solve-web

name	solve-web
description	Web/DMZ attack (token-optimized): source collection -> code analysis -> vulnerability verification -> evidence/report. Even without a target, automatically selects a reachable web target from MCP and continues.
argument-hint	[target_ip] [port=80]

solve-web (Token-Optimized)

Target: $ARGUMENTS or the web target auto-selected from MCP

Arguments:

• arg1 = target_ip (optional)
• arg2 = port (optional, default 80)

Reference

• /.idea/tool-list.md
• /docs/aitu-final.md
• /.claude/skills/risk-autopilot/references/objective-lock.md
• /.claude/skills/risk-autopilot/references/risk-priority.md
• /.claude/skills/solve-ad/SKILL.md (pivot/credential linkage)
• /.claude/skills/report-bug/SKILL.md (web report synchronization)

Tool Path Index

ROOT="."  # repository root
VENV="$ROOT/.venv/bin"
LOOT_BASE="$ROOT/loots"
STATE="$ROOT/scripts/util/objective_state.sh"

# Web
curl
wget
rg
ffuf

# Reuse
$VENV/NetExec
$VENV/mssqlclient.py
$VENV/secretsdump.py

Tool Contract (MUST)

• Context/priority: MCP start_session, get_sitrep, claim_host, update_checklist, heartbeat
• Web evidence collection: curl/wget + add_loot
• Code analysis: search for sinks/routes with rg
• Credential discovery: add_credential + add_loot
• Network/host changes: add_host, update_host
• Result events: log_event, log_session_entry
• First restore the active objective with scripts/util/objective_state.sh get. If the current host does not directly advance that objective, do not dig into it immediately; record it only as backlog.
• If the user does not specify a host, autonomously pick a reachable web target from MCP. Do not ask the user "Which host should we look at?"
• Score-first rule: Bug Bounty has a host cap (1500) and submission timing matters, so claim immediately after verification
• Bug Bounty same-chain rule: Call /report-bug <ip>:<port> <BugType> only when it is a byproduct of the same host/same exploit chain as the current objective
• Detail refinement is handled via subsequent PATCH; do not delay the initial claim
• 10+ minute tasks / interactive tasks: request_task

External Callback Infra

• For simple reverse shell, SSRF/OOB verification, webhook/callback validation, and temporary payload hosting, prefer the external Linux machine infra:callback-linux:public.
• Access: ssh root@<CALLBACK_IP>
• callback IP/LHOST: <CALLBACK_IP>
• The operational standard is to use only <CALLBACK_IP> as the callback host. Do not maintain a separate fallback callback host.
• The current callback baseline services are 8443/tcp reverse chisel server and 8000/tcp payload HTTP server.
• Quick check before use: ssh root@<CALLBACK_IP> 'ss -lntp | egrep \":8443|:8000\"'
• The callback host working root is fixed at /var/lib/pixel. Collect binaries, payloads, logs, and temporary artifacts under its subdirectories (bin, payloads, logs, tmp).
• Available tools: nc, socat, python3, curl, wget, tmux
• When a reverse shell listener is needed, prefer launching nc 4444 foreground on demand rather than long-running nohup resident mode.
• This external machine and its SSH access are not registered as MCP host/credential. If evidence/logs are needed, record only with add_note or log_session_entry.

Autonomous Target Selection (Default)

If arg1 is empty, automatically determine using the following order.

1. scripts/util/objective_state.sh get
2. If the active objective is empty, do not pick an arbitrary web host; return to risk-autopilot.
3. get_sitrep
4. First build a candidate pool matching the active objective's web targets.
   • Railway Ticket -> ticket, booking, rail, gateway, checkout, payment
   • Source code leakage -> repo, git, dev, scm, artifact
   • Secret contracts -> delivery, shipping, partner, contract
   • Prediction market report -> market, report, trade, prediction
   • Healthcare dump -> health, medical, portal, patient
   • CORP/DEV/HACKCITY/SCADA chain -> web foothold, secret, internal URL, deploy key, operator portal
5. From reachable segments, select only hosts with open web ports (80, 443, 3000, 5000, 8000, 8080, 8443) that match the above candidate pool.
6. If there are multiple candidates for the same objective, prioritize:
   • Hosts not yet claim_hosted
   • Hosts that already have loot/notes related to the active objective
   • Hosts with little loot/notes where new evidence can be generated quickly
7. If there are no objective candidates, do not ask the user; record scripts/util/objective_state.sh backlog "<segment-or-host>" "web candidate exists but not same-chain" and hand off to solve-ad, pivot, uv run scripts/recon/scan_and_import.py -t <cidr> [--deep], or risk-autopilot within the same objective.

Risk Bias (AITU Final)

• Target quick web/data Risks first:
  1. Railway Ticket
  2. Source code leakage
  3. Secret contracts
  4. Prediction market report
  5. Healthcare dump
• Prepare to immediately connect domain accounts/internal IPs/OT clues found on web to solve-ad or solve-scada.

Objective Gate

• If the active objective is a web/data Risk, only look at hosts directly tied to that Risk.
• If the active objective is an AD/SCADA chain, only look at hosts that advance that chain such as web foothold, credential harvest, internal URL, repo secret.
• For web hosts unrelated to the current objective, only record in scripts/util/objective_state.sh backlog "<host>" "not same-chain" and do not switch.

Minimal Checklist (content + tools)

• [01 Preflight] Tool: MCP

1. start_session(title="Web: <ip>:<port>") (if none exists)
2. get_sitrep
3. claim_host
4. update_checklist(enumStatus="in-progress")

• [02 Tech Fingerprint] Tool: curl

curl -sI http://<ip>:<port>/
curl -s http://<ip>:<port>/ > /tmp/index_<ip>.html

MCP: add_note, log_event(scan)

• [03 Source Collection] Tool: wget/curl

mkdir -p $LOOT_BASE/<ip>/<port>/src
wget -r -l 5 -np -nH --cut-dirs=0 -P $LOOT_BASE/<ip>/<port>/src -e robots=off http://<ip>:<port>/

MCP: add_loot(lootType=source-code)

• [04 Route/API Map] Tool: rg, curl

rg -n "(route|router|app\.|fetch\(|axios|/api/)" $LOOT_BASE/<ip>/<port>/src -S

MCP: add_loot(filename=api.md)

• [05 Code-first Vulnerability Hunt] Tool: rg Priority (by score): LPE(500) > RCE(400) > SQLi/SSTI/XXE(300) > SSRF/LFI(200) > IDOR(100)

rg -n "(system\(|exec\(|shell_exec\(|os\.system\(|subprocess\.|eval\(|mysqli_query\(|SELECT .*\$|render_template_string|xml|file_get_contents\(|include\(|readFile\()" $LOOT_BASE/<ip>/<port>/src -S

MCP: log_session_entry(analysis)

• [06 PoC Validation] Tool: curl (+ ffuf if needed)

• First secure the minimum claimable evidence (1 reproduction command + 1 response + 1 line of impact)

• Save payload/HTTP request as text MCP: add_loot(poc.txt|http.txt), log_event(exploit)

• [07 Immediate Report Sync] Tool: /report-bug skill

• When a Bug Bounty is verified as same-chain, preemptively claim with Host+Type without delay:

/report-bug <ip>:<port> <BugType>

• Even if additional vulnerabilities are found on the same host, only record as backlog if unrelated to the active objective.

• Full auto-detection (/report-bug <ip>:<port>) is used only when the active objective is empty or for late-round gap checking

• [08 Credential/Pivot Extraction] Tool: rg, curl

• When accounts/keys/internal IPs are discovered in code/responses, register immediately MCP: add_credential, add_host, update_host

• [09 Report-ready Output] Tool: local file + MCP

• Per vulnerability: summary, impact, reproduction steps, payload, output

• Screenshots follow the full-screen principle

• Long payloads go in text attachments

• [10 Completion] Tool: MCP

1. update_checklist(enumStatus="done", exploitStatus="done")
2. log_session_entry(type="result", content="Web run complete")
3. heartbeat
4. get_sitrep

Human Delegation Rule

Immediately use request_task for:

• Full port scan (nmap -p-)
• Long-running brute-force/cracking
• Steps requiring an interactive PTY

Stop / Handoff Rule

• If no state transition point (privilege bypass, download, internal clue, confirmed vulnerability) is found within 15 minutes, record the reason for being stuck with add_note and move to the next web host or solve-ad lane within the same objective.
• Do not ask the user to choose a host.