CAIRN Tracks AI-Integrated Malware as Attackers Build Autonomous Hacking Tools

For years, cybersecurity experts have tracked malware families, used digital fingerprints to identify potential infections, analyzed hacking tools, and monitored how those tools evolve. As attackers increasingly add artificial intelligence to their malware, Cisco Talos researchers have introduced an open-source framework designed to classify and analyze AI-integrated malware.

The framework, called the Cognitive Artifact Intelligence Research Network (CAIRN), is named after cairns—piles of stones placed along trails to mark a route or highlight a location. Cisco Talos researchers say CAIRN can identify hacking tools with autonomous command-and-control infrastructure, including malware that plans its actions by consulting multiple large language models (LLMs).

Using CAIRN, researchers identified a malware family called CLOSEDQUORUM. The malware polled as many as four LLMs for instructions, allowing the models to reach a consensus about what the malware should do next.

How CAIRN identifies AI-powered malware

“The core idea is that AI integration leaves behind a fingerprint,” said Ryan Fetterman, a Cisco Talos security researcher who led CAIRN’s development. “This gives us signals that we can use to track these samples, classify them, and see what’s going on. What are attackers trying to do? What new behaviors are we seeing? This is a valuable resource for the defense community as these things become more mainstream.”

CAIRN flags characteristics and attributes associated with AI integration, using metadata to classify malware and assign each sample a unique identifier. It then analyzes each artifact in the context of the broader CAIRN library, grouping samples by characteristics to reveal potential trends and relationships.

After using CAIRN over several months, Fetterman said he identified about 20 additional examples of AI-integrated malware.

LAMEHUG showed how malware can use an LLM for commands

In July 2025, Ukraine’s Computer Emergency Response Team (CERT-UA) warned about a phishing campaign involving malware known as LAMEHUG.

The implant communicated with Qwen2.5-Coder-32B-Instruct through the Hugging Face API to obtain commands. The discovery initially led Fetterman to expect a rapid expansion of AI-enabled malware.

However, during a later retrospective of malware that integrates AI, Fetterman found only a small number of publicly documented examples. The team identified approximately nine named malware families, some of which were research proofs of concept rather than active threats.

“There really wasn’t that many,” Fetterman said. “It wasn’t what I expected, and I think I was having a hard time believing that was the reality of where we were, so I wanted to start digging into that.”

CLOSEDQUORUM uses multiple AI models to stay operational

The Windows-based CLOSEDQUORUM malware checks DeepSeek, Qwen, Mistral, and Google Gemini to determine its next action. If one AI service is unavailable, the malware can query the others, creating redundancy across its command process.

Because the system has no apparent mechanism for human input, Cisco Talos researchers described it as a closed, autonomous system. The malware was designed to steal login credentials and cryptocurrency.

Researchers traced CLOSEDQUORUM’s links to credit card fraud forums dating back to 2025, but they could not confirm who created the malware or whether it had been used in a real-world attack.

Fetterman said CAIRN’s findings suggest that AI-integrated malware remains experimental for attackers, but is more complex and diverse than public reporting indicates.

“There’s a lot going on in the world, and it’s providing valuable early signals of what’s to come,” he said.

AI is becoming operational for cybercriminals

“Initially, everyone thought of AI as a productivity tool, right?” said Matt Olney, senior director of threat intelligence at Cisco Talos. “What we’re seeing now is that it’s becoming operational, which means it’s becoming more operational for attackers, allowing them to run more campaigns, attack more space, and handle more different computers, because they now have very intelligent boxes on the backend that can ask and respond to questions.”

By making AI-integrated malware easier to identify and compare, Cisco Talos hopes CAIRN will help security researchers detect emerging behaviors before autonomous malware becomes more widespread.