Illustration for: Malware Now Lets Four AI Models Vote On The Attack

Malware Now Lets Four AI Models Vote On The Attack

Cisco Talos found a Windows malware sample that hands its command-and-control decisions to a panel of four AI models, which vote on the next move -- the first publicly documented malware with no human operator directing it.

By the Numbers

4 commercial LLMs
AI providers queried
Plurality vote
Decision method
16.4MB, Go binary
Sample size
None yet
Confirmed in-the-wild use
TC
By the Markets Desk
Edited by Trace Cohen · Early-stage VC & angel · Founder, New York Venture Partners
3 min read
ShareXLinkedInEmail

THE RUNDOWN

1

CLOSEDQUORUM is the first publicly documented Windows implant to delegate tactical command-and-control decisions entirely to commercial LLMs rather than a human operator or hardcoded logic.

2

The malware queries up to four providers -- DeepSeek, Qwen, Mistral and Google Gemini -- and resolves their votes by plurality, meaning it keeps functioning even if one or two providers refuse or rate-limit the request.

3

Talos has found no evidence of in-the-wild deployment yet, which means the security industry has a rare head start: a real design pattern to defend against before criminals scale it.

4

Talos's own response -- releasing CAIRN, an open-source toolkit that hunts for AI-malware fingerprints without executing the binary -- is a concrete new category of tooling every AI-security vendor now has to match or differentiate against.

TC

The VC Read · Trace's Take

Trace Cohen

The plurality-vote design is the tell that this was built by someone who understood AI safety mechanics well enough to route around them -- a single-provider dependency dies the moment one lab patches it, four independent providers voting doesn't. Diligence item for anyone backing an AI-security startup on the strength of a 'we detect AI-generated malware' pitch: ask specifically whether their detection depends on any one provider's refusal behavior, because CLOSEDQUORUM is a working example of why that assumption breaks.

Analysis

Cisco's Talos threat-intelligence unit disclosed a malware sample it calls CLOSEDQUORUM on September 22, describing it as the first publicly documented Windows implant to delegate command-and-control decisions to a panel of commercial AI models rather than to a human operator. The sample is a 16.4MB, 64-bit Windows executable compiled in Go.

How The Malware Actually Decides What To Do

CLOSEDQUORUM sends host context -- details about the infected machine -- to up to four commercial AI providers: DeepSeek, Qwen, Mistral and Google Gemini. Each model is prompted as an "advanced malware strategist" and asked to pick the next move from a constrained menu that includes credential theft, browser-stored password harvesting, cryptocurrency wallet theft, code injection and persistence mechanisms. Each of the four models returns a structured choice, and the implant resolves the four votes through simple plurality -- whichever option gets the most votes wins. That design means the malware keeps operating even if one provider blocks the request, rate-limits it, or refuses on safety grounds; it just needs a majority of the panel to cooperate.

“Each of the four models returns a structured choice, and the implant resolves the four votes through simple plurality -- whichever option gets the most votes wins.”

Talos says it has no confirmation the sample has been deployed against real targets -- this appears to be a proof-of-concept or early-stage tool discovered through Talos's own hunting rather than an active campaign causing documented harm. That is a meaningfully different, and better, starting point than most malware disclosures, which typically surface only after victims have already been hit.

Talos's Answer: A Toolkit, Not Just A Warning

Alongside the CLOSEDQUORUM disclosure, Talos open-sourced CAIRN (Cognitive Artifact Intelligence Research Network), a research toolkit built specifically to hunt, classify and track AI-integrated malware. CAIRN searches for the fingerprints AI-driven malware leaves behind -- prompt templates, provider API-key prefixes, jailbreak terms, orchestration logic -- without downloading or executing the actual malicious binary, a metadata-first approach that lets researchers scan at scale without the operational risk of running live malware samples.

This places Talos in a different competitive position than most cybersecurity vendors currently selling AI-security products: rather than pitching a product built for a hypothetical threat, Cisco is releasing free tooling built around a threat it just discovered itself, in the same disclosure. That's a stronger credibility signal than most vendor claims in this category, similar in spirit to Microsoft's EvilTokens takedown earlier this week, which paired a real enforcement action with named arrests rather than only a warning.

Why The Four-Provider Design Matters More Than It Looks

The plurality-vote architecture is the detail worth sitting with. A single-model dependency would make malware like this trivially killable -- one provider adds a refusal pattern for the specific prompt structure, and the malware goes dark. By spreading decision-making across four independent commercial providers with different safety training, different refusal thresholds and different content policies, CLOSEDQUORUM's authors built in redundancy against exactly the kind of provider-side mitigation that would otherwise be the fastest fix. It's the malware equivalent of a distributed system designed for fault tolerance, applied to evading AI safety guardrails specifically.

For AI labs, this creates an uncomfortable shared-responsibility problem: no single provider's safety team can unilaterally solve this, because the malware doesn't depend on any one of them individually. DeepSeek, Qwen, Mistral and Google would each need to independently detect and block this exact use pattern, or coordinate on shared threat intelligence about it -- something the industry has talked about wanting for years but rarely does at the speed adversaries iterate.

What to watch: whether any of the four named providers confirm they've since blocked the specific prompt pattern CLOSEDQUORUM uses, and whether Talos or another researcher finds this design pattern reused in an actual in-the-wild campaign in the coming months -- the gap between a documented proof-of-concept and a live incident is usually measured in weeks, not years, once the technique is public.

ShareXLinkedInEmail

Key Sources

2 sources

THE WIRE in your inbox— Tech, startup & VC news with Trace's take. Free, no spam.