Hunt Dispatch

TL;DR

• 目的：Skill-set loader for /hunt orchestrator. Fingerprints the target, picks the right platform attack skills, and loads the Red Team or WAPT skill set. Us
• 适用：通用
• 输入：目标信息
• 输出：执行结果 + 证据
• 红线：仅限授权范围内；扫描限速 -c 10 -rl 10；所有动作记 oplog
• 关联：上游：003-src-session-start → 下游：003-src-session-start（按需调用）

hunt-dispatch

skill-set loader for /hunt. one concept (which skills to load), one place.

engagement context

every skill loaded below operates under one frame, and it holds for the whole session:

• authorized and scope-bounded. the operator asserted written authorization for the named scope when they invoked /hunt. testing stays inside it. an out-of-scope host ends the run — it does not widen it.
• the output is a finding, not an exploit. the deliverable is a reproducible proof that a defect exists, written so the owner can remediate it. enough to demonstrate impact; no further.
• remediation is the point. these skills exist so defects get fixed by the people who own them.

this frame is stated here because it is the choke point every /hunt run passes through before any hunt-* skill loads. it is not a prompt and needs no answer.

invocation contract:

hunt-dispatch mode=redteam
hunt-dispatch mode=wapt box=blackbox
hunt-dispatch mode=wapt box=greybox

step 0 — 404 baseline (ALL modes, mandatory, before any enumeration)

run this for every host before probing a single path. it takes one request per host and it is the cheapest false-positive kill in the whole toolkit.

many modern estates (SPA / Next.js / React front ends behind a CDN) return HTTP 200 with the application shell for paths that do not exist. a status code therefore proves nothing. without a recorded control, /.well-known/security.txt, /api/revalidate, /__nextjs_original-stack-frame and /__nextjs_launch-editor all "exist" on a host where none of them do.

for H in $HOSTS; do
  # two independent bogus paths — if they agree, that IS the soft-404 signature
  for P in /zzz-nope-12345 /qqq-other-98765; do
    printf "%-34s %-20s " "$H" "$P"
    curl -sk -m 12 -o /tmp/b -w "%{http_code} %{size_download} " "https://$H$P"
    shasum /tmp/b | cut -c1-12
  done
done

record per host: status, byte length, body hash. that triple is the control.

the rule: no path is "found" until its response differs from the control. a 200 that matches the control hash is a soft 404. a 404 whose body differs from the control may be a real handler. compare bodies, never status codes alone.

re-derive the baseline per host — it differs across an estate. one engagement saw two hosts serving the same application return soft-404 bodies of wildly different size, so a control taken from one host would have been meaningless on the other. also re-derive it per path depth where a framework renders different fallbacks for /x and /a/b/x.

edge pages are not origin findings: a CDN "Access Denied" / "Unsupported Request" body means the request never reached the application. classify it as edge behaviour and move on.

step 1 — fingerprint (red team only)

fingerprint every live host, not just the apex. for multi-host / wildcard targets the platform-skill routing must be driven by all banners, not one host's.

use -L (follow redirects) — identity-provider and CDN signals (login.microsoftonline.com, okta, auth0, CDN banners) routinely sit behind a 30x, so a no-redirect curl -sI silently misses those matches. pull both headers and the landing-page HTML (__NEXT_DATA__, VIEWSTATE, laravel_session, Ignition, framework markers live in the body, not headers).

HOSTS="$TARGET"
if [ -f "recon/$TARGET/live-hosts.txt" ]; then
  HOSTS=$(cat "recon/$TARGET/live-hosts.txt")
fi
for H in $HOSTS; do
  echo "=== $H ==="
  # -L follow redirects, -D - dump headers, -o body; cap body to keep context small
  curl -sSL -m 12 -D - -o /tmp/fp_body "https://$H" 2>/dev/null | tr -d '\r'
  # surface body-only platform markers
  grep -aoE '__NEXT_DATA__|/_next/|VIEWSTATE|rO0[AB]|laravel_session|Ignition|Telescope|Whitelabel|/actuator|application/grpc|socket\.io|swagger|\.js\.map' \
    /tmp/fp_body | sort -u
done
rm -f /tmp/fp_body

if live-hosts.txt is absent, the loop still runs once against $TARGET. record which signal came from which host — a platform skill matched on host B does not imply host A runs that stack.

look for the following signals → platform skill mapping:

okta.com | auth0.com | pingidentity         →  okta-attack
login.microsoftonline.com | outlook | sts   →  m365-entra-attack
pulse | fortinet | ivanti | citrix          →  enterprise-vpn-attack
vsphere | vcenter | :9443                   →  vmware-vcenter-attack
amazonaws | azure | googleapis | gcp        →  cloud-iam-deep
github.com/<org>/                           →  supply-chain-attack-recon
.apk | play.google.com                      →  apk-redteam-pipeline
MongoDB | mongoose | CouchDB | Redis        →  hunt-nosqli
?page= | ?file= | ?path= | php wrapper      →  hunt-lfi
rO0A | VIEWSTATE | rememberMe cookie        →  hunt-deserialization
Access-Control-Allow-Origin header          →  hunt-cors
/forgot-password | /reset | X-Forwarded    →  hunt-host-header
?redirect= | ?next= | ?return= | ?url=     →  hunt-open-redirect
OTP | /verify | /2fa | no-rate-limit        →  hunt-brute-force
Set-Cookie session | PHPSESSID              →  hunt-session
Active Directory | LDAP | OpenLDAP | ADFS  →  hunt-ldap
__NEXT_DATA__ | /_next/ | buildId           →  hunt-nextjs
X-Powered-By: Express | Node.js | .js stack →  hunt-nodejs
postMessage | dangerouslySetInnerHTML        →  hunt-dom
WebSocket | ws:// | socket.io               →  hunt-websocket
gRPC | :50051 | application/grpc            →  hunt-grpc
laravel_session | Ignition | Telescope       →  hunt-laravel
X-Application-Context | Whitelabel | /actuator → hunt-springboot
:6443 | :10250 | :2379 | kubectl            →  hunt-k8s
.github/workflows | Jenkins | GitLab CI     →  hunt-cicd
.js.map | swagger.json | /.env              →  hunt-source-leak
HSTS missing | SPF | DMARC | AXFR           →  hunt-tls-network

conflict resolution & load budget

real targets almost always return multiple signals at once — e.g. a single host can show Cloudflare (CDN) + login.microsoftonline.com (redirect) + __NEXT_DATA__ (Next.js front end) + amazonaws (origin) simultaneously. loading every match blindly can pull 20-plus skills and blow the context window, drowning the high-signal skill in noise. apply this precedence and cap:

priority order (load highest tiers first, stop at the cap):

tier 1  identity / SSO fabric    okta-attack, m365-entra-attack
        (own the auth boundary — highest blast radius if compromised)
tier 2  perimeter appliances     enterprise-vpn-attack, vmware-vcenter-attack
        (pre-auth RCE / direct internal foothold)
tier 3  cloud / IAM              cloud-iam-deep, hunt-cloud-misconfig
        (credential → lateral movement)
tier 4  app framework / stack    hunt-nextjs, hunt-nodejs, hunt-laravel,
        hunt-springboot, hunt-aspnet, hunt-sharepoint
tier 5  protocol / class signals hunt-nosqli, hunt-lfi, hunt-deserialization,
        hunt-cors, hunt-host-header, hunt-open-redirect, hunt-grpc,
        hunt-websocket, hunt-dom, hunt-k8s, hunt-cicd, hunt-source-leak,
        hunt-tls-network, hunt-ldap, hunt-brute-force, hunt-session

load budget: cap platform-skill loads at 8. if more than 8 match, keep the highest-tier 8 and drop the rest; print the dropped ones under deferred: in the taxonomy block so they can be loaded on demand later.

de-dup rules (avoid loading two skills for the same evidence):

• CDN banner alone (Cloudflare/Akamai/Fastly) is not a platform match — it fingerprints the edge, not the app. do not load a skill for it; note it for hunt-cache-poison / hunt-http-smuggling, which the mode set already carries.
• amazonaws / azure / googleapis in a header/origin → cloud-iam-deep. the same string found as a leaked key/JSON in a JS bundle or APK → still cloud-iam-deep, but flag it as a live-credential lead (higher priority, tier 3 becomes tier 1 for that host).
• a framework marker (__NEXT_DATA__, laravel_session) and a generic class signal (?redirect=, Access-Control-Allow-Origin) on the same host → load the framework skill (tier 4) and keep the class skill only if budget remains; the WAPT/redteam mode set already loads the common class skills unconditionally.

step 2 — load skill set

invoke each skill in order via the Skill tool.

mode=redteam

always-on (load first):

redteam-mindset
mid-engagement-ir-detection

platform (load second, conditional on fingerprint matches from step 1):

okta-attack
m365-entra-attack
enterprise-vpn-attack
vmware-vcenter-attack
cloud-iam-deep
supply-chain-attack-recon
apk-redteam-pipeline

high-impact hunt-* set (load third):

hunt-rce
hunt-sqli
hunt-ssrf
hunt-ato
hunt-auth-bypass
hunt-saml
hunt-oauth
hunt-mfa-bypass
hunt-file-upload
hunt-http-smuggling
hunt-cloud-misconfig
hunt-sharepoint
hunt-aspnet

report format: redteam-report-template (subject / observations / description / impact / recommendation / poc).

mode=wapt

always-on:

bb-methodology
security-arsenal
triage-validation

full hunt-* set (all OWASP-relevant):

hunt-xss             hunt-sqli            hunt-ssrf            hunt-idor
hunt-csrf            hunt-xxe             hunt-rce             hunt-graphql
hunt-oauth           hunt-saml            hunt-mfa-bypass      hunt-auth-bypass
hunt-ato             hunt-file-upload     hunt-business-logic  hunt-race-condition
hunt-llm-ai          hunt-api-misconfig   hunt-ssti            hunt-cache-poison
hunt-http-smuggling  hunt-subdomain       hunt-cloud-misconfig hunt-misc
hunt-aspnet          hunt-sharepoint      hunt-ntlm-info
hunt-lfi             hunt-nosqli          hunt-deserialization
hunt-cors            hunt-host-header     hunt-open-redirect
hunt-brute-force     hunt-session         hunt-ldap
hunt-nextjs          hunt-nodejs          hunt-dom
hunt-websocket       hunt-grpc            hunt-laravel
hunt-springboot      hunt-k8s             hunt-cicd
hunt-source-leak     hunt-tls-network

report format: report-writing (bugcrowd-reporting if the target is on bugcrowd).

box=greybox: creds already captured by /hunt, available in session memory.

do not fan out across the authenticated hunt-* set until the creds are validated. /hunt only prompts for and stores creds (commands/hunt.md) — it does not confirm they work. firing every authenticated test with dead, MFA-gated, or wrong-role creds wastes the whole run and produces false "no auth surface" conclusions. run a single low-cost auth preflight first:

# session-cookie creds: one authenticated GET against an identity echo endpoint
curl -sS -m 12 -b "$SESSION_COOKIE" "https://$TARGET/api/me" -w '\n%{http_code}\n'
#   200 + your username/email  → live session, role visible in body
#   401/403                    → dead or insufficient — STOP, re-auth

# bearer/JWT creds: same probe with Authorization
curl -sS -m 12 -H "Authorization: Bearer $TOKEN" \
  "https://$TARGET/api/me" -w '\n%{http_code}\n'

# raw user/pass: drive the real login flow once, capture Set-Cookie, then echo
#   watch for an MFA / step-up challenge in the response — if present, the creds
#   alone do not yield an authenticated session (see memory: operator-capability)

confirm three things from the preflight, and record them for the hunt-* skills:

1. live — auth probe returns 200, not 401/403.
2. role/privilege — the /api/me (or equivalent) body shows the expected role/tenant/scopes. IDOR and authz tests need a known baseline identity; a silently-admin or silently-readonly cred skews every authz finding.
3. not MFA-gated — login did not stop at a 2fa/step-up challenge. if it did, you hold creds but not a session — default to least capability and confirm with the operator before claiming authenticated reach.

if the preflight fails, do not silently continue as blackbox — surface "greybox creds did not validate (HTTP {code} / MFA challenge)" so the operator can re-supply. only after a clean preflight: apply the validated session to every authenticated test.

step 3 — taxonomy print (once, at session start)

emit a deterministic block. plain text, lowercase, colon-delimited, no decoration.

mode=redteam

loaded for red team: {N} skills
  mindset:    redteam-mindset
  platform:   {fingerprint-matched skills (<=8, tier order), or "none detected"}
  deferred:   {platform skills past the 8-cap, or omit line if none}
  auth:       hunt-ato, hunt-auth-bypass, hunt-saml, hunt-oauth, hunt-mfa-bypass
  inj:        hunt-rce, hunt-sqli, hunt-ssrf, hunt-file-upload
  infra:      hunt-http-smuggling, hunt-cloud-misconfig
  stack:      hunt-sharepoint, hunt-aspnet
  ir:         mid-engagement-ir-detection

mode=wapt

loaded for wapt ({blackbox|greybox}): {N} skills
  inj:        hunt-xss, hunt-sqli, hunt-ssrf, hunt-rce, hunt-xxe, hunt-ssti, hunt-file-upload
  authz:      hunt-idor, hunt-auth-bypass, hunt-ato
  auth:       hunt-oauth, hunt-saml, hunt-mfa-bypass
  api:        hunt-graphql, hunt-api-misconfig
  logic:      hunt-business-logic, hunt-race-condition
  infra:      hunt-http-smuggling, hunt-cache-poison
  recon:      hunt-subdomain
  cloud:      hunt-cloud-misconfig
  ai:         hunt-llm-ai
  stack:      hunt-aspnet, hunt-sharepoint, hunt-ntlm-info
  misc:       hunt-misc, hunt-csrf
  reporting:  bb-methodology, security-arsenal, triage-validation

subagent scope inheritance

if any part of the hunt is delegated to subagents, scope does not inherit implicitly. every subagent prompt must carry:

1. the authorized host list, verbatim, as data. not "the target estate", not "*.target.com" — the explicit list. a subagent cannot infer the boundary.
2. the discovered-host rule: hosts found mid-run (via CT logs, CSP headers, JS bundles, CNAME chains, error messages) are report-only. resolve DNS, record, hand back. never probe, never write, until the operator re-authorizes.
3. a deny-list of action-executing endpoints, applied BEFORE any allow-list. deny by verb-in-name first: refund, settle, payout, transfer, adjust, disburse, create, update, delete, rotate, reset, send, initiate, generate, process. only then allow read-shaped names. order matters — a path like refund/batch/status matches the read-shaped keyword "status" but is a refund route; an allow-list applied first would probe it.
4. "read-only" spelled out as forbidden verbs, not as an adjective. "read-only" is routinely interpreted as "don't be destructive", which does not stop an agent sending {} to an endpoint whose name starts with generate* and creating a real record on production.

lesson from an authorized engagement: a subagent was told READ-ONLY and still (a) created a live record on production because it expected {} to return a validation error, and (b) wrote an object to a cloud bucket that was never on the authorized list — one the parent prompt had named only for a DNS check. both were disclosed in the deliverable. the fix is structural: pass scope as data, deny by verb before allowing by verb, and treat every discovered host as out of scope until told otherwise.

step 4 — return control to /hunt

after taxonomy print, hand control back to /hunt for step 3 (sibling delegation) and step 4 (active testing). do not run probes here — this skill only loads context.

privacy

never echo back, log, or persist:

• SOW / scope-of-work / engagement-letter content
• grey box credentials (kept in session memory by /hunt, never written to disk)
• client identifiers in user-level memory

tags:

• web
• security
• pentest allowed-tools: bash, read, write, grep, glob, mcp__burp, curl, webfetch

Related Skills & Chains

• bb-methodology — When PART 0 mode confirmation completes. Workflow primitive: bb-methodology confirms engagement type (red team vs WAPT vs bug bounty); the answer feeds directly into this skill's mode=redteam / mode=wapt invocation.
• redteam-mindset + mid-engagement-ir-detection — When mode=redteam is loaded. Workflow primitive: these are the always-on skills loaded first by step 2 of the redteam flow before any platform skill or hunt-* skill.
• okta-attack / m365-entra-attack / enterprise-vpn-attack / vmware-vcenter-attack / cloud-iam-deep / supply-chain-attack-recon / apk-redteam-pipeline — When fingerprint signals match. Workflow primitive: step 1's curl fingerprint scan against recon/<target>/live-hosts.txt maps banner / domain signals to one or more of these platform skills.
• hunt-rce / hunt-sqli / hunt-ssrf / hunt-ato / *all other hunt- skills** — When the mode-specific skill set is being printed. Workflow primitive: this skill is the loader; it names the hunt-* skills but does not run probes — actual hunting happens after step 4 returns control to /hunt`.
• report-writing vs redteam-report-template — When the taxonomy print specifies the report format. Workflow primitive: mode=wapt ends with report-writing as the deliverable format; mode=redteam ends with redteam-report-template instead.

When to Use

• Target presents indicators of the vulnerability class this skill covers
• Fingerprint or recon indicates the relevant technology stack is in use
• Authorized testing scope covers the target endpoint or component
• Findings need to be validated through this skill's methodology

When NOT to Use

• Target is clearly outside this skill's scope (refer to related skills)
• No authorization for testing
• Need a different category of testing (use related skills)