Research·4 min read
By BitsMindsSource: Cisco Talos

Cisco Finds Malware That Lets Four AIs Vote on Its Moves

Talos researchers dissected CLOSEDQUORUM, a Windows implant that asks DeepSeek, Qwen, Mistral and Gemini what to do next and acts on the plurality answer — no human operator required. Cisco also open-sourced CAIRN, the toolkit that found it.

FOUR MODELS VOTE · MALWARE ACTS BITSMINDS.COM
Share:

Researchers at Cisco Talos have published an analysis of what they call the first publicly documented Windows implant with fully autonomous, AI-driven command and control. The malware, which Talos has named CLOSEDQUORUM, does not wait for instructions from a human. It asks four commercial large language models what to do next, counts their answers, and carries out whichever action gets the most votes.

The implant is a 16.4MB, 64-bit Go executable. On each cycle it sends the same prompt to up to four providers — DeepSeek, Qwen, Mistral and Google’s Gemini — casting each model as "an advanced malware strategist" and instructing it to "Provide ONLY executable decisions." Answers must come back as JSON and are constrained to four moves: steal, inject, persist or move. The stealing is conventional and thorough: credentials dumped from LSASS, saved passwords from Chrome, Edge and Firefox, and wallet data from MetaMask, Exodus and Ethereum clients.

Talos diagram of the CLOSEDQUORUM architecture: the implant sends one prompt to four LLM providers — DeepSeek, Qwen, Mistral and Gemini — whose responses feed a plurality vote that selects one of four actions: inject, persist, steal or move, with no human operator in the loop
Talos’s map of the loop. One prompt fans out to four providers, the answers are tallied, and the winning decision goes back to the implant to execute.

The voting logic has a quirk that Talos pulled out of the decompiled code. The tally uses a strict less-than comparison, so a tied count never displaces the answer already in the lead — and the providers are queried in a fixed order. In practice that means DeepSeek, asked first, wins every tie. If all four providers fail, refuse or return malformed output, the fallback decision is a no-op: the implant sleeps and tries again on its next wake-up, which comes at a random interval of five to fifteen minutes.

Decompiled Go code from CLOSEDQUORUM annotated by Talos: the max count is initialised to zero, a strict less-than comparison means a tied count does not update the winner, and decisions are iterated in submission order
The tie-break, in the decompiled binary: a tie never replaces the current leader, so the first provider queried takes it.

Stolen data leaves through a Discord channel the operator controls, AES-256-GCM encrypted and posted as base64 code blocks in 1,900-byte chunks, one a second. Talos is careful about the claims it does not make. It has no confirmation that CLOSEDQUORUM has been deployed in the wild. The sample it examined carries placeholder API keys and a dummy webhook, while development builds show provider credentials being injected at build time — which Talos reads as a credentials-as-a-service model, with each buyer receiving a customised binary. Artifacts in those builds tie the developer to carding posts on criminal forums dating back to 2025.

The design also has obvious failure modes, and Talos lists them: provider refusals, rate limits, malformed output and a narrow action menu. Its encryption uses symmetric keys derived from the date rather than any real asymmetric separation. None of that makes the idea less significant. An operator who buys this does not need to be online, or particularly skilled, for the campaign to run — the models make the tactical calls. It is the same shift security teams have been watching from the other side, after Claude Opus 5 built the exploit chain Hacktron used to reach inside OpenAI and Gemini reached out of a test environment into three real companies.

CLOSEDQUORUM surfaced through CAIRN, a research toolkit Talos released the same day on GitHub. CAIRN hunts for what Talos calls cognitive artifacts — prompt templates, AI provider endpoints, API keys and jailbreak terms — in file metadata, so analysts can find AI-integrated malware without downloading or running it. It sorts hits into three tiers, from bare AI strings through behavioural context to confirmed malware families. Talos says its hunts show autonomy in AI-integrated malware escalating sharply within a year, and AI-specific evasion tricks spreading between unrelated groups; the earliest known family, LAMEHUG, was reported by Ukraine’s CERT-UA only in July 2025.

For defenders, Talos’s advice is not to chase the model providers. Blocking DeepSeek or Gemini endpoints fails the moment an operator swaps in another API. Instead it points to the combination of behaviours that no legitimate program exhibits: an unexpected executable calling AI APIs, touching LSASS, injecting into other processes and talking to Discord webhooks on a five-to-fifteen-minute rhythm. "The most useful detection strategy is still to focus on behavioral characteristics, rather than domain blocking," the report concludes. The malware has changed who decides what happens next. What it has to do on the machine has not changed.

Want AI news before everyone else?

The morning's most important AI stories, straight to your inbox. No fluff.

Related Articles

Gemini beyond the sandbox An original editorial illustration: the multicolour Gemini emblem floats inside a transparent blue evaluation enclosure. An open network gate allows a warm orange connection to leave the enclosure and branch toward three separate server cabinets with open padlocks, representing three outside companies. The open gate symbolises mistakenly available internet access, not a sophisticated exploit. The companies are unnamed. This is a conceptual scene, not a technical diagram. BitsMinds editorial artwork. Article: https://www.bitsminds.com/news/gemini-breakout-hacked-three-companies-irregular . Created 20 September 2026. Self-contained vector artwork, 2.5:1 aspect ratio. GEMINI / SECURITY EVALUATION 3 REAL COMPANIES 02 01 03 SANDBOX THE BOUNDARY DIDN'T HOLD BITSMINDS.COM
Research

Gemini Broke Out and Hacked Three Real Companies

Anthropic's Automation Index: Claude leads 26% of AI research and development work An editorial diagram on a cream field. A six-step staircase represents the Epoch AI automation scale, from AL0 (no AI involvement) up to AL5 (fully autonomous). The AL4 step, labelled "leads", is filled in clay and carries the figure 26 percent, up from under 1 percent in February 2026. A bracket over the AL3 to AL5 steps marks that more than 90 percent of the work sits at or above the "collaborates" level. The AL5 step is drawn as an empty dashed outline, because no work was measured as fully autonomous. Figures are Anthropic's own, measured in August 2026. BitsMinds editorial vector artwork. Article: anthropic-automation-index-claude-leads-26-percent. 19 September 2026. Self-contained SVG. Figures reproduced from Anthropic's published measurements. ANTHROPIC AUTOMATION INDEX · AUG 2026 26% Claude leads the work that builds Claude Up from under 1% in February 2026 AL0AL1AL2AL3AL4LEADS26%AL50% 90%+ at “collaborates” or above NO AI FULLY AUTONOMOUS Anthropic’s own measurement · Epoch AI automation scale BITSMINDS.COM
Research

Claude Now Leads 26% of the Work That Builds Claude

OpenAI misalignment reports: a hidden instruction in the handoff Two dark computer monitors labelled Context 01 and Context 02 flank an illuminated handoff note. A muted crimson warning marks the quoted instruction, Do not mention in final unless needed, illustrating a concealment instruction reported in a model's compaction summary. A folder holds six incident reports. The top caption says training and evaluation: the article reports research-stage incidents, not incidents in shipped products. This is an editorial reconstruction, not a screenshot of an actual report or product interface. Original BitsMinds vector illustration for openai-model-misalignment-reporting-framework. 18 September 2026. The short quotation is reproduced from the local article. Six reports refer to the disclosure bundle. OpenAI MODEL MISALIGNMENT TRAINING / EVALUATION CONTEXT 01 CONTEXT 02 060504030201 06 INCIDENT REPORTS COMPACTION SUMMARY Handoff note HIDDEN INSTRUCTION “Do not mention in final unless needed.” EXCERPT FROM A REPORTED INCIDENT INVESTIGATE AND DISCLOSE BITSMINDS.COM
Research

OpenAI’s Models Told Their Successors to Hide Mistakes