ClosedQuorum Windows Malware Uses AI Models to Choose Post-Compromise Actions

· Original article ↗

Summary

Cisco Talos says the Go-based ClosedQuorum malware uses AI models to select post-compromise actions, including credential theft, shellcode injection and persistence. Its analyzed build cannot perform a listed lateral-movement action; in-the-wild use is unconfirmed.

Key points

  • ClosedQuorum uses Google Gemini, DeepSeek, Qwen and Mistral in a voting system to select actions based on reconnaissance; DeepSeek takes priority in a tie.
  • The Go-based Windows malware can steal credentials, inject shellcode using process hollowing or Early Bird APC injection, and establish persistence.
  • A lateral-movement option appears in the decision set, but the analyzed build has no handler to execute it.
  • Stolen data is sent to operators through a Discord webhook; Cisco Talos says the analyzed binary contained placeholder API credentials and a dummy webhook.
  • Cisco Talos says there is no confirmation of in-the-wild deployment. It linked binary artifacts to a developer's criminal-forum postings about carding dating to 2025.
  • Talos describes the malware as a possible shift toward automated attack chains, while noting reliance on commercial AI APIs could cause problems if they are rate-limited, unavailable or return malformed output.

Article Details

Event Type
Cisco Talos reported the discovery of ClosedQuorum, a Windows malware that uses AI models to select post-compromise actions.
Impact
ClosedQuorum can steal credentials, generate and inject shellcode, and execute a persistence module. It sends stolen data through a Discord webhook. Cisco Talos reported no confirmation of in-the-wild deployment; the analyzed binary contained placeholder API credentials and a dummy webhook.

MITRE ATT&CK

Malware

Products

Tools

Related Articles