LATEST
AI Tools

Claude-BugHunter: 83 Skills That Turn Claude Code Into a Bug Hunter

Updated November 21, 2026
15 min read
Editor's Note: This article was human-reviewed for accuracy
Security researcher working across multiple monitors during a bug bounty hunt

Quick Answer & Key Takeaways

  • • Claude-BugHunter is a free skill bundle that turns Claude Code into a bug bounty hunter, built from 681 publicly disclosed HackerOne reports.
  • • It packs 83 skills and 15 slash commands for recon, web hunting, enterprise red-team work, and reporting, and it costs nothing beyond your Claude plan.
  • • The catch is authorization: it only works on assets you own or have written permission to test, and Anthropic's own safeguards can still refuse or downgrade the work.

Claude Code is a strong coding agent, but out of the box it knows nothing about bug bounty hunting. It has no method for recon, no checklist for a specific vulnerability class, and no idea what a HackerOne triager expects in a report. Claude-BugHunter fills that gap with procedure instead of prompts: 83 skills and 15 slash commands distilled from 681 publicly disclosed HackerOne reports, packaged as a free install for Claude Code. On November 18, 2026 it hit number one on GitHub Trending, and by November 21 it had passed 14,000 stars.

That rise is worth a closer look, because the bundle is both useful and easy to misuse. It is built for authorized security testing only, and Anthropic's own safeguards can refuse or quietly downgrade the work even when you have permission. This review covers what the bundle contains, how to install it, what it really costs, and the catches that matter before you point it at anything.

What Claude-BugHunter actually is

Claude-BugHunter is a skill bundle for the Claude Code skills system, built by security researcher Sachin Sharma (ElementalSoul). A skill, in Claude Code terms, is a folder with instructions the agent loads automatically when the topic comes up. Install the bundle once and Claude Code stops answering security questions from general knowledge. Instead it loads a specific skill for the task at hand, follows that skill's workflow, and uses its payloads, bypass tables, and chain templates.

The bundle's knowledge comes from real reports. Its 58 web-hunting skills were built from 681 disclosed HackerOne reports, and 433 of those patterns are individually cited, so you can trace a technique back to the report it came from. The project also ships an engagement engine that maps a target's attack surface and routes each finding to the skill that handles it, plus a cbh command-line runner for terminal-native use. It was tested against practice targets including OWASP Juice Shop, DVWA, Hacker101, and testphp.vulnweb.com, according to the README.

One counting note up front. The README says 83 skills. The repository description on GitHub says 82. The README's own category table adds up to 83, so this article uses 83 and flags the one-skill gap rather than pretending it does not exist.

Why it shot to number one on GitHub Trending

The repository is not new. It was created on May 5, 2026, and sat quietly for months. Then, in the third week of November, it went viral. The star counts below come from the GitHub API and from trending trackers that read the same public data.

DateStarsWhat happened
November 16, 20266,401Appears on GitHub Trending roundups with 221 stars that day
November 17, 20267,681Up 1,094 stars in 24 hours; coverage starts appearing
November 18, 202610,113Number one on GitHub Trending daily, up 1,005 stars that day
November 20, 202613,227Up 1,349 stars that day; third for the week with 6,100 gained
November 21, 202614,058GitHub API count at publication, with 1,680 forks

That is roughly 7,700 stars in seven days, which puts it among the fastest-growing security projects on GitHub that week. The timing is not a coincidence. Agent skills became the dominant GitHub trend of the autumn, and this bundle arrived as the security-flavored entry: the same packaging idea as the big coding-skill repos, applied to a field where a missed step costs a bounty. A hands-on tutorial from AI Frontier Post, published November 18, walks through a full install and a practice hunt against OWASP Juice Shop, which helped the surge along.

What is inside the bundle

The 83 skills fall into five groups, and the split tells you what the bundle is really for. Most of the weight sits in web hunting, with a smaller enterprise layer on top and a reporting layer at the end.

CategorySkillsWhat it covers
Hunt, web app vulnerability classes58XSS, SQL injection, SSRF, IDOR, file upload, JWT, OAuth, GraphQL, SSTI, XXE, CSRF, CORS, open redirect, and more
Enterprise platform attack10Microsoft 365 and Entra ID, Okta, vCenter, enterprise VPN appliances, cloud IAM, Android APK pipelines, supply-chain recon
Reporting and validation6Triage validation with a 7-Question Gate, report writing, HackerOne and Bugcrowd formats, evidence hygiene
Recon and OSINT5Subdomain enumeration, identity-fabric mapping, certificate transparency, JavaScript analysis, secret scanning
Methodology and mindset4Bug bounty method, red-team operator discipline, local toolkit setup

On top of the skills sit 15 slash commands, including /hunt, /recon, and /report. The /hunt command scaffolds an engagement folder with scope files and state tracking, so findings, evidence, and submission IDs stop living in scattered chat transcripts. The workflow runs in phases, recon, map and rank, hunt, validate, report, and it is deliberately non-linear: you loop back to recon when a finding opens a new surface.

The bundle also carries an explicit scope list. In scope: bug bounty hunting, web application penetration testing, external red-team engagements, cloud misconfiguration, and reporting. Out of scope by design: internal Active Directory attacks, command-and-control frameworks, post-exploitation and persistence, antivirus evasion, iOS and hardware work, and binary exploitation. The README is blunt that if your red team needs domain-takeover chains, this bundle will not help in those phases.

How to install it

There are two install paths, and the difference matters. The plugin route is faster and updates itself. The clone route gives you the engagement engine and the cbh CLI.

Option A, the plugin (recommended). From inside Claude Code, run these two commands:

/plugin marketplace add elementalsouls/Claude-BugHunter
/plugin install claude-bughunter@elementalsouls

All 83 skills and 15 commands load under the claude-bughunter: namespace, and they update when you bump the plugin version. No files are copied into your ~/.claude/ directory.

Option B, the clone. Clone the repository and run the installer for your system:

git clone https://github.com/elementalsouls/Claude-BugHunter.git
cd Claude-BugHunter
bash scripts/install.sh        # macOS / Linux
pwsh ./scripts/install.ps1    # Windows

This copies the skills and commands into ~/.claude/ (or %USERPROFILE%\.claude\ on Windows) and wires up the hunt engagement scaffolder. The terminal-native cbh runner installs standalone with pipx install git+https://github.com/elementalsouls/Claude-BugHunter, so plugin users can add it without a full clone.

The skills are plain SKILL.md files, so they load in more than Claude Code. One install command covers five harnesses:

HarnessWhere skills land
Claude Code~/.claude/skills/
OpenCodeReads ~/.claude/skills/ and ~/.agents/skills/
OpenAI Codex CLI~/.agents/skills/
Hermes Agent~/.hermes/skills/
Google AntiGravity~/.gemini/config/skills/

The slash commands and the /hunt engine stay Claude Code only. A HackerOne account is optional. The install asks for a HackerOne API username and token, which only let the bundled HackerOne MCP server read your private programs and your own reports. Leave both blank and every skill, command, and the MCP server's public tools still work. The AI Frontier Post tutorial reports the install takes about ten minutes.

What it actually costs

The bundle itself is free. The code is MIT licensed and the skill content is CC BY 4.0, so there is no purchase, no subscription, and no paid tier inside the project. The real cost is the agent that runs it, because the skills are instructions, not a model. You need Claude Code, and Claude Code is included in Anthropic's paid plans.

Claude planMonthly price (USD)Approx. GBPApprox. CADWhat it buys for this use
Pro$20 billed monthly, or $17 with annual billing ($200 up front)about £15.30, or £13.00 annualabout CA$28.70, or CA$24.40 annualClaude Code included, with standard usage limits shared across chat and terminal
Max$100about £76.40about CA$143.305x Pro usage, plus $100 a month in API credits
Ultra$200about £152.80about CA$286.6010x Pro weekly usage, no 5-hour limit for verified accounts, $200 in API credits

Those prices come from Anthropic's pricing page, read at publication, and the GBP and CAD figures are approximations at November 20, 2026 market rates (1 USD = 0.764 GBP, 1.433 CAD). Taxes are not included in Anthropic's listed prices. The practical point is that a Pro subscription is the realistic floor for regular use, and heavy agentic hunts will burn through Pro's usage limits fast enough that Max is the honest recommendation for anyone doing this weekly. There is also a pay-as-you-go route through API credits, which the Max and Ultra plans partly cover with their monthly credit allowances.

Two cost details are easy to miss. First, the optional Burp Suite integration assumes you already have Burp, and Burp Suite Professional is a separate paid product. Second, the recon tooling the skills call, such as subfinder, httpx, and nuclei from ProjectDiscovery, is free and open source, so the only recurring bill is the Claude plan.

The authorization rules are the whole point

This is where the bundle earns its reputation, and where most coverage stops too early. Claude-BugHunter is built for assets you own or have written authorization to assess: bug bounty programs with a defined scope, penetration test engagement letters, CTF challenges, and your own infrastructure. The README states that plainly, and the workflow enforces it in two places.

First, the triage-validation skill runs a 7-Question Gate before anything is submitted, and two of the seven questions ask whether the asset is in scope and whether the finding sits on the program's accepted-impact list. Second, the reporting skills include researcher-side hygiene, such as PII redaction in screenshots and evidence handling, that signals legitimate authorized testing to a target's fraud team. The bundle explicitly excludes weaponizing zero-days against unauthorized targets, post-exploitation tooling, malware development, and mass-targeting infrastructure. Its security policy file states the same posture.

There is also a limit that comes from Anthropic, not from the bundle. Anthropic's models apply real-time cyber safeguards that block vulnerability exploitation and offensive security tooling development by default. That means even authorized, in-scope work can hit a refusal that has nothing to do with this bundle. For legitimate dual-use work, Anthropic runs a free, application-based Cyber Verification Program that adjusts the safeguards. Mass data exfiltration and ransomware development stay prohibited and are not adjustable, even after enrollment. Do not reword an offensive engagement to look defensive to get past a classifier; the README says to enroll instead.

A related quirk is worth knowing before a long session. On Opus 5, a narrow set of higher-risk cyber requests, which Anthropic names as exploit generation, binary-based vulnerability scanning, and penetration testing, falls back to Opus 4.8 rather than being refused. You get a notice and the response is labeled with the model that answered, but in a long agentic run that is easy to scroll past, so it can look like Opus 5 quietly got worse. You can turn off automatic model switching in Settings, under Capabilities. Starting an engagement with /hunt helps too, because it states the authorized, scope-bounded frame on its first turn.

The catches you should know before installing

The scope wall has a gap. The scope rules are enforced in code only on the engine path, where a Python allowlist decides which targets the agent may be sent to. If you drive the bundle in plain English, staying in scope depends on the model remembering to check. An independent review from Fervor AI makes exactly this point and recommends starting every engagement from a scope file, putting a network allowlist under the agents, and enrolling in the Cyber Verification Program. Use the engine path when you can.

The main branch moves. The latest release tag is v2.1, from June 5, but the main branch keeps moving and the project pushed commits as recently as November 20. If you clone, pin a specific commit so a push cannot change your tooling mid-engagement. The plugin route avoids this problem because it tracks a version.

There is no independent benchmark of finding quality. The star count measures attention, not results. No independent test has published how many valid findings the bundle produces per engagement, or how it compares with a skilled human working without it. The 681-report lineage is real and auditable, but a pattern library is not a guarantee of valid reports. Treat the bundle as a force multiplier for a researcher who already knows the field, not as a replacement for one.

Eight of the 83 skills are vendored. The methodology, validation, reporting, and payload library skills come from shuvonsec/claude-bug-bounty, with attribution in the project. That is normal open source practice and it is disclosed, but it means part of the bundle's value is repackaged upstream work rather than original research.

Enterprise coverage is current, not permanent. The enterprise platform chains cover 2024 to 2026 CVEs, and the roadmap promises a quarterly refresh of the VPN CVE matrix. If that refresh slips, the enterprise skills will age faster than the web hunting skills, which are built on disclosed reports that stay relevant longer.

How it compares

Claude-BugHunter is not the only free skill bundle in this space, and the comparison clarifies what it is and is not. Cloudflare's security-audit skill, covered on this site on November 3, is also free and also turns an agent into a security auditor, but it is aimed at auditing code and configurations rather than hunting live targets. ECC, the 291,000-star toolkit covered here on October 28, optimizes the coding agent harness itself rather than adding security knowledge. OpenCode, covered the same day, is a separate coding agent that happens to load the same skill format, which is why BugHunter works there too.

ToolWhat it doesPriceAPI access
Claude-BugHunterBug bounty and external red-team skills for Claude CodeFree (MIT / CC BY 4.0); Claude plan requiredNo public API; HackerOne MCP optional
Cloudflare security-audit skillFree AI security audits of code and configsFreeNo public API
ECCToolkit that optimizes the Claude Code harnessFree and open sourceNo public API
Claude CodeThe agent the bundle plugs intoPro $20/mo and upYes, Anthropic API with usage credits

The practical read is that these tools stack rather than compete. ECC tunes the harness, BugHunter adds the security knowledge, and Cloudflare's skill covers the audit side. The sandbox question is separate again: if you want the agent's commands contained, run it inside a sandbox such as Eclipse Enclave, the free sandbox for AI coding agents covered here on November 19. For a fuller picture of the agent itself, see our Claude Code explainer.

Who should use it, and who should skip it

Use it if you already hunt bounties or run authorized external tests and you want Claude Code to follow a repeatable method instead of improvising. The reporting skills alone are worth the install: VRT-aware severity, program-specific formats, and evidence hygiene are the parts of bounty hunting that waste the most time. It also fits security teams that want a shared, auditable workflow across researchers.

Skip it if you have no authorized target. Without a scope, the bundle is a library with nowhere to go, and the authorization gates will fight you at every step. Skip it too if you need internal network testing, Active Directory attacks, or post-exploitation work, because those are deliberately out of scope and will stay that way. Beginners should treat it as a study aid on practice targets like OWASP Juice Shop first, not as a shortcut to a first bounty.

Common mistakes

Pointing it at a target you have not verified is in scope. The 7-Question Gate exists because a scope mistake can get you banned from a program, or worse. Confirm the program, the scope file, and the accepted-impact list before the first recon command.

Running it in plain English with no scope file. The engine path enforces scope in code. The plain-English path does not. Write the scope down, start with /hunt, and keep the engagement folder as the source of truth.

Installing from main and never pinning. A moving branch means your tooling can change between engagements. Pin a commit when you clone, or use the plugin route and its versioning.

Ignoring the model fallback. If Opus 5 silently hands off to Opus 4.8 mid-session, your results change and you may not notice. Watch for the notice, or disable automatic switching before a long run.

Treating stars as proof of findings. Fourteen thousand stars say the bundle is popular. They say nothing about how many of its findings survive triage. Measure it on a practice target before trusting it on a live program.

FAQ

Is Claude-BugHunter free?

Yes. The code is MIT licensed and the skill content is CC BY 4.0, so the bundle costs nothing. The real cost is Claude Code itself, which starts at $20 a month on the Pro plan, or $17 a month with annual billing. API usage is the pay-as-you-go alternative.

Does it work with tools other than Claude Code?

Yes. The skills are plain SKILL.md files, so they also load in OpenCode, OpenAI Codex CLI, Hermes Agent, and Google AntiGravity. The 15 slash commands and the /hunt engagement engine stay Claude Code only.

Do I need a HackerOne account to use it?

No. A HackerOne API username and token are optional and only let the bundled MCP server read your private programs and your own reports. Every skill and command works without them, including the public duplicate search and program scope tools.

Is it legal to use?

The bundle is built for authorized testing: assets you own, bug bounty programs with a defined scope, penetration test engagements, and practice targets. It explicitly excludes unauthorized targets, and Anthropic's safeguards can refuse offensive work even when you have permission unless you enroll in the free Cyber Verification Program. This is general information, not legal advice.

How many skills does it actually have, 82 or 83?

The README counts 83 skills and its category table adds up to 83. The repository description says 82. The gap is one skill in the count, and the README's figure is the one this article uses.

What should I practice on first?

OWASP Juice Shop, DVWA, and Hacker101 are the practice targets the project itself names. They are legal, self-contained, and built for exactly this kind of learning. Run a full engagement there, including the validation gate and a report, before touching a live program.

The bottom line

Claude-BugHunter earned its trending spot for an unglamorous reason: it replaces vibes with procedure. Eighty-three skills, 681 disclosed reports, and a validation gate turn Claude Code from a chatbot that knows about security into an operator that follows a method. The price is honest, free plus a Claude plan, and the authorization rules are written into the workflow rather than bolted on.

The honest limits are just as clear. Scope is enforced in code only on the engine path, the main branch moves, and no independent benchmark has measured how many of its findings survive triage. Install it, pin a commit or use the plugin, start on OWASP Juice Shop, and let the validation gate do its job. Used that way, it is a useful free addition to a security researcher's toolkit.

Sources

15 min read
7 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more