repo-forensics (Claude Skill) — 설치 & 라이브 데모

왜 쓰나요

핵심 기능

18 scanners covering distinct threat surfaces
Skill-specific: SKILL.md poisoning, Unicode smuggling, OpenClaw patterns
MCP-specific: tool poisoning, rug-pull enablers
Correlation engine connects findings across scanners to catch compound attacks
Zero dependencies — pure audit, safe to run on anything

라이브 데모

실제 사용 모습

repo-forensics-skill.replay ▶ 준비됨

0/0

설치

작성자 저장소 · Claude Skills 공식 문서

사용 사례

실전 활용법: repo-forensics

Audit a third-party skill before you clone it into ~/.claude/skills

👤 Anyone installing skills from strangers on GitHub ⏱ ~15 min beginner

언제 쓸까: Someone sent you a cool-looking skill link. Before clone, you want to know if it's safe.

흐름

Point at the target repo URL

Use repo-forensics. Scan https://github.com/<author>/<skill> with the skill_threats, secrets, dataflow, and lifecycle scanners. Report findings.✓ 복사됨

→ Per-scanner report with severity flags
Read correlated findings

Any correlations across scanners — e.g. a secret-leak path plus a lifecycle hook?✓ 복사됨

→ Correlation summary

결과: A go / no-go decision grounded in scanner output, not vibes.

함정

Trusting a green scan as full proof of safety — Scanners are strong floor, not a ceiling — still read SKILL.md

함께 쓰기: github · filesystem

Audit an MCP server for tool poisoning and rug-pull patterns

👤 Teams onboarding a new MCP into a production agent stack ⏱ ~20 min intermediate

언제 쓸까: Before adding a new MCP server to an agent with real permissions.

흐름

Run MCP-focused scanners

Scan this MCP server repo with mcp_security, skill_threats, integrity, and dast.✓ 복사됨

→ Findings focused on tool descriptions and install scripts
Check git history for identity spoofing

Run git_forensics on the same repo — any spoofed commits?✓ 복사됨

→ Author / signing analysis

결과: An MCP security dossier before you wire it into your agent.

Run post-incident scanners after a suspicious install

👤 Anyone who just ran npm/pip install and feels uneasy ⏱ ~15 min intermediate

언제 쓸까: Install finished with unexpected behavior; you want to know if something landed.

흐름

Run post_incident + entropy scans

Run post_incident scanner on ~/.npm and the current project. Also entropy scan the installed packages.✓ 복사됨

→ Flags for suspicious cached artifacts or encoded payloads

결과: Evidence of whether to rotate secrets and rebuild.

함정

Running scans from the potentially-compromised shell — Run from a fresh shell or another machine when in doubt

조합

다른 MCP와 조합해 10배 효율

repo-forensics-skill + github

Pull a PR's diff and scan only the added files

Fetch PR #123 and run repo-forensics just on the files it adds.✓ 복사됨

repo-forensics-skill + filesystem

Scan a locally cloned repo and save a JSON report

Scan ./suspicious-repo and write findings.json next to it.✓ 복사됨

도구

이 MCP가 노출하는 것

도구	입력	언제 호출
skill_threats	repo path	Any skill audit
mcp_security	repo path	MCP audits
secrets	repo path	Always
dataflow	repo path	Suspicious egress concerns
lifecycle	repo path	Pre-install
git_forensics	repo path	Checking maintainer identity
entropy	repo path	Obfuscation suspected
sast	repo path	General code risk

비용 및 제한

운영 비용

API 쿼터: None
호출당 토큰: Moderate — scanner output summarization can grow
금액: Free
팁: Run only the scanners relevant to the threat model (skills → skill_threats; MCP → mcp_security)

보안

권한, 시크릿, 파급범위

자격 증명 저장: None — scanner is read-only

데이터 외부 송신: Local-only by default

절대 부여 금지: Write access to the repo under audit

This is a defensive tool. Don't use scanner output to stress-test your own targets — it's not a pentest kit.

문제 해결

자주 발생하는 오류와 해결

Scanner reports 0 findings on obviously sketchy repo

Ensure you ran against the full repo path, not a subset; check scanner selection

확인: repo-forensics list-scanners

False positives flood the report

Filter by severity; focus on high/critical; use correlation to prioritize

대안

repo-forensics 다른 것과 비교

대안	언제 쓰나	단점/장점
claude-cybersecurity	Your code needs OWASP-style app security review rather than supply-chain vetting	Different focus — app logic vs install-time threats
gitleaks / trufflehog	You only need secret scanning	Single-scanner, not correlation

repo-forensics