template Injection, Deserialization, Path Tra, XXE
Wave 3: JWT, Crypto, Race Conditions, Business Logic

Phase 3: Exploit ()

findings with mode-aware safe Proof-of-Concept payloads.

Runs automatically after Audit during /start.

Safe Payloads Only:

SQL: SLEEP(5), AND 1=1

Command: sleep 5, whoami

XSS: alert(1), alert(document.domain)

SSTI: {{7*7}} → 49

Phase 4: Report (Executi)

Synthesize all findings into professional security report.

| Command | Output |
|---------|--------|
| /report | deli/SECURITY_REPORT.md |

Report Includes:

Executi & Risk O

Engagement mode and co

Technologies Analyzed (language, framework, infrastructure)

Exploits with PoC

Infrastructure Security (Docker, CI/CD, Cloud, K8s)

AI/LLM Security Findings

Supply Chain Summary

Language-specific Remediation Guidance

Strategic Recommendations

---

Specialist Deep-Di

Perseus provides 8 enhanced specialist skills with multi-language support. These run automatically during /start when relevant signals are detected. Use /specialist to force-run all of them.

| Skill | Co |
|-------|----------|
| api | OWASP API Top 10, GraphQL, WebSocket, OAuth, Cache, gRPC |
| injection | NoSQL, LDAP, XPath, SSTI, Command, Log4j, Expression Language |
| crypto | JWT (8 languages), Hashing, Encryption, Key Management |
| supply-chain | CVEs (8 package managers), Typosquatting, Dependency Confusion |
| file | Path Tra, Upload Bypass, XXE, Zip Slip (8 languages) |
| logic | Business Logic, Race Conditions, AI/LLM Security, Price Manipulation |
| client | React, Next.js SSR, Ser, Vue, Angular, Svelte |
| config | Docker, CI/CD, Cloud (AWS/GCP/Azure), Kubernetes |

---

Command Reference

Short Commands (Aliases)

| Command | Description |
|---------|-------------|
| /start | Full automated assessment with smart auto-detect |
| /scan | Phase 1: Reconnaissance |
| /report | Phase 4: Executi |

Specialist Command

| Command | Description |
|---------|-------------|
| /specialist | Run all specialist skills |

Full Commands

| Command | Description |
|---------|-------------|
| /perseus:start | Full automated assessment |
| /perseus:scan | Reconnaissance |
| /perseus:report | Executi |

Full Specialist Command

| Command | Description |
|---------|-------------|
| /perseus:specialist | Run all specialist skills |

---

Output Structure

After a full assessment, the deli/ directory contains:

---

Project Structure

---

Running Tests

:

Metadata files (plugin.json, manifest.json)

Core skills (6 skills)

Specialist skills (9 skills)

Short commands (4 commands)

Perseus commands (4 commands)

---

Safety & Ethics

Perseus is designed for defensi testing only:

All analysis is performed on your own codebase

Safe payloads only (no destructi)

PRODUCTION_SAFE is the default mode

Aggressi is restricted to staging/lab modes

LAB_RED_TEAM requires isolated en and non-production data

Kill-switch can stop acti with ABORTED-SAFETY

No data exfiltration

E reporting (no hallucinations)

Equi to running security linters or SAST tools

---

Troubleshooting

Hook Blocking Issue

Problem: Perseus scan/start fails with error like:

Cause: The security-guidance plugin blocks files containing security-related keywords, even in documentation.

Solution 1: Automatic (Recommended)

Restart your Claude Code session. Perseus auto-patches the security hook on session start:

Solution 2: Manual Patch

If auto-patch doesn't work, run manually:

Solution 3: Patch All Hook Locations

The security hook may exist in multiple locations (cache + marketplaces). Patch all:

Solution 4: Disable Security Hook (Temporary)

Deli Not Created

Problem: deli/ folder is empty after scan.

Cause: Hook blocked file writing (see above).

Solution: Fix the hook issue, then run /scan again.

Skills Not Found

Problem: /scan or /start says skill not found.

Solution: Run the post-install script:

This creates all necessary symlinks automatically.

Session Start Hook Not Running

Problem: Auto-patch doesn't happen on session start.

Solution: hooks.json exists and is valid:

Should contain SessionStart configuration.

---

Changelog

v2.2.2 (2026-02)

Simplified slash commands to 3 main entries plus /specialist

Renamed command files to Windows-safe filenames

Updated docs and structure to match new command set

v2.2.1 (2026-02)

Added engagement modes: PRODUCTION_SAFE, STAGING_ACTIVE, LAB_FULL, LAB_RED_TEAM

Added mode-aware and specialist safety gates

Added kill-switch behavior and ABORTED-SAFETY outcomes

Added new deli: engagement_profile.md, .md

Impro with co and context-aware risk weighting

v2.0.0 (2026-02)

Multi-Language Support: Added support for 8 languages (JS, Go, PHP, Python, Rust, Java, Ruby, C#)

Smart Auto-Detect: /start now auto-detects language, framework, and infrastructure

Infrastructure Security: Added Docker, CI/CD, Cloud (AWS/GCP/Azure), Kubernetes analysis

AI/LLM Security: Added prompt injection, RAG security, tool use

Enhanced Specialists: All 8 specialists now support multiple languages

Impro: Added infrastructure, AI, and supply chain sections

v1.0.0 (2026-01)

Initial release with core phases and specialists

---

Documentation

| Platform | Guide |
|----------|-------|
| Claude Code | docs/README.claude.md |
| Codex | docs/README.codex.md |
| OpenCode | docs/README.opencode.md |

---

License

MIT

项目信息

开发语言：Shell

Star 数：68

技术要点

该项目属于安全软件方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕AI-powered security assessment SKILLS for your codebase. Multi-language (JS, Go, Python, Rust, Java, PHP, Ruby, C#). Works with Claude Code, Codex, OpenCode, etc.展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。