Claude Code Taught to Hunt Vulnerabilities: 83 Skills Built from 681 HackerOne Reports
Claude-BugHunter is a set of 83 skills and 15 commands that turns Claude Code from a chat assistant into a helper for vulnerability hunters. The skills are built from 681 public HackerOne reports and load on their own when you describe a test target. The repository gained 124 stars in one day on GitHub Trending. You no longer have to re-explain the methodology for every project, but the toolkit should only be applied to targets you are authorized to test.
AI-processed from GitHub Trending AI; edited by Hamidun News
If you have asked a language model to look for holes in a web application, you know the problem: a prompt like "act as a pentester" returns generic advice about the OWASP Top 10 and nothing you can immediately check against a specific service. The other route is also clumsy: you collect bookmarks of dozens of analyzed reports and reread them before every new target. Claude-BugHunter offers a third option: package the experience of bug hunters into a set of skills that Claude Code pulls in by itself. In one day, the project gained 124 stars on GitHub Trending.
What's inside and where the numbers come from
This is a self-contained package for Claude Code. The GitHub Trending listing says 82 skills, but the README itself says 83 skills and 15 slash commands; we rely on the README. The foundation of the set is 681 public reports from HackerOne, a platform where researchers disclose vulnerabilities to companies and receive rewards. For 433 of those 681 patterns, the authors have added individual links to their source, so they can be verified.
The patterns are organized into 24 main vulnerability classes: XSS, SQL injection, SSRF, IDOR, OAuth, JWT, file upload, and others. For each class there are detection signals, payload examples, bypass tables, and attack chain templates. A separate layer covers enterprise platforms: M365 and Entra, Okta, vCenter, SharePoint, SSL VPN appliances, and cloud access rights. It includes CVE chains from 2024 to 2026 and privilege escalation after credentials are obtained.
You don't need to call skills by name. You write in plain language, for example: "I'm testing acme.com, it's an authorized target on HackerOne, do reconnaissance and prioritize the surface." The relevant skills load based on the topic. In the demo example from the README, the agent enumerates subdomains, finds live hosts, and suggests starting with the GraphQL API. The authors note that this fragment is illustrative and not a recording of an actual session.
How it works and what it learned from
The methodology is structured as a non-linear process of six phases: reconnaissance, surface mapping and ranking, discovery, validation, and reporting. Task boundaries are enforced by code. Before a report is submitted, a "seven-question gate" runs, including whether the asset is in scope and whether that type of impact is accepted by the program. As a result, the model is less likely to write up a finding the platform will reject.
According to the authors, the set grew out of real engagements and addressed specific failures. Drafts were written before validation and hours were lost. The standard severity table undervalued P3 findings as P4. Screenshots leaked cookies and victims' personal data. During an external test, a client's SOC closed a confirmed SQL injection within 30 minutes, while a third-party attacker blocked 14 accounts. Skills for evidence handling and for detecting defender reactions during testing were added for such situations.
Installation, limits, and a warning about Anthropic
The simplest way to install the package is as a Claude Code plugin using /plugin marketplace add and /plugin install. The second option is to clone the repository and run the installation script. The skills also work in OpenCode, OpenAI Codex CLI, Hermes Agent, and Google AntiGravity; slash commands and the orchestration engine remain Claude Code only. There is a Burp integration via MCP. A HackerOne account is optional: an API token is needed only to read your private programs and reports.
The boundaries are firmly set. The set deliberately excludes attacks on internal Active Directory, C2 frameworks, post-exploitation, antivirus and EDR evasion, iOS, hardware, and binary exploitation. It covers the external perimeter, not a full red team. It may only be used against assets with written authorization: in-scope bug bounty targets, a penetration testing contract, CTFs, or your own infrastructure.
There is also a practical risk. The authors warn that Anthropic's models block some requests about vulnerability exploitation and offensive tool development in real time. If you work legitimately, they advise joining the Anthropic Cyber Verification Program. Rephrasing the task as "defensive" to get around the classifier is not recommended.
For practitioners, the takeaway is simple. If you run bug bounty or external tests, the ready-made skills will save time on methodology and report writing. But this is a tool for an experienced researcher: decisions about scope and validating findings remain your responsibility. The methodology and documentation are licensed under CC BY 4.0.
- Nov 21, 2026Anthropic Paid $3.4 Billion for 1.7 GW of Gas Generators
- Nov 21, 2026OpenAI Names 4 Reasons GPT-6 Astra Hides Its Reasoning from Monitors
- Nov 21, 2026Australia's Central Bank: AI Hits Young Workers Aged 15–24 First in IT and Customer Support
- Nov 21, 2026Anthropic: Claude responds appropriately to self-harm queries in 94% of app conversations and 60% via API
Key Questions Answered
What's inside and where the numbers come from?
This is a self-contained package for Claude Code. The GitHub Trending listing says 82 skills, but the README itself says 83 skills and 15 slash commands; we rely on the README. The foundation of the set is 681 public reports from HackerOne, a platform where researchers disclose vulnerabilities to companies and receive rewards. For 433 of those 681 patterns, the authors have added individual links to their source, so they can be verified.
How it works and what it learned from?
The methodology is structured as a non-linear process of six phases: reconnaissance, surface mapping and ranking, discovery, validation, and reporting. Task boundaries are enforced by code. Before a report is submitted, a "seven-question gate" runs, including whether the asset is in scope and whether that type of impact is accepted by the program. As a result, the model is less likely to write up a finding the platform will reject.
What are the key details on Installation, limits, and a warning about Anthropic?
The simplest way to install the package is as a Claude Code plugin using /plugin marketplace add and /plugin install. The second option is to clone the repository and run the installation script. The skills also work in OpenCode, OpenAI Codex CLI, Hermes Agent, and Google AntiGravity; slash commands and the orchestration engine remain Claude Code only. There is a Burp integration via MCP. A HackerOne account is optional: an API token is needed only to read your private programs and reports.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.