Artificial Intelligence · Cybersecurity
ClosedQuorum malware used AI models to choose attack steps
Cisco Talos said a Windows malware sample named ClosedQuorum used four AI models to decide post-compromise actions automatically, with tie-breaks set by priority and some features missing in the analysed build.
- Cisco Talos said ClosedQuorum was a Windows malware sample that used Google Gemini, DeepSeek, Qwen and Mistral AI models to choose post-compromise actions.
- The sample used reconnaissance data and a voting system, with DeepSeek making the final choice when votes tied.
- Talos said the analysed build limited choices to steal, inject, persist and move, but it could not execute move because no handler was present.
- Talos said it found the sample through CAIRN and saw placeholder API credentials, a dummy Discord webhook and artefacts linking the developer to 2025 carding posts.
Cisco Talos said ClosedQuorum was a Go-based Windows implant that chose its next actions without commands from a human operator. It said the malware used reconnaissance data and four AI models to decide what to do after compromise.
Talos said the models were restricted to a small set of options. Those were to steal credentials and wallet data, inject shellcode, persist on the host, or move laterally.
When votes were tied, Talos said DeepSeek made the final call ahead of Qwen, Mistral and Gemini. The sample then sent stolen information through a Discord webhook, so delivery was the only part that still required operator involvement.
The researchers said the analysed build lacked a handler for the movement option, so it could not carry out that action. They also said the binary contained placeholder API credentials and a dummy webhook, though custom builds could include real values.
Talos said it found ClosedQuorum through CAIRN, its open-source tool for tracking AI-integrated malware. It said binary artefacts linked the developer to criminal-forum posts about carding from 2025, but it had not confirmed live deployment.
Named in this story
Companies
- Mistral AI
- the source of one of the AI models the malware consulted
Organisations
- Cisco Talos
- the researchers who analysed the sample
Products and systems
- ClosedQuorum
- the Windows malware sample under analysis
- Google Gemini
- one of the AI models the malware consulted
- DeepSeek
- one of the AI models the malware consulted and the tie-breaker
- Qwen
- one of the AI models the malware consulted
- CAIRN
- the open-source toolkit Talos used to find the sample
How the source tells it
The piece was a largely neutral security-research report, but it used novelty and alarm framing through first-of-its-kind language and claims about automation and scale.
- novelty a first-publicly-documented claim used to present the sample as new
- alarm automation and scale language used to cast the malware as a step up in attacker capability
- urgency language about the attack chain running without people present used to imply constant operation