A Claude Code skill bundle for bug hunting and external red-team work - 82 skills, 15 slash commands, 681 disclosed-report patterns curated across 24 core vulnerability classes, plus enterprise identity + infrastructure attack matrices.
Disclosed-report pattern libraries
Each hunt-<class>.md here is the cited grounding behind the matching
skills/hunt-<class>/SKILL.md. The SKILL.md holds the methodology an operator
runs; these files hold the public disclosures that prove the patterns are real.
They are distilled + cited, never verbatim — the technique is the value, not the
victim.
The report_count convention
A skill's report_count: frontmatter = the number of distinct disclosed-report
URLs cited in its pattern library (grep -oE 'reports/[0-9]+' → unique). It is
grep-auditable: the number always equals what you can count in the file. Nothing
is asserted that a reader can't click.
- Counts that went up = real cited reports added.
- Counts that went down (e.g.
hunt-source-leak31→7) = a prior number was asserted without a backing corpus and has been reconciled to what's actually cited. This is honesty, not lost coverage — the SKILL.md methodology body is unchanged; only the grounding count is now truthful.
The repo-wide 681 headline is a different measure — the count of distilled
patterns across the core classes — and is kept as-is. 433 is the
now-auditable count of individually cited disclosed reports. Both numbers,
stated side by side, mean what they say.
Two-tier grounding (web + API)
Deep API bugs (mass-assignment, BOLA/BFLA, shadow/zombie APIs, GraphQL depth) are
rarely disclosed publicly, so H1 reports under-cover them. Those skills are
grounded on a second tier cited under sources: in the SKILL.md
(PortSwigger / Assetnote research, named CVEs) rather than report_count. The two
tiers are labeled distinctly and never conflated: report_count = disclosed
reports only; sources: = research / CVE grounding.
The credibility gate
research/reports/verify_citations.py is the check that makes these files
trustworthy. For every cited report it fetches the public subject (title +
<meta> summary), re-runs the class router over that content, and flags any
report whose actual subject does not corroborate the skill it's cited under:
python3 research/reports/verify_citations.py # verify all
python3 research/reports/verify_citations.py --skill hunt-saml
Current state: 433 unique cited reports — 0 mismatch, 0 weak. Every citation is content-verified. Re-run this after adding any citation.
Genericization rule
Targets are always genericized — victim hostnames/company names are dropped, the
technique and the report URL are kept. scripts/scan_identifiers.py
(leak-guard) runs over every change; a real target name in prose is a bug, not a
citation.