| name | pentest-engagement |
| description | Run a professional penetration engagement OR a network vulnerability scan from a scope. WEB mode (apex domains / app URLs) — mandatory surface expansion, systematic OWASP attack-class coverage, reversible active exploitation, authoritative validation, Transilience PDF. NETWORK mode (a list of IPs/CIDRs, e.g. 1500 hosts) — machine-prudent BATCHED tool-first nmap sweep (host discovery + bounded common+less-common port/service scan + CVE surfacing) with a uniform per-IP output tree, then bounded deep-dives on the highest-value hosts. The general (non-CTF) analogue of hackthebox/htb-solve. |
| context | fork |
Pentest Engagement
Orchestrates a scoped pentest end-to-end via the pentest-engagement workflow. It is the breadth-complete, coverage-gated counterpart to the flag-shaped htb-solve — same engine (coordinator-loop, now with interleaved per-finding validation built into the loop), but driven by an attack-class coverage matrix instead of a flag, with surface expansion and root-cause severity baked in.
When to use
A real (non-CTF) engagement defined by a scope — either:
- WEB — web / API / cloud apps defined by apex domains / asset URLs, or
- NETWORK — a list of IPs / CIDRs / ranges (e.g. 1500 hosts) to scan for live services and vulnerabilities.
The workflow auto-detects the mode in Setup (engagement_kind): predominantly IPs/CIDRs → network; apex domains / app URLs → web. For HackTheBox/CTF use hackthebox (htb-solve) instead.
Run it
WEB — from a scope file or inline:
Workflow('pentest-engagement', { scope_file: 'projects/pentest/<engagement>-scope.md' })
Workflow('pentest-engagement', { scope: { engagement_name, apex_domains:[], assets:[...], creds_env:[...], roe, business_tier } })
NETWORK — inline IP/CIDR list or a scope file containing one (a plain newline list of IPs/CIDRs is accepted):
Workflow('pentest-engagement', { targets: ['10.0.0.0/24', '192.0.2.0/24', '198.51.100.7'] })
Workflow('pentest-engagement', { scope_file: 'projects/pentest/<engagement>-ips.txt', scan_profile: 'standard' })
Options (shared): maxConcurrent (default = prudent, derived from CPU cores — ~half the cores, capped 2–8; never hundreds/thousands of parallel tasks), dryRun, max_experiments, business_tier, report (default true).
Options (network): scan_profile light (bounded 1-1024 + curated less-common, for large/fast sweeps) / standard (full-range -p- on every reachable host, DEFAULT; two-stage SYN→-sV on found-open ports, host-count-guarded, message-bus + non-443-TLS aware) / full (-p- + bounded UDP); udp (top-50 UDP; off for light/standard, auto-on for full); slice_size (hosts per scan worker, auto ≈64 IP-list / 2 CIDR-heavy); deepen_top (deep-dive the N highest-value hosts, default 10, 0 to skip); geo_vantages (≤2 gcp zones for the 2nd-vantage allowlist re-probe; overrides the US+EU default), auto_provision (default true; false = detect+flag only, no cloud spend). On a source-IP/geo-allowlist signature the workflow auto-provisions a 2nd-geography vantage and re-probes the filtered hosts before concluding "no surface."
Write the scope file per reference/scope-file-format.md. Credentials are referenced by env-var name only and read from the repo .env via python3 tools/env-reader.py — never inline secret values.
Phases (what the workflow does)
- Setup —
env-reader creds, parse scope, classify kind (web|network), read CPU cores → prudent parallel-task cap, OUTPUT_DIR = projects/pentest/<date>_<engagement>/, STARTED Slack (gated).
- Expand (WEB, the #1 fix) — MANDATORY CT-log / passive-DNS / origin-discovery across every in-scope apex (
crt.sh, certspotter, subfinder, origin-discovery for CDN/WAF-fronted hosts). Scope = the discovered surface, not the handoff. Builds the per-asset work list + seeds each asset's coverage matrix.
Scan (NETWORK, replaces Expand) — slice the IP/CIDR set into machine-prudent batches; one nmap worker per slice (agents scale with slices ≈ dozens, never with IP count) runs the SAME pipeline: host discovery (reachability is unknown) → bounded common+less-common port/service scan → CVE surfacing (nmap --script vulners, nuclei, each CVE-ID enriched via tools/nvd-lookup.py) → writes a uniform per-IP tree hosts/<ip>/{recon,host.json,findings} + a merged recon/inventory/.
- Assess (single interleaved stage — no separate downstream validation pass) — WEB: each asset →
coordinator-loop (coverage mode), which validates each candidate the instant it is materialized on fresh blind agents (strict per-finding cure/drop loop) before search continues. NETWORK: bounded coordinator-loop deep-dives on only the deepen_top highest-value hosts, same interleaved per-finding validation (everything else is the uniform tool-scan, not a per-host agent). Coverage-by-VALID: a class is covered only by a VALID/REPAIRED finding, a justified N/A, or a genuine negative — a class whose candidates were all rejected/dropped stays pending and search keeps going.
- Correlate —
attack-path-stitcher + risk-prioritiser across all validated findings → ranked org roadmap.
- Report (deterministic — no agent authors the report) — JS hands the resolved engagement block + exact commands to ONE finalize runner:
tools/report_data_build.py merges the namespaced interim finding-JSONs into the canonical (the sole-owner assembly), then the renderer runs — → the canonical PDF skill, → (or a Markdown fallback). JS then : assembled ∧ (transilience: ∧ ∧ the provenance tag), retry-once → . Only / findings appear ( — is confirmed-only by construction); () and uncured () never appear and there is no gaps/assurance section. The finalize runner also runs (swept-host tail) + over the whole engagement, writes , and the deliverable includes a deterministic section (surface-unit × attack-class). : it requires the report to assemble+render AND the coverage gate to report (every applicable cell covered) — for BOTH web and network (network additionally requires scan-completion). Any untested applicable cell → / .
Post-run (main loop, outside the workflow): because coordinators must not call AskUserQuestion, the invoking agent shows the returned summary and asks whether to post the deliverable_zip to Slack; on yes, python3 tools/slack-send.py --channel "$PENTEST_SLACK_CHANNEL_ID" (gated on a successful, COMPLETE engagement).
Report options: report_format transilience (default) | custom; custom_report_cmd (the custom renderer, receives the report_data.json path + reports/ dir); prior_report (a prior PDF or report_data.json — its title/sector/scope are metadata-only, never seeding the work list) + version to mint the cover version + "Supersedes" line.
Determinism (be honest)
The decision layer is a frozen pure-JS computeVerdict (parity-guarded, fixture-pinned), its operands are frozen per engagement via NVD/KEV --cache-dir snapshots (artifacts/nvd-cache/, artifacts/kev-snapshot.json), and the adversarial quorum is raised to 3. Given identical inputs the verdict is provably identical. But a fresh live run is highly reproducible, not 100% — LLM sampling produces the booleans/numbers that feed the verdict, and live-target drift moves the inputs; that ceiling is inherent and stated plainly, not papered over. Byte-identical results are a guarantee 100% only on replay of a frozen evidence set (the Phase-2 replay cache, artifacts/validation-cache/) — that is the sole context in which "same request → same result" is a guarantee rather than a strong tendency.
Boundaries
- Orchestrator only — never run the coordinator loop inline; the coverage/bookkeeping discipline needs the workflow boundary.
- A missing credential is not a global block — the unauthenticated surface is always tested; only a no-reachable-asset scope blocks.
- Reversible own-org/own-tenant writes are authorized by default (create-then-delete is non-destructive); destructive ops, DoS, brute force, and out-of-scope tenants are prohibited (set in RoE).
References
