Particle.news
Get it on Google Play
Download on the App Store
7 ARTICLES
·
yesterday

ClosedQuorum Malware Lets Four Commercial AI Models Vote on Its Next Move

Talos says the design hands tactical control to third‑party AI models, creating a new class of autonomous attacks that defenders can spot by tracking AI API use and related behaviors.

An image about Researchers found malware that lets AI models vote on how to attack you
Image
Image
Image

Overview

  • Cisco Talos publicly disclosed ClosedQuorum on Tuesday, Sept. 22 and released CAIRN plus YARA and Snort rules, SHA‑256 hashes, and other indicators to help hunt AI‑integrated malware.
  • The Go‑based implant sends host details to up to four commercial models — DeepSeek, Qwen, Mistral, and Google Gemini — tallies their structured responses, and executes the action with the most votes.
  • Models may choose one of four actions — steal, inject, persist, or move — where steal triggers LSASS memory dumps, saved‑browser credential theft, and crypto‑wallet extraction, and results are sent to an operator via a Discord webhook.
  • Talos found the analyzed public build is inert because it uses placeholder API keys and a dummy webhook but logged evidence that developer builds can be compiled with real credentials and linked the developer to 2025 carding forum posts.
  • Researchers say the malware illustrates “effort displacement” by removing the human decision loop while also exposing defensive weak points such as reliance on commercial AI APIs, predictable tie‑breaking, TLS‑inspectable prompts, and correlated behaviors defenders can monitor.