Claude-BugHunter's README Advertises 681 Attack Patterns and Cites 433, and Counts 83 Skills Where the Repo Says 82
Security / news
Claude-BugHunter's README Advertises 681 Attack Patterns and Cites 433, and Counts 83 Skills Where the Repo Says 82
The 14.8k-star Claude Code bundle for bug hunting is MIT-licensed and carries scope warnings, but its own figures disagree and Anthropic's access tiers decide who can run it.
Claude-BugHunter, a bundle of Claude Code skills for bug hunting that was climbing GitHub's trending list on Nov. 21, advertises 681 attack patterns drawn from public HackerOne reports. Its own README says only 433 of them are individually cited and auditable.
The repository, elementalsouls/Claude-BugHunter, is credited to Sachin Sharma (ElementalSoul), who is listed as working on bug hunting and generative A.I. security research. The page shows 14.8k stars, 1.7k forks, 355 commits, 4 open issues and 8 open pull requests. Stars measure interest, not adoption, and nothing on the page reports how many people have used it against a real target.
What is in the bundle
The skills are Markdown instruction files that load by topic. The README lists 58 hunt-* skills for web-application bug classes, attack chains for enterprise platforms (M365/Entra, Okta, vCenter, SSL VPNs, SharePoint), reporting and recon skills, and 15 slash commands. It says the skills also load in other harnesses, including OpenCode and Codex CLI. Code in Python and shell is MIT; methodology, wordlists and docs are CC BY 4.0.
The bundle does not run anything itself. Its README example walks through subdomain enumeration and live-host probing against a named HackerOne target and mentions Burp MCP, httpx and nuclei. The authors label that sample session as illustrative, not a recording.
Where the README disagrees with itself
| Claim | Where | Figure |
|---|---|---|
| Skill count | Repo description | 82 |
| Skill count | README | 83 |
| Disclosed-report patterns advertised | README | 681 |
| Patterns individually cited and auditable | README | 433 |
| Concrete bugs, CVEs or bounties credited to the author | README | none given |
The README calls the bundle "battle-tested across authorized red-team and bug-hunting engagements" but names no example. The 248 uncited patterns are the gap a reader cannot check. Sponsors are disclosed on the page.
Scope warnings and the access question
The README is explicit on authorization: "These skills are intended for assets you own or have written authorization to assess." Its triage-validation skill includes a scope check that asks whether the asset is in scope, and it excludes "weaponizing 0-days against unauthorized targets, post-exploitation tooling, malware development."
It also says Anthropic's safeguards block "vulnerability exploitation or offensive security tooling development" by default and points users to the Cyber Verification Program. One line is aimed at users who hit those blocks: "Do not reword an offensive engagement to look defensive to get past a classifier — enroll instead." A separate section explains how to phrase requests so they read as remediation, which the authors frame as accurate description.
Anthropic's own program page shows what enrolling means. Red Team Access, the tier that covers authorized penetration testing and exploit development, is open to organisations only, on Enterprise and API plans, with up to 75 approved users. Defense Access, which individuals can hold, covers security operations, malware analysis and vulnerability triage and validation. Red Team applicants must attest to phishing-resistant MFA and background checks and show a program such as a SOC 2 Type II report or a paid bug bounty running at least 12 months. Anthropic's review target for Red Team Access is about two weeks.
A freelance hunter on an individual plan therefore cannot enroll in the tier the README's offensive skills assume. Anyone considering the bundle should read each skill before loading it and confirm written scope first. Two other agent-skill repositories the site has examined, CLI-Anything and Superpowers, are covered separately.
Sources
More in Security
- 01Signal Now Blocks Siri AI on iOS 27 and Auto-Enables Screen Security on Android, but Cannot Stop macOS ScreenshotsMeredith Whittaker's Nov. 20 post lists what Signal changed on each platform and where its protection runs out, then asks Apple, Google and Microsoft for an opt-out.
- 02OpenAI's Cyber Tier Answers 95.0% of Advanced Requests; Anthropic's Lowest Tier Blocked 46 of 50 TrialsBoth labs now gate their least-restricted cyber models behind vetting. The two published figures measure different things, and both come from the vendors.
- 03Apache Roller's Second CVE Widens the Affected Range to 4.0.0, and CISA Lists It With Three Others Due Nov. 23A pre-authentication deserialization flaw first listed only 6.1.5. A vCenter bypass in the same batch was used by a suspected PRC-nexus actor, per Google.
- 04Windows Server 2028 Preview Hardens Server Core by Default in Build 29706, but NTLM Is Still OpenHundreds of settings are on from first boot, Microsoft warns Server 2025 workloads may break, and the 33 percent of servers still on 2019 are untouched.