Bug Hunter — Autonomous, Authorization-Gated Bug Bounty Agent

Recon → active scan → known-vuln intel → reproduction-based triage → report. Every network call is scope-gated and rate-limited — nothing is tested without explicit authorization. Zero pip dependencies (Python 3 stdlib only).

Install

Claude Code (skill — auto-triggers)

ln -sfn "$(pwd)" ~/.claude/skills/bug-hunter   # run from this repo root

Restart Claude Code, then trigger with: hunt example.com, bug bounty on *.example.com, scan target acme.com, or /bug-hunter. Verify with /skills.

Codex / any other agent (no skill system)

Tell the agent: "Read SKILL.md and follow it." See AGENTS.md — Codex reads it automatically.

Optional power-ups (auto-detected if installed)

brew install nuclei subfinder        # nuclei = biggest coverage boost

Run

1. Authorize + initialize (only against targets that permit automated testing):

python3 scripts/init_hunt.py \
  --domains example.com "*.example.com" \
  --confirmed-by "you@email / program-name" --rate 3 --note "scope source"
# prints the hunt dir on the last line -> capture as $HUNT

2. Hunt (autonomous, resumable):

python3 scripts/orchestrate.py --hunt-dir $HUNT

Or phase-by-phase:

python3 scripts/recon.py     --hunt-dir $HUNT
python3 scripts/web_scan.py  --hunt-dir $HUNT
python3 scripts/cve_intel.py --hunt-dir $HUNT
python3 scripts/triage.py    --hunt-dir $HUNT
python3 scripts/report.py    --hunt-dir $HUNT

Extend scope mid-hunt (after re-confirming authorization):

python3 scripts/init_hunt.py --hunt-dir $HUNT --add \
  --domains newasset.example.net --confirmed-by "you"

Output ($HUNT/)

File	Contents
scope.json	authorized targets, who confirmed, rate limit
ledger.json / worklog.jsonl	resumable phase state / audit trail
recon.json	hosts, live URLs, tech, collected URLs
evidence/	per-request capture (request+response+repro curl)
findings/	raw_ / validated_ / rejected_ findings
reports/INDEX.md	severity-sorted summary + per-finding reports

What it detects

Missing security headers · CORS reflect+credentials · open redirect · exposed files (.git/.env/actuator…) · secrets in JS · subdomain takeover · GraphQL introspection · SQLi (error + double-confirmed time-based) · reflected XSS · nuclei templates · NVD/Exploit-DB leads.

Access control (IDOR/BOLA), business logic, and SSRF chains are agent-driven — the scripts flag hints; the human/agent reasons and tests via the scope-gated client. See knowledge/methodology.txt.

Architecture

lib/scope.py     authorization + rate limit (anchored-dot host matching)
lib/nethttp.py   hardened HTTP client; scope-checks + throttles + captures evidence
lib/findings.py  finding model, severity/CWE, dedup, markdown
lib/hunt.py      resumable ledger + worklog
scripts/*.py     init_hunt · recon · web_scan · cve_intel · triage · report · orchestrate

Safety / legal

Only test programs that explicitly permit automated testing — many bug bounty policies forbid or rate-limit scanners. The scope gate and rate limit are there to protect you; do not weaken them. Out-of-scope hosts are refused before any packet is sent.

autonomous-bug-bounty