Cisco Talos Uncovers 'CLOSEDQUORUM' Malware Using Autonomous Multi-LLM Decision Engines

Executive Summary

Cisco Talos researchers have recently identified a groundbreaking evolution in malicious software: CLOSEDQUORUM. Unlike traditional malware that relies on hardcoded logic or human-operated command-and-control (C2) servers, CLOSEDQUORUM utilizes an integrated, autonomous decision-making engine powered by multiple Large Language Models (LLMs). This shift marks a transition from AI as a mere productivity tool for attackers to AI as a core component of the attack infrastructure itself.

Threat Analysis

CLOSEDQUORUM is designed to operate with minimal human intervention. By querying multiple AI services—specifically DeepSeek, Qwen, Mistral, and Google Gemini—the malware can adapt its tactics in real-time based on the environment it infects. This redundancy ensures that if one AI service is blocked or unavailable, the malware seamlessly pivots to another, maintaining its operational tempo. The malware focuses primarily on the theft of sensitive credentials and cryptocurrency assets.

Technical Details

The malware was detected using Cisco Talos’s new CAIRN (Cognitive Artifact Intelligence Research Network) framework, which identifies digital fingerprints left by AI-integrated code. CLOSEDQUORUM functions by sending environmental data to the aforementioned LLMs, which then provide instructions on the next malicious step. The malware evaluates the responses from these models for reliability before execution, effectively creating a 'consensus' mechanism that minimizes the risk of triggering security alerts or failing to execute commands. This autonomous loop allows the malware to perform complex tasks like network mapping and privilege escalation without manual input.

Attribution Assessment

While the specific threat actor behind CLOSEDQUORUM remains under investigation, the sophistication of the integration suggests a highly capable group with the resources to maintain persistent API access to multiple frontier AI models. The use of diverse, international LLMs indicates an attempt to bypass regional restrictions and ensure high availability for the malware's decision-making process.

Implications

The emergence of CLOSEDQUORUM signals a critical shift in the threat landscape. As AI-powered malware becomes more autonomous, the time window for defenders to respond to an active breach is shrinking. Traditional signature-based detection is increasingly ineffective against malware that can dynamically rewrite its own logic or adapt its behavior based on LLM-generated advice.

Recommendations

Organizations must prioritize the implementation of AI-aware security monitoring. Security teams should: 1) Deploy behavioral analytics that can detect anomalous API traffic patterns associated with unauthorized LLM queries. 2) Utilize frameworks like CAIRN to audit internal systems for AI-integrated malicious artifacts. 3) Strengthen endpoint detection and response (EDR) policies to restrict unauthorized outbound connections to AI service providers from sensitive production environments.