Welcome back, aspiring cyberwarriors!

Some of you have already seen the headlines about malware calling out to AI APIs. Recently, Talos found a Windows implant called CLOSEDQUORUM that leaves its tactical command and control decisions to a panel of commercial LLMs while the operator sleeps. Malware like that doesn’t leave the traces we’re used to hunting. Now you find prompt templates, provider endpoints, API key prefixes, jailbreak wording and sometimes text written to talk an AI sandbox out of flagging the file. Talos calls these cognitive artifacts and developers who wire an LLM into their tooling leave some behind.

What is CAIRN

Cisco Talos built a free research tool called CAIRN. The name is a metaphor. A cairn is a pile of stones hikers leave on a trail so the next person can see where to go. CAIRN looks for the same kind of marker inside malware that talks to AI. It works from metadata, which means you never download the sample and you never run it. Keep in mind, it needs a VirusTotal API. It sorts what it finds into three tiers, from a simple AI string match, to behavior and families. It also clusters samples that look alike even when they don’t share an obvious string. 

CAIRN is a research tool and not a live threat feed. It shows us where hackers are heading. 

Let’s set it up and cover some basic commands. 

Setting Up

We start by copying the repository and installing the python libraries.

kali > git clone https://github.com/Cisco-Talos/Cognitive-Artifact-Intelligence-Research-Network.git

kali > cd Cognitive-Artifact-Intelligence-Research-Network

kali > python -m venv .venv
kali > source .venv/bin/activate          

kali > pip install -e ".[dev]"

The env file needs VirusTotal API.

kali > cp .env.example .env
kali > nano .env

Once it’s there, you can start with some basic commands.

Basic Commands

The first thing we do is open the YARA file that ships with CAIRN and check that it is readable and put together the way the tool expects. If it prints a rule count and says the pack is valid, the rulers are loaded. You run it after install or after you edit a rule, so you know the next hunt will not fail because the rule file is broken.

kali > cairn validate-rules

Next we print the list of saved VirusTotal searches that come with CAIRN. We will use these names in the following command. The command only reads a config file on disk.

kali > cairn filters

Each name on that list is a hunt idea you can use, like Python files that mention an LLM library, or files that mention a hosted API.

kali > cairn pull --filter python-ai-scripts --limit 10

Next, you can summarize the findings. It reads only your database and tells you how many files you got and what rules were in use. 

kali > cairn summarize

We can then open the map.

kali > cairn explore

Here is how the clusters look like. Samples are grouped into clusters even when they share no obvious string overlap. These are semantic discoveries that analyze the meaning. 

Malware That Uses AI

Let’s take a look at the new developments in malware that were published by Cisco Talos. Their team was analyzing CLOSEDQUORUM.

They say it’s the first publicly documented Windows implant to apply this model to tactical command and control (C2). After it’s deployed, it uses AI models to decide what to do next and then executes their collective decision to harvest user credentials and crypto wallets. It doesn’t need a human to be ruling over the process, as the AI does everything. 

Traditional C2 architecture needs the hackers to have server infrastructure. That infrastructure is attributable and blockable. Instead of this, CLOSEDQUORUM calls up to four commercial LLM provider endpoints used by thousands of legitimate applications daily.

Here is how it looks decompiled. 

You see the providers, models and prompts. 

Hackers also may give detailed technical prompt templates. Here TARGET: %s is substituted at runtime. 

The response is then deserialized into a Go struct. 

Structure of the returned model’s decision.

Summary

Malware that calls out to AI is already here. There’s no server to attribute or block, since the traffic goes to provider endpoints that legitimate apps hit all day. Now we also hunt for cognitive artifacts like prompt templates, provider endpoints, API key prefixes and jailbreak wording. CAIRN helps with that by working from VirusTotal metadata, so you never download or run the sample. It groups findings by AI strings, behavior and families and clusters similar samples by meaning. It shows us where hackers are heading, and they’re heading toward AI doing the thinking for them.

If you like Malware Analysis and want to improve your skills in this field, sign up for our Reverse Engineering Malware training.

The post Malware Analysis: Analyzing AI-Backed Malware with CAIRN first appeared on Hackers Arise.