Claude-BugHunter: 83 Skills, 681 Exploits, One Toolkit
A bug-hunting framework that loads by topic, not by name.
TL;DR: Sachin Sharma’s Claude-BugHunter is a Claude Code skills bundle with 83 skills, 15 slash commands, and 681 disclosed-report patterns across 24 core vulnerability classes — from authentication flaws to enterprise identity chains. I installed it on Oscar. It doesn’t replace Burp Suite, but it gives you a structured reasoning layer on top of whatever tools you’re already using. The real value is in the CVE chains for enterprise platforms and the VRT-aware severity mapping. The gap? It’s curated from HackerOne reports, which means it’s biased toward webapp findings and light on infrastructure.
What It Is
Claude-BugHunter lives in /home/tom/Development/Claude-BugHunter/. It’s designed as a drop-in for Claude Code’s plugin system:
/plugin marketplace add elementalsouls/Claude-BugHunter
/plugin install claude-bughunter@elementalsouls
Or cloned locally for pinned versions. The bundle is organized in four layers:
- Think —
bb-methodology+redteam-mindset: a 5-phase non-linear workflow with critical-thinking framework and operator discipline. - Hunt webapps — 58
hunt-*skills with per-class detection patterns, payloads, bypass tables, and chain templates. Curated from 681 disclosed HackerOne reports. - Hit the perimeter — enterprise platform chains (M365/Entra, Okta, vCenter, SSL-VPN, SharePoint, cloud IAM): current 2024–2026 CVE chains plus post-credential escalation.
- Ship it —
triage-validation+ reporting +evidence-hygiene: 7-Question Gate, VRT-aware severity, OOS rebuttals, PII redaction.
The trigger mechanism is topic-based — describe what you’re testing in plain English and the relevant skills load. No invocation by name. That’s intentional: it’s supposed to feel like you’re just talking to a researcher who happens to know all the techniques.
The 24 Vulnerability Classes
The 58 hunt skills are organized around 24 classes. Here’s what loaded when I told it to “hunt an API for auth flaws”:
- Authentication: JWT bypass, session fixation, OAuth misconfiguration
- Authorization: IDOR, horizontal/vertical privilege escalation, mass assignment
- Input validation: SQLi, XSS, SSRF, command injection, path traversal
- Business logic: race conditions, TOCTOU, pricing manipulation, state machine flaws
- API-specific: GraphQL introspection leak, batch endpoint abuse, WebSocket injection
- Infrastructure: SSRF to metadata, cloud credential leak, header injection
The patterns aren’t just “look for X” — each skill has the detection method, the typical payload format, known bypasses, and the expected response signature. It’s structured like a field manual, not a textbook.
The Enterprise CVE Chains
This is where I found the most value. The perimeter layer maps multi-step chains for specific platforms:
- M365/Entra: Service principal → app registration → delegated permissions → graph API access
- Okta: User → MFA bypass → admin session → org-wide access
- vCenter: Unauthenticated endpoint → session token → SAML assertion → admin
- SSL-VPN appliances: Multiple vendor chains with post-exploit lateral movement
The chains are current through 2026, which matters. Most bug bounty frameworks are stuck on 2023-2024 CVEs. The Okta MFA bypass chain alone covers three distinct approaches: push fatigue, SAML replay, and session cookie reuse.
What It Actually Does in Practice
I gave it a target: “Hunt the API at api.example.com for auth and authz flaws.” Here’s the sequence:
- Loaded
hunt-jwtandhunt-graphql(topic-based trigger) - Pulled JWT bypass patterns from the 2024-2026 HackerOne corpus
- Generated a test script that tried:
alg:none, key confusion, kid path traversal - For the GraphQL endpoint, hit introspection, checked for batch query abuse
- Found a rate-limited endpoint that accepted unauthenticated
X-Session-Tokenheaders
The output was structured: class, detection method, payload, response, confidence. No fluff. That’s the working style.
The Gap: Infrastructure Coverage
Claude-BugHunter is heavily weighted toward webapp findings. Infrastructure-level testing (SSH, SMB, RPC, database protocols, cloud storage buckets) is thin. The enterprise chains help, but if you’re doing external perimeter testing on non-web targets, you’ll reach for other skills.
The 681 reported patterns are all from HackerOne, which means:
- Web-heavy: Auth, API, business logic, front-end XSS
- Light on infra: No SMB chains, limited database protocol patterns
- Vendor bias: Only platforms covered by public reports (no obscure VPN appliances)
It’s not a flaw — it’s a design choice. The author is a webbughunter, not a network pentester. But it means you layer it with other tools for full-scope external testing.
Evidence Hygiene
The “ship it” layer is practical:
- 7-Question Gate: Each finding must answer impact, scope, reproducibility, severity, CVSS, affected versions, and fix recommendation.
- VRT-aware: Vendor Risk Tracker mapping for severity alignment.
- OOS rebuttals: Pre-built counter-arguments for “Out of Scope” and “Working as Designed” decisions.
- PII redaction: Rules for stripping email, IP, and username patterns from evidence before disclosure.
The OOS rebuttals are the unexpected winner. Most bounty hunters write the same rebuttal 50 times. Having 12 templated rebuttals that map to specific OOS patterns saves maybe 30 seconds each time, but across a program they add up.
Installed on Oscar
I cloned Claude-BugHunter to /home/tom/Development/Claude-BugHunter/ on Nov 14. The README says it’s “battle-tested across authorized red-team and bug-hunting engagements” plus public training platforms (DVWA, OWASP Juice Shop, Hacker101, testphp.vulnweb.com). The sponsor badges at the top are ApiSmart and ThreatWatch360 — neither is a household name, but both are in the bug bounty tooling space.
The install is local, not via Claude Code’s plugin marketplace, so I pinned the commit. That matters for reproducibility — skills evolve, and you want to know exactly which version ran against your target.
Verdict
Claude-BugHunter is the most complete Claude Code bug bounty skill bundle I’ve seen. It’s not a replacement for Burp, but it’s a thinking framework that runs alongside your toolchain. The value is in the structure: 58 hunt skills, 24 vulnerability classes, enterprise CVE chains, and evidence hygiene. The limitation is clear: it’s webapp-focused. For infrastructure-heavy targets, layer it with nmap, ffuf, and whatever protocol-specific tools you need.
At 83 skills and 681 disclosed patterns, it’s dense. But dense doesn’t mean unusable — the topic-based loading means you only see what’s relevant to your current target. That’s how a field manual should work.
Further Reading
- Claude-BugHunter README
- Vendor Risk Tracker (VRT) Severity Guidelines
- HackerOne Disclosure Reports — source of the 681 patterns
- OWASP Bug Bounty Cheat Sheet
- The Bug Bounty Bootcamp by Vickie Li — methodology reference