Cisco Talos finds malware that puts its next move to a four-model vote

Cisco Systems Inc.’s Talos Threat Intelligence group has unveiled an open-source toolkit for detecting a new kind of malware dubbed CLOSEDQUORUM. This Windows credential stealer defies traditional command-and-control operations, instead directing its actions through a multi-model voting system involving Google Gemini, DeepSeek, Qwen, and Mistral—a departure from any previously documented Windows malware that relies on a single decision-making model.

The malware, a 16.4 MB Go binary, solicits advice from these four commercial large language models sequentially every five to 15 minutes. Each model is given a system prompt instructing it to play the role of an "advanced malware strategist" and provide only executable decisions. The model that secures the majority of votes – Mistral holds the tie-break – dictates the malware’s course of action, which could involve credential theft, code injection, persistence establishment, or lateral movement.

All these actions culminate in encrypted data exfiltration via Discord webhooks, bolstered by stealth tactics such as suppressing Windows event tracing.

The toolkit, known as the Cognitive Artifact Intelligence Research Network (CAIRN), is available on GitHub. It employs metadata analysis, searching for signs of AI integration like calls to model providers, Python framework imports, local model runtimes, and linguistic text designed to mislead automated analysts. The CAIRN filters categorize findings into three tiers, from the generic presence of AI components to the specific identification of malware families, utilizing an embedding model that groups samples sharing no common strings at all.

This new breed of malware reflects a significant evolution in attack tactics. While initially an optional feature, language models have since become central to the malware’s operation, evolving from simple payloads to a fully autonomous decision-making component. This shift underscores a trend of delegating more complex tasks to artificial intelligence, a development that poses a formidable challenge for defenders.

With models continuously responding to task requests indefinitely, the human element in the attack chain is diminished, allowing the malware to operate autonomously beyond human attention spans, working hours, and cognitive capacities.