| name | hunt |
| description | Hunt a target codebase or web app for a specific vulnerability class, primed with real exploitation patterns from the local disclosed-report corpus. Use when the user wants to find real vulnerabilities — "look for IDOR/SSRF/XSS/auth bypass in this code", "audit this for <vuln class>", "find bugs in <target>". Combines corpus-derived patterns with source/endpoint analysis and reports concrete candidate findings. |
hunt
Find real vulnerabilities of a given class in a target (source code, or a running
app's endpoints), using the disclosed-report corpus to know exactly what to look
for and how it gets exploited.
Workflow
-
Scope. Confirm the target (which repo/dir, or which host/endpoints) and the
vuln class(es) to hunt. If the user named a target host, confirm they have
authorization (bug bounty scope, pentest, own asset) before active testing.
-
Prime from precedent. Pull real exploitation patterns:
python3 .claude/skills/_lib/search_reports.py "<vuln class>" --full --limit 25
Read the most detailed reports to learn the source/sink patterns, dangerous
APIs, and bypasses that worked in the wild.
-
Map the attack surface.
- Source: grep for the sinks/sources the corpus flagged (e.g. for SSRF: HTTP
client calls taking user input; for SQLi: string-built queries; for XSS:
unescaped output,
dangerouslySetInnerHTML, v-html, template injection;
for IDOR: object lookups by user-supplied id without authz checks).
- Running app: enumerate endpoints/params and probe the patterns from the
playbook (use
recon-playbook if a full methodology helps).
-
Investigate candidates. For each potential finding, trace data flow from
source to sink. Confirm the missing control. Distinguish real exploitability
from false positives — verify the guard truly isn't there.
-
Report findings as a ranked list. For each:
- Title, vuln class, severity (rough), confidence.
- Location —
file:line or endpoint.
- Why it's exploitable — the data flow and the missing/insufficient control.
- Closest precedent —
#<id> — <title> from the corpus.
- Next step — how to confirm / build a PoC (hand off to
poc).
Rules
- Only do active testing against targets the user is authorized to test.
- Prefer precision: a few well-traced, exploitable findings beat a wall of grep
hits. Mark anything unconfirmed as "needs verification".
- Tie each finding to a concrete code path or request, not a vague suspicion.
- Hand off confirmed findings to
write-report, and PoC building to poc.