A significant shift is underway in the cybersecurity threat landscape as malicious software transitions from human-operated campaigns to fully autonomous operations powered by artificial intelligence. Security researchers have uncovered a sophisticated Windows hacking tool capable of navigating compromised machines without receiving commands from a human handler. Named CLOSEDQUORUM, the software determines its operational steps inside a target network by consulting four independent large language models, executing moves only after establishing a consensus among the AI platforms. The discovery was made possible through a newly engineered detection framework known as the Cognitive Artifact Intelligence Research Network, or CAIRN.
Tracking Digital Fingerprints with CAIRN
The monitoring framework draws its name from cairns, the balanced stone markers traditionally constructed by trail hikers to guide paths and highlight notable terrain features. Ryan Fetterman, the Cisco Talos security researcher who spearheaded the construction of CAIRN, explains that integrating artificial intelligence into software inherently leaves behind detectable attributes, similar to physical fingerprints. CAIRN scans the metadata of suspicious code files for these distinctive signatures, assigning each sample a unique identifier to systematically classify and cross-reference it against an expansive artifact repository.
According to Fetterman, this automated categorization provides the defensive cybersecurity community with early intelligence regarding emerging adversary techniques. By understanding what malicious actors are attempting and recognizing emergent behaviors before they proliferate, defenders gain critical foresight. Over recent months of operating CAIRN, researchers have tracked approximately 20 previously uncataloged instances of malware incorporating AI functionality, demonstrating that experimental adoption among cybercriminals is broader than early evidence suggested.
Evolution from Early Phishing to Multi-Model Networks
The trajectory toward autonomous malicious code traces back to earlier experimental deployments. In July 2025, the Ukrainian cyber emergency response team CERT-UA issued an alert regarding a targeted phishing wave powered by a malware variant designated as LAMEHUG. That particular implant relied on Hugging Face API connections to fetch operational commands directly from an LLM known as Qwen2.5-Coder-32B-Instruct. While the incident prompted industry expectations of an immediate, sweeping surge in AI-driven malware families, subsequent tracking revealed a slower public emergence.
During a retrospective analysis conducted this summer, Fetterman initially located only about nine documented malware families that incorporated artificial intelligence, with multiple instances representing benign academic proofs of concept rather than operational threats. That surprising scarcity prompted the deeper technical investigation that culminated in CAIRN, ultimately demonstrating that the malicious development sphere is far more active and multifaceted than conventional reporting indicated.
Autonomous Consensus Strategy of CLOSEDQUORUM
The architecture of the CLOSEDQUORUM malware demonstrates an unprecedented level of operational redundancy. Operating on Windows systems, the malware queries DeepSeek, Qwen, Mistral, and Google Gemini simultaneously to agree upon each sequential task within a compromised machine. If any individual AI endpoint becomes unreachable or unresponsive, the implant continues querying the remaining providers, maintaining a resilient command loop that operates entirely without a mechanism for manual human input.
Investigators traced historical links connecting elements of the software to underground cybercrime forums focused on payment card fraud dating back to 2025. The code is structured primarily to harvest sensitive user login credentials and exfiltrate cryptocurrency assets. However, researchers noted that attribution remains unconfirmed, leaving the developer identity unresolved, while definitive proof of active deployments against real-world enterprise environments has not yet been established.
Operationalization of Threat Capabilities
Matt Olney, senior director of threat intelligence at Cisco Talos, emphasized that artificial intelligence has moved well past its initial reputation as a simple efficiency mechanism for coding or document writing. Malicious developers are now fully operationalizing these capabilities directly inside attack workflows. By embedding intelligent response engines capable of reasoning through situational queries in the background, threat actors can dramatically broaden the scope of their campaigns, manage disparate computer architectures simultaneously, and escalate attack frequency without increasing their operational personnel.
