Claude-BugHunter: Use 681 disclosed reports to guide Claude Code vulnerability hunting
A Claude Code skill bundle for bug bounty and authorized red teams, turning 681 reports into auto-loaded hunting and delivery workflows.
GitHub elementalsouls/Claude-BugHunter Updated 2026-11-16 Branch main Stars 6.4K Forks 970
Python PowerShell Shell Bug bounty External red teaming Claude Code Burp MCP Enterprise identity and infrastructure

🧭 Decision Guide

Try it if you

  • You use Claude Code to test web applications and need detection patterns organized around XSS, SSRF, IDOR, and other vulnerability classes.
    The README's “What's inside” lists 58 hunt-* skills, including XSS, SSRF, and IDOR, and says they are based on 681 HackerOne reports.
  • Your authorized assessment covers M365, Entra, Okta, vCenter, or enterprise VPNs.
    The README's “Hit the perimeter” layer lists M365/Entra, Okta, vCenter, SSL-VPN appliances, SharePoint, and cloud IAM, with 2024–2026 CVE chains.
  • You need more than vulnerability discovery, including validation, VRT severity, out-of-scope rebuttals, and PII redaction.
    The README's “Ship it” layer lists triage-validation, the 7-Question Gate, VRT-aware severity, OOS rebuttals, and PII redaction.
  • You want skills to trigger from natural-language descriptions instead of memorizing every skill name.
    The README's “What is this?” section says relevant skills trigger by topic when the user describes what they are testing in plain English.

Skip it if you

  • You need to test an unauthorized target, or your work does not permit external red-team or bug-bounty activity.
    The README describes the project as intended for authorized red-team and bug-hunting engagements and says authorized work should use CVP.
  • Your work is limited to pure OSINT reconnaissance and does not need the project's vulnerability-hunting and delivery layers.
    The README's “About” section describes Claude-OSINT as the sister project for the recon phase, while this project picks up afterward.
  • You cannot accept the model switching from Opus 5 to Opus 4.8 for high-risk cybersecurity requests.
    The README's “Why your model switched mid-session” section explicitly says Opus 5 falls back to Opus 4.8 for some high-risk requests.

Requirements

  • Claude Code skills system is required; the README defines the project as a drop-in skill bundle for that system.
  • The use case should be an authorized red-team or bug-hunting engagement, or one of the listed training platforms such as DVWA, OWASP Juice Shop, Hacker101, and testphp.vulnweb.com.
  • The README's tool inventory includes PortSwigger Burp Suite + MCP Server extension, ProjectDiscovery tools, SecLists, and Assetnote Wordlists, but does not state which are hard requirements.

First step (verbatim from README)

/hunt

Watch out

  • The Opus 5 to Opus 4.8 fallback notice can be easy to miss during a long run.
    The README's “Why your model switched mid-session” section says the notice and response-model label can be easy to scroll past in a long agentic run.
  • The README lists both 82 and 83 skills, so the bundle size needs verification.
    Project metadata says 82 skills; the README title and “What's inside” say 83 skills.
  • External red-team use requires explicit authorization, scope, and remediable findings.
    The README says /hunt declares an engagement frame of authorized, scope-bounded, and remediable findings on its first turn.
  • The project claims enterprise CVE chains, but the provided material does not list the individual CVE identifiers.
    The README only states that enterprise platform chains include current 2024–2026 CVE chains.

Alternatives

  • Claude-OSINT:Your work centers on offensive-osint and osint-methodology during the recon phase rather than post-recon vulnerability discovery and workflow.
    README “About”
  • ProjectDiscovery toolset:You need standalone reconnaissance and scanning tools such as subfinder, dnsx, httpx, katana, or nuclei rather than Claude Code skill orchestration.
    README “Tool inventory”

Not stated in the README

  • The complete Quickstart installation commands, supported operating systems, and Claude Code version are not provided.
  • The README does not specify Claude model access requirements, API configuration, operating cost, or context limits.
  • No performance benchmarks, false-positive rate, successful-finding rate, or comparison with ordinary Claude Code prompting is provided.
  • The README does not state whether Burp MCP, ProjectDiscovery, SecLists, and Assetnote Wordlists are required dependencies.
  • The README does not explain the discrepancy between 83 skills and the metadata count of 82.
  • The provided material omits the details of the Authorization, Scope, Out of scope, and Runs on multiple agent harnesses sections.

💡 Deep Analysis

6
Yes I already use Claude-OSINT and need Claude Code on both macOS/Linux and Windows. If I install Claude-BugHunter as well, are duplicate skills, uninstallation, and cross-platform coexistence manageable?
For: A security engineer already using Claude-OSINT who needs cross-platform installation without conflicting skills

Yes, because the README explicitly provides coexistence and cross-platform mechanisms for Claude-OSINT.

  • It states that offensive-osint and osint-methodology are canonically maintained in Claude-BugHunter and re-exported by Claude-OSINT, with both copies kept byte-identical.
  • macOS/Linux use install.sh, while Windows uses install.ps1; the installers record a manifest and skip copying an identical skill again.
  • During uninstall, a skill still owned by the other bundle is retained, so removing one project does not break the other project’s OSINT capabilities.
  • This is a better fit than manually copying 83 skills across two operating systems and projects. However, content conflicts, update precedence, and synchronization between versions still require human confirmation.
  • About: "Installing both is safe"
  • About: "offensive-osint ... are canonically maintained here and re-exported there"
  • About: "install.sh on macOS/Linux, install.ps1 on Windows"
  • About: "records a manifest, the script skips re-copying an identical skill"
  • About: "uninstalling one never breaks the other"
Not stated in the README:The excerpt does not include the full arguments or installation paths for install.sh and install.ps1;It does not explain manifest conflict handling when different project versions coexist;It does not state the minimum PowerShell version or execution-policy requirements on Windows
Yes I mainly test XSS, SSRF, IDOR, and SQLi within HackerOne scopes, and I often face incomplete validation, VRT severity mistakes, and cookies leaking into screenshots. Does this bundle fit my bug-bounty workflow?
For: An independent bug hunter using Claude Code on HackerOne programs who must submit reproducible web vulnerabilities aligned with VRT

Yes, because it directly addresses the breaks between hypothesis, validation, and submission in a bug-bounty workflow.

  • The README says its 58 hunt-* skills are curated from 681 disclosed HackerOne reports and cover XSS, SQLi, SSRF, IDOR, and other classes across 24 core vulnerability categories.
  • triage-validation includes the “7-Question Gate” and VRT-aware severity handling, targeting premature drafts and incorrect P3/P4 decisions.
  • evidence-hygiene includes PII redaction; the project history specifically records screenshots leaking cookies and victim PII.
  • Skills load automatically by topic, so the hunter does not need to remember individual skill names.

It is not a standalone scanner, however. Actual requests, impact confirmation, and scope decisions still depend on Claude Code, external tools, and human review.

  • README: "58 `hunt-*` skills curated from 681 disclosed HackerOne reports"
  • README: "across 24 core vulnerability classes"
  • README: "triage-validation + reporting + evidence-hygiene"
  • README: "the 7-Question Gate, VRT-aware severity ... PII redaction"
  • Why this exists: "screenshots leaked cookies and victim PII"
Not stated in the README:The supplied README excerpt does not include the Quickstart installation command or minimum Claude Code version;It does not state which HackerOne VRT revision or custom program rules are supported;It does not provide the exact Claude Code and Burp MCP configuration steps
It depends I currently practice web security only on DVWA, OWASP Juice Shop, and Hacker101, and I am just beginning with Claude Code, Burp Suite, and authorized testing. Will the 83 skills make this too difficult to use?
For: A security learner using Claude Code workflows against DVWA, OWASP Juice Shop, and Hacker101 training environments

It depends: the bundle can support learning when the targets are explicitly authorized, but it is not designed as a single beginner teaching tool.

  • The README names DVWA, OWASP Juice Shop, Hacker101, and testphp.vulnweb.com as public training or validation environments, so it is not limited to production red-team work.
  • At the same time, it contains 83 auto-loaded skills, 15 slash commands, enterprise identity matrices, and complex attack chains; the project insight explicitly identifies the skill count and platform breadth as a significant beginner learning burden.
  • Effective use assumes familiarity with Claude Code, HTTP/web security, Burp Suite, and authorized testing. Natural-language triggering reduces command memorization, but not the need to understand payloads, scope, and impact.
  • The project is restricted to authorized owned assets, bounty scopes, or formal red-team engagements; listing a training platform does not authorize testing arbitrary third-party systems.
  • README: "public training platforms (DVWA, OWASP Juice Shop, Hacker101, testphp.vulnweb.com)"
  • README: "83 skills · 15 slash commands"
  • README: "All triggered automatically by topic"
  • Project insight, user_experience.learning_curve: high learning difficulty for beginners
  • Project insight, usage_limitations: limited to explicitly authorized owned assets, bounty scopes, or formal red-team engagements
Not stated in the README:The excerpt does not provide a beginner learning sequence, sample sessions, or exercise list;It does not map specific skills or slash commands to DVWA, Juice Shop, or Hacker101;It does not state the required Claude Code, model, or Burp Suite versions for the training environments
Yes During red-team reconnaissance I often obtain AWS keys, JWTs, or GCP JSON credentials, and then need to determine privileges and analyze cloud-IAM post-credential escalation. Is this project more suitable than using only OWASP web skills?
For: A red-team engineer responsible for enterprise cloud identity security and frequently handling AWS keys, JWTs, and GCP JSON credentials

Yes, because it connects post-credential privilege analysis to web, identity, and infrastructure attack chains instead of offering only OWASP payloads.

  • The README’s enterprise attack layer includes cloud IAM and explicitly mentions post-credential escalation, matching AWS keys, JWTs, and GCP JSON scenarios.
  • Its “Why this exists” section records that reconnaissance produced these credentials but the team could not determine what they granted or how to escalate; cloud-iam-deep was added for that gap.
  • The bundle also covers M365/Entra, Okta, vCenter, VPN, and SharePoint, allowing cloud-identity findings to be considered within a broader enterprise attack surface.
  • Platform skills do not guarantee applicability to a particular tenant, version, or permission configuration. The excerpt also does not explain whether credentials are processed locally, exposed to the model, or which cloud APIs are covered.
  • README: "enterprise platform chains ... cloud IAM"
  • README: "current 2024–2026 CVE chains + post-credential escalation"
  • Why this exists: "when recon yielded credentials (AWS keys, JWTs, GCP JSON), it was unclear what they granted or how to escalate"
  • Why this exists: "cloud-iam-deep"
  • Project insight, usage_limitations: tenant, version, permission model, and authorization scope must be checked
Not stated in the README:The excerpt does not list the specific services and APIs supported for AWS, GCP, or other clouds;It does not state whether JWTs, cloud keys, and GCP JSON are automatically redacted or written into engagement folders;It does not clarify whether cloud-iam-deep can call cloud CLIs/APIs or only generate analysis and test hypotheses
It depends In an external red-team engagement I need to test M365/Entra, on-prem SharePoint, Cisco SSL VPN, and vCenter, then deliver a 50KB+ Markdown and DOCX report with screenshots. Can this project serve as my primary workflow assistant?
For: An external red-team service team covering M365/Entra, SharePoint, Cisco SSL VPN, and vCenter

It depends: it is well suited to organizing this red-team work, but it cannot replace attack tools, authorization controls, or validation of the final report.

  • The README’s enterprise layer explicitly lists M365/Entra, Okta, vCenter, SSL-VPN appliances, SharePoint, and cloud IAM, including 2024–2026 CVE chains and post-credential escalation.
  • redteam-mindset and mid-engagement-ir-detection address monitored targets, SOC discovery, and account-lockout situations rather than applying bug-bounty defaults blindly.
  • The README describes redteam-report-template for enterprise delivery, including 50KB+ Markdown, DOCX, and embedded screenshots, which matches your deliverable constraint better than a bounty template.
  • However, this remains a Claude Code skill bundle, not a validation platform. Credential privileges, CVE applicability, production impact, and final document quality are not automatically guaranteed.
  • README: "enterprise platform chains (M365/Entra, Okta, vCenter, SSL-VPN appliances, SharePoint, cloud IAM)"
  • README: "current 2024–2026 CVE chains + post-credential escalation"
  • Why this exists: "client SOC patched confirmed SQLi within 30 min"
  • Why this exists: "external attacker locked 14 accounts"
  • Why this exists: "50KB+ MD + DOCX with embedded screenshots"
Not stated in the README:The excerpt does not say whether DOCX is generated automatically or whether the project only supplies report content templates;It does not specify supported product versions, tenant configurations, or the exact CVE list for each platform;It does not define the full integration boundary with existing Burp, scanners, or credential-management systems
It depends I already use Burp Suite MCP and ProjectDiscovery for reconnaissance and proxy testing, and I want Claude Code to route automatically to web, enterprise-platform, and reporting skills based on the attack surface. Can this project reduce manual coordination?
For: A security engineer integrating Burp Suite MCP, ProjectDiscovery, and existing reconnaissance tools into an external red-team workflow

It depends: it can reduce manual coordination of skills and engagement artifacts, but it does not establish end-to-end automation for tool calls, request control, or result import.

  • The README lists Burp MCP integration and says the bundle can slot into an existing engagement workflow, which fits a team retaining its proxy and reconnaissance tools.
  • The engagement engine maps a target’s attack surface to the skill handling the finding, while 15 slash commands expose workflow entries such as /hunt, /recon, and /report.
  • Topic-based automatic loading removes the need to invoke each skill manually when moving from web applications to M365/Entra, cloud IAM, or reporting.
  • The project insight still says actual detection depends on Claude Code, the model, external tools, and user context. Rate limits, account lockouts, production impact, and sensitive-data exposure are not automatically solved by MCP integration.
  • README: "Burp MCP integration"
  • README: "a deterministic engagement engine ... maps a target's attack surface and routes each finding to the skill that handles it"
  • README: "Also ships 15 slash commands (`/hunt`, `/recon`, `/report`, …)"
  • README: "All triggered automatically by topic"
  • Project insight, usage_limitations: actual detection depends on Claude Code, model capability, external tools, and user-provided context
Not stated in the README:The excerpt does not specify Burp MCP server configuration, permission model, or supported Burp versions;It does not say whether ProjectDiscovery has a native adapter, input/output format, or command wrapper;It does not clarify whether the engagement engine executes external tools or only routes skills

✨ Highlights

  • 58 hunt skills cover 24 web vulnerability classes
  • 681 HackerOne reports inform its detection patterns
  • M365, Okta, and vCenter include 2024–2026 chains
  • The README says 83 skills, while metadata says 82

🔧 Engineering

  • Claude Code auto-loads 83 skills without name-based invocation
  • 15 slash commands include /hunt, /recon, and /report
  • The engine routes findings to skills by attack surface
  • triage-validation provides the 7-Question Gate and VRT-aware severity

⚠️ Risks

  • The README explicitly targets authorized red-team and bug-bounty work
  • High-risk requests on Opus 5 may fall back to Opus 4.8
  • The README's 83-skill count conflicts with project metadata showing 82
  • The provided README material lacks Quickstart commands and version requirements

👥 For who?

  • Researchers using Claude Code for web-application bug bounty work
  • Teams assessing M365, Entra, and Okta in red-team engagements
  • Security engineers needing Burp MCP and reporting-hygiene workflows