Claude-BugHunter: a Claude Code skill bundle that turns the agent into a scoped bug-hunting operator
A Claude Code skill bundle for bug hunting and external red-team work - 82 skills, 15 slash commands, 681 disclosed-report patterns curated across 24 core vulnerability classes, plus enterprise identity + infrastructure attack matrices.
At a glance
- What is it?
- Claude-BugHunter by Sachin Sharma is an MIT-licensed bundle of 83 skills, 15 slash commands and hundreds of disclosed-report patterns for authorized bug bounty and red-team work. Skills load automatically by topic, and the knowledge layer ports to five agent harnesses, though the slash commands and engagement engine stay Claude Code only.
- Who is it for?
- Install Claude-BugHunter if you do authorized bug bounty or red-team work and want Claude Code to carry the methodology, payloads and reporting discipline rather than reconstruct them each time. Use the plugin route for the fastest path to the skills, add the cbh CLI with pipx if you want the terminal runner, and stay strictly inside your scope, because this bundle is built for authorized testing and public training targets, not opportunistic use.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What Claude-BugHunter is
Claude-BugHunter is a self-contained skill bundle for the Claude Code skills system. The README's claim is direct: install it once and Claude Code stops behaving like a chatbot and starts behaving like a senior bug-hunting researcher or red-team operator that knows the techniques, chain templates, VRT mappings and platform CVE chains, and stays in scope. The counts it advertises are 83 skills, 15 slash commands and 681 disclosed-report patterns, 433 of them individually cited and auditable, across 24 core vulnerability classes.
It was built by Sachin Sharma, whose focus the README gives as bug hunting and GenAI security research, and it describes itself as battle-tested across authorized red-team and bug-hunting engagements plus public training platforms such as DVWA, OWASP Juice Shop, Hacker101 and testphp.vulnweb.com. The audience is clear from that list: security professionals doing authorized testing, not anyone looking for a shortcut around a target's permission.
The four layers of the bundle
The README organises the skills into four stacked layers, which double as a workflow. The Think layer, bb-methodology and redteam-mindset, provides a five-phase non-linear workflow, a critical-thinking framework and red-team operator discipline. The Hunt webapps layer is 58 hunt-* skills curated from the 681 disclosed HackerOne reports, giving per-class detection patterns, payloads, bypass tables and chain templates.
The perimeter layer covers enterprise platform chains: M365 and Entra, Okta, vCenter, SSL-VPN appliances, SharePoint and cloud IAM, with 2024 to 2026 CVE chains and post-credential escalation. The Ship it layer, triage-validation, reporting and evidence-hygiene, adds a seven-question gate, VRT-aware severity, out-of-scope rebuttals, PII redaction and red-team deliverables. The design principle is that all of this triggers automatically by topic: you describe what you are testing in plain English and the relevant skill loads, with no invocation by name.
Installing it: plugin versus copy install
The README gives two routes. The recommended one is the Claude Code plugin, run from inside Claude Code:
/plugin marketplace add elementalsouls/Claude-BugHunter
/plugin install claude-bughunter@elementalsoulsAll 83 skills and 15 commands load namespaced under claude-bughunter: and update when you bump the plugin version, with no files copied into ~/.claude/. The alternative is a copy install from a clone:
git clone https://github.com/elementalsouls/Claude-BugHunter.git
cd Claude-BugHunter
bash scripts/install.shOn Windows the equivalent is pwsh ./scripts/install.ps1. The README's comparison table explains the trade-off: the plugin is the fastest path to the skills and slash commands but the cbh CLI is a separate pipx install and the hunt scaffolder is clone-only, while the copy install gives you all three from the clone. The terminal-native runner installs standalone with pipx install git+https://github.com/elementalsouls/Claude-BugHunter, so plugin users can add it without a full clone.
Running across five agent harnesses
The skills are plain Agent Skills in the SKILL.md format, so they are not tied to Claude Code alone. The README says the same format is loaded by Claude Code, OpenCode, OpenAI Codex CLI, Hermes Agent and Google AntiGravity, and one command installs to all of them:
bash scripts/install.sh --all --burp-mcpEach harness has its own target directory: ~/.claude/skills/ for Claude Code, ~/.agents/skills/ for Codex CLI via --agents, ~/.hermes/skills/ with --hermes and ~/.gemini/config/skills/ with --antigravity, while OpenCode reads the Claude and agents paths. The --all flag detects installed harnesses and copies to each, and --burp-mcp wires the Burp MCP server into each.
There is an important limit stated plainly: the full knowledge layer ports to all five harnesses, but the slash commands and the /hunt engagement engine stay Claude-Code-only by design. So on Codex, Hermes or AntiGravity you get the skills' knowledge but not the command surface or the engagement scaffolder.
What a session looks like, and the caveats
The README shows an illustrative transcript in which the user names an in-scope HackerOne target and asks for recon; Claude loads web2-recon, offensive-osint and bb-methodology, enumerates subdomains with subfinder and crt.sh, finds live hosts with httpx, fingerprints the stacks and ranks the surface, flagging a GraphQL API with introspection on as the place to start. It is a good picture of the intended flow: the agent chooses and loads skills from your plain-English description and proposes the next probe.
The README is explicit that this block is an illustrative transcript, not a recording, and points to asciinema for capturing a real one. Two caveats follow. First, the numbers in the banner and the README body differ slightly, 82 versus 83 skills and the report-pattern citation counts, so treat the headline figures as approximate and check the searchable skill catalog for what is actually installed. Second, and more important, the tool is scoped for authorized work: the README repeatedly frames targets as in-scope and lists public training platforms, and the discipline layer exists precisely to keep testing inside scope and to redact sensitive data.
Licence, the cbh CLI and maintenance
The bundle is MIT licensed, with a separate LICENSE-CONTENT file, which matters here because the value is as much in the curated report patterns as in the code, and the two may carry different terms. pyproject.toml describes the packaged component precisely: cbh, a terminal-native recon, classify, triage and report runner, at version 2.1.0, requiring Python 3.9 or newer, with no required dependencies and requests as an optional extra since it falls back to urllib.
The recent releases are v2.1 on 2026-06-05, adding community skills, a CI gate and security fixes, and v2.0 on 2026-05-25, which added report curation, five missing surfaces, chains and engagements. The last push was on 2026-09-14. The repository carries a SECURITY.md, an ENGAGEMENTS.md, an eval/ directory and a self-hosted star-history chart, and the CI gate mentioned in the v2.1 notes suggests the skills are checked rather than merged blind, which is reassuring for a bundle whose whole purpose is to run security tooling against live targets.
Editorial conclusion
Install Claude-BugHunter if you do authorized bug bounty or red-team work and want Claude Code to carry the methodology, payloads and reporting discipline rather than reconstruct them each time. Use the plugin route for the fastest path to the skills, add the cbh CLI with pipx if you want the terminal runner, and stay strictly inside your scope, because this bundle is built for authorized testing and public training targets, not opportunistic use.
Frequently asked questions
What is Claude-BugHunter and what is it for?
It is a Claude Code skill bundle for authorized bug bounty and red-team work, with 83 skills, 15 slash commands and hundreds of disclosed-report patterns. Skills load automatically by topic so the agent applies the right techniques, chains and reporting discipline while staying in scope.
Do I have to invoke each skill by name?
No. The README says all skills trigger automatically by topic: you describe what you are testing in plain English and the relevant skill loads, with no invocation by name.
Can I use it with agents other than Claude Code?
The knowledge layer ports to five harnesses, Claude Code, OpenCode, OpenAI Codex CLI, Hermes Agent and Google AntiGravity, via a single install command. But the slash commands and the /hunt engagement engine remain Claude Code only.
Community notes