Claude-BugHunter bundles 83 bug bounty skills for Claude Code built on disclosed reports
A trending MIT-licensed skill bundle turns Claude Code into a scoped bug bounty and external red-team assistant, built on 681 disclosed-report patterns, with scope and authorization checks written into the workflow.
Claude-BugHunter, a repository under the elementalsouls account on GitHub, is a skill bundle for Claude Code aimed at bug bounty hunting and external red-team work. It was created on 2026-05-05, is published under the MIT license and had 7280 stars and 1125 forks at the time of the GitHub trending snapshot, 1057 of the stars gained in the last day. The repository description counts 82 skills, while the README counts 83 skills, 15 slash commands and 681 disclosed-report patterns across 24 core vulnerability classes, 433 of them individually cited.
The bundle relies on skills that load automatically by topic. A user describes the target and what is being tested in plain English, and the matching skills load without being invoked by name.
Four layers
The README splits the bundle into four layers:
- Think: methodology and red-team mindset skills that set the workflow and operator discipline.
- Hunt webapps: 58 hunt skills curated from 681 disclosed HackerOne reports, with detection patterns, payloads, bypass tables and chain templates for each vulnerability class.
- Hit the perimeter: chains for enterprise platforms such as M365 and Entra, Okta, vCenter, SSL-VPN appliances, SharePoint and cloud IAM, covering current 2024 to 2026 CVE chains and post-credential escalation.
- Ship it: triage, validation and reporting skills, including VRT-aware severity, rebuttals for out-of-scope rulings, PII redaction and evidence hygiene.
By category, 58 skills cover web application vulnerability classes such as XSS, SQLi, SSRF, IDOR, SSTI, XXE and CSRF, 10 cover enterprise platforms, 6 cover reporting and validation, 5 cover recon and OSINT, and 4 cover methodology and mindset. The bundle also ships 15 slash commands, including /hunt, /recon and /report, and a deterministic engagement engine that maps the attack surface of a target and routes each finding to the skill that handles it. The README describes the workflow as recon, map and rank, hunt, validate and report, with scope enforced in code.
Install paths
The recommended route is the Claude Code plugin system:
/plugin marketplace add elementalsouls/Claude-BugHunter
/plugin install claude-bughunter@elementalsouls
This loads the skills and commands under a claude-bughunter namespace. A copy install from a clone puts the files into the Claude Code configuration folder and also provides the cbh command-line runner and the hunt engagement scaffolder:
git clone https://github.com/elementalsouls/Claude-BugHunter.git
cd Claude-BugHunter
bash scripts/install.sh
The skills use the same SKILL.md format that other agent harnesses load, and the install script can also copy them into the skill folders of OpenCode, OpenAI Codex CLI and Google AntiGravity, with a separate flag that wires in the Burp MCP server. According to the README, the knowledge layer ports to all of them, while the slash commands and the /hunt engine stay specific to Claude Code by design.
Scope and authorization
The bundle is limited to the external attack surface, meaning anything reachable from the internet without first compromising an internal endpoint: web apps, APIs, OAuth and JWT flows, cloud misconfiguration, recon and OSINT, and reporting for HackerOne, Bugcrowd, Intigriti and Immunefi. The README lists what is deliberately left out: internal Active Directory attacks, C2 frameworks, post-exploitation and persistence, AV and EDR evasion, iOS, hardware, RF and ICS work, and binary exploitation.
The skills are intended for assets the user owns or has written authorization to assess. A 7-Question Gate in the triage skill asks, among other things, whether the asset is in scope and whether the impact is on the accepted-impact list of the program before anything is submitted. The README states that the bundle excludes weaponizing 0-days against unauthorized targets, post-exploitation tooling, malware development and mass-targeting infrastructure.
The README also warns that Anthropic models apply real-time cyber safeguards that can block offensive security work by default, even when it is authorized, and points such users to the free, application-based Cyber Verification Program instead of rewording requests to look defensive. It adds that on Opus 5 some higher-risk cyber requests fall back to Opus 4.8 rather than being refused, which can look like a quality drop in a long agentic session.