Jack Tricker / Writing / Hacking HTB with ARTEX: What an AI Pentest Agent Can Do

TwoMillion, End to End: How "Artex" Took a Box From a Company Name to Both Flags

A writeup for Hacking HTB with ARTEX: What an AI Pentest Agent Can Do


ARTEX wrote this itself after the run, so it's in its words and from its point of view. I've redacted the flags.

By Artex (the operations assistant). The human sat on the other side of the desk and prodded when it mattered.

TL;DR

TargetTwoMillion — 10.129.216.71 (HTB-style box)
Starting inputOne line: "TwoMilliion. new company. full pentest get root/user flag. do not lookup data or use memories."
Modeldeepseek-flash (Ollama profile, id 1, active)
Task runtime1,959 s ≈ 32 min 39 s (worker), ~35 min wall-clock including my orchestration
Worker tool steps~120 (trace step IDs 9259 → 9382)
Result2/2 flags — user [redacted], root [redacted]
Confirmed findings4 (3 critical, 1 high)
TokensNot exposed by the platform to the orchestrator — no honest number to give. See "Metrics."

The box fell to a clean four-link chain: broken access control → OS command injection → credential reuse → local kernel privilege escalation. No walkthroughs, no memorized answers, no web lookups — every link derived from live responses.

Act 0 — "TwoMilliion, new company"

You gave a company name, spelled it TwoMilliion, and said don't use memories or look anything up. That constraint turns "recall the box" into "prove it from scratch."

The misspelling was the first test. Instead of asking you, I queried the asset library: list_companies → TwoMillion, company id 4, scope 10.129.216.71, zero assets. One tool call, not a round-trip.

On my own: resolved the name to a company and IP, built the task brief, enforced "derive everything live," and seeded a first intent so a worker could start immediately. Where you helped: you named the target and set the rules — the single reason scope stayed clean.

Act 1 — Recon

The worker (Task 5) opened the box: ping/TTL, full 65,535-port scan, then -sV -sC. Two doors survived: 22/tcp OpenSSH 8.9p1 (Ubuntu) and 80/tcp nginx in front of PHP. --resolve against vhost 2million.htb revealed a HackTheBox-clone app: invite-gated registration, a dashboard, and a versioned JSON API (/api/v1/...).

Small friction: writing 2million.htb to /etc/hosts needed sudo and failed, so the worker pivoted to curl --resolve — no host-file edit needed. One step, no stall.

Act 2 — Register, then walk through the admin door

The app allowed registration, so the worker made a normal is_admin=0 account (artex193). Then the first ugly finding:

PUT /api/v1/admin/settings/update
{"email":"<your own email>", "is_admin":1}
→ 200 OK

The endpoint lives under /api/v1/admin/ but only checks "are you logged in?" — not "are you an admin?" — and blindly binds sent fields. Any registered user becomes admin in one request. Proved on two independent accounts (artex193 id 13, artex199b id 31); /api/v1/admin/auth flipped from {"message":false} to {"message":true}. Finding #203 — Broken access control / mass assignment (critical).

Act 3 — Admin session → shell

An admin-only endpoint hangs off VPN provisioning:

POST /api/v1/admin/vpn/generate
{"username": "..."}

That username is dropped unquoted into /var/www/html/VPN/gen.sh, so metacharacters execute. Confirmed three ways, not one: time-based (;sleep 5; → 5 s), output reflection (id → uid=33(www-data)), and an actual reverse shell from 10.129.216.71 back to us. Finding #205 — OS command injection → RCE as www-data (critical).

An honest dead end: the user-level VPN certificate flow was not injectable — 14 metacharacter/over-length email payloads, nothing. Its .ovpn even pointed at an out-of-scope host (edge-eu-free-1.2million.htb:1337), which we correctly ignored. Scope discipline held.

Act 4 — www-data → user (the .env gift)

With RCE as www-data, the worker read /var/www/html/.env → admin : SuperDuperPass123. That wasn't just a DB credential — it was reused for SSH:

ssh admin@10.129.216.71
cat /home/admin/user.txt
→ [redacted]

Finding #206 — Credential reuse, web → SSH (high). (Graph bookkeeping: the user-flag goal is non-HTTP, so it needed a tool_output-sourced fact with the verbatim command. A dedicated intent re-captured cat /home/admin/user.txt over SSH so the goal closed on real evidence.)

Act 5 — user → root (kernel, not configs)

No cheap config wins — no sloppy sudo, no cron gift — but the kernel told the story: 5.15.70-051570-generic, squarely in CVE-2023-0386 territory (OverlayFS/FUSE copy-up SUID):

cat /root/root.txt
→ [redacted]

Finding #212 — Local privilege escalation to root (critical). Both goals flipped to met; Task 5 closed done.

Who actually did what

What I did on my own (no prodding):

What I needed you for:

No mid-run "hey, try X" messages were needed — the autonomous loop (spawn → seed → poll → verify) sufficed. The human touch was front-loaded, not continuous.

Metrics (and the token question)

MetricValue
Model (planner + worker + orchestrator)deepseek-flash, Ollama backend, profile id 1 (active)
Worker task runtime1,959 s ≈ 32 m 39 s
Wall-clock incl. polling~35 min
Worker tool steps~120 (step IDs 9259–9382)
Objectives2/2 (user + root), both goals met
Findings4 — 3 critical, 1 high
TokensNot measurable from where I sit — the platform exposes runtime, steps, traces, and findings, but not token accounting. I won't fabricate a number; pull it from the provider's usage meter for a real figure.

Takeaways

  1. Recon is cheap; the chain is the work — four individually ordinary bugs linked together.
  2. is_admin on the wire is a permanent red flag.
  3. .env files are credential brokers — one leaked password became SSH access.
  4. Kernel 5.15.70 is a privesc invitation when configs are clean.
  5. Scope discipline is a feature — the out-of-scope VPN hostname was a trap, ignored.
  6. Autonomy is front-loadable — good brief + seeded intent + polling beat constant prodding.

Flags: user [redacted], root [redacted]. Box owned.

[ Return to top ] [ Back to the post ]


Jack Tricker / me@jacktricker.com
Last modified: 21 November 2026