Cisco Talos researchers release open-source framework that spots LLM-driven malware operating

without human input Cisco Talos researchers this week published an open-source framework called CAIRN — short for Cognitive Artifact Intelligence Research Network — built to classify and analyze malware that incorporates artificial-intelligence components. The tool has already yielded a first discovery: a Windows malware family dubbed CLOSEDQUORUM that runs a fully autonomous command-and-control loop, pulling instructions from four large language models simultaneously and reaching consensus among them with no human input mechanism.

CAIRN was developed by Ryan Fetterman, a security researcher at Cisco Talos. It focuses on the artifacts left behind by AI integration — digital fingerprints in the metadata of malicious files. The framework tags those characteristics, assigns each sample a unique identifier, and analyzes every artifact in the context of the entire library to cluster samples by shared traits and surface trends. Since adopting the tool, Fetterman says he has uncovered roughly twenty additional AI-integrated malware samples, well beyond the nine documented families he found in an earlier retrospective survey.

CLOSEDQUORUM consults DeepSeek, Qwen, Mistral and Google Gemini to decide its next action on a target system. If one service is unreachable, the malware simply queries the others, a redundancy that makes the control loop hermetically sealed. Researchers traced connections to credit-card fraud forums from 2025; the malware is designed to harvest credentials and cryptocurrency. The developers remain unidentified.

The find contrasts with expectations set a year ago. In July 2025, Ukraine's CERT-UA — the national computer emergency response team — warned of a phishing campaign that used a malware strain called LAMEHUG, which communicated with Qwen2.5-Coder-32B-Instruct via the Hugging Face API. Fetterman anticipated an explosion of AI-driven malware at the time, but a year later he found the landscape still relatively sparse. "The landscape is more complex and diverse than what has been publicly reported," he says now, adding that the activity remains mostly experimental on the attackers' side.

CAIRN is now available to the defense community as an early-warning tool for tracking the evolution of autonomous malware. The ability to follow the fingerprints of LLM integration in malicious code could buy critical time — provided attackers are still experimenting rather than standardizing the approach.