Personal WebsiteLinkedin

Blurry

Enumeration

nmap -v -A -O -p- -Pn blurry.htb -oN nmap

Port 80

There is a redirect to app.blurry.htb when we hit the port 80, so add the subdomain to /ect/hosts and start the enumeration:

ClearML is a lifecycle management platform for machine learning (ML) and experiments. It provides tools to track, manage and automate machine learning projects, making it easier for data science and engineering teams to manage their experiments, data and models.

Insert any username and login.

Previously we have seen a redirect to app.blurry.htb so we can try to enumerate different subdomains using gobuster:

There are so many functionalities, so starts viewing which projects are already created:

This project highlighted with red is composed by different automated task that runs occasionally:

From here we can extract useful information about the box and a possible user of the machine -> jippity.

Exploitation (user)

Searching online for some ClearML vulns, we will find it:

LogoGitHub - xffsec/CVE-2024-24590-ClearML-RCE-ExploitGitHub

We just need to setup the python venv and install dependencies:

and then we can run the exploit:

We need to set up the ClearML client by copy&paste the API configuration from Settings of our account:

Then run the exploit and obtain the reverse shell as jippity user and the user flag:

Obtain a second cleaner reverse shell using pwncat-cs. This is necessary due to log output inside the first reverse shell:

To mantain an access and avoid to exploit the ClearML every time we can dump SSH key of jippity:

Privilege escalation (root)

There isn't any other interesting users to exploit so go further with root:

The more simple command sudo -l from SSH session or the LinPeas script are our friends:

We can run as root the bash script evaluate_model by passing as a parameter any .pth (PyTorch trained model) file present inside the /models folder.

We have write permission inside the /models directory so:

We need to install torch because it will allow us to create a model containing bash spawning process. This script defines a PyTorch model with malicious intent by overriding the __reduce__ method to execute a shell command when the model is deserialized. The __reduce__ method is a special method used by Python's pickle module to define how an object is reduced to a serialized form. By overriding this method, the script executes a malicious command during deserialization.

Run the script and upload the evil.pth to victim machine under /models.

Now run the sudo command and enjoy the root shell 🎉:

PreviousFreelancerNextEditorial

Last updated