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 harvest → coverage-gap → curation pipeline
Ingests real disclosed bug-bounty reports and turns them into candidate pattern entries that you curate into skills. It never auto-writes skill content — it collects, finds coverage gaps, and drafts skeletons. You own the curation.
Why it's shaped this way (honest constraints)
- HackerOne carries the value. Public GraphQL gives full metadata; each report
page exposes a ~400-char summary via its
<meta description>. Full writeup bodies are gated behind login even for disclosed reports (vulnerability_informationis null unauthenticated) — pass your own H1 session cookie (--h1-cookie) to fetch them, or read the full report at its URL during curation. - Bugcrowd is metadata-only. CrowdStream exposes program/priority/target/researcher — no title or technique. Collected as a reference feed (raw/bugcrowd/), not a pattern source.
- Intigriti has no public feed (disclosure is approval-gated).
harvest_intigriti.pyis an honest stub; curate Intigriti write-ups by hand. - Never verbatim. Reports are copyrighted by their authors; the repo distills
patterns + cites the URL and genericizes targets to
target.com. Raw corpus, drafts, and gaps stay gitignored. - Rate-limited, resumable, personal-research scale. No official bulk API exists.
The loop
# 1. collect (network; writes to gitignored raw/)
python3 research/reports/harvest_h1.py --sort bounty --limit 100
python3 research/reports/harvest_h1.py --sort votes --limit 200
# optional full bodies (your H1 session): --h1-cookie "$H1_COOKIE"
python3 research/reports/harvest_bugcrowd.py --pages 5 # metadata reference only
# 2. see what your skills DON'T cover yet (network-free)
python3 research/reports/report_coverage.py # -> research/reports/gaps.md
# 3. draft skeletons for a chosen skill (network-free; writes to gitignored drafts/)
python3 research/reports/draft_patterns.py --skill hunt-ato
# 4. CURATE (you): edit drafts/hunt-ato.md — fill the TODOs, confirm technique at the
# source, keep targets generic. Move approved entries into
# docs/disclosed-reports/hunt-ato.md, update skills/hunt-ato/SKILL.md, and bump
# report_count by the number actually added. New-technique candidates -> new skill.
# 5. guard + regenerate before committing
python3 scripts/scan_identifiers.py # no real client/target identifiers
python3 scripts/lint_skills.py # grounding + structure
python3 scripts/check_doc_counts.py
python3 scripts/gen_skill_catalog.py ; python3 scripts/gen_skill_index.py
Files
| File | Role |
|---|---|
harvest_h1.py |
HackerOne collector — metadata + summary snippet (public); full body via --h1-cookie |
harvest_bugcrowd.py |
CrowdStream metadata reference feed (not a pattern source) |
harvest_intigriti.py |
honest stub (no public feed) |
classify_reports.py |
report → vuln class → hunt-* skill (CWE + title heuristics) |
report_coverage.py |
gap detector — ranked uncovered reports per skill + new-technique candidates |
draft_patterns.py |
auto-draft skeleton pattern entries (genericized, TODO markers) |
test_pipeline.py |
network-free sanity check (classify + genericize + skeleton) |
raw/, drafts/, gaps.md, processed.json are gitignored runtime artifacts.
Honest ceiling
This accelerates discovery and drafting; it does not replace judgment. Pattern
quality, target genericization, report_count, and the "681" headline stay human-set.
Most marginal reports repeat known patterns — the win is filling under-grounded skills
and surfacing genuinely new techniques, which the gap report ranks for you.