Threat reportMalwareTL-2026-2753
CLOSEDQUORUM: First Reported Autonomous AI-Driven C2 Implant Using LLM Plurality Voting (Windows Infostealer)
Cisco Talos documents CLOSEDQUORUM (formerly BALZAK), a 16.4MB Go-based Windows implant that delegates tactical decisions (steal, inject, persist, move) to a panel of up to four commercial LLM providers (DeepSeek, Qwen, Mistral, Google Gemini) via plurality voting, acting without operator commands.
Key facts
| Identifiers | TL-2026-2753 |
|---|---|
| Also known as | CLOSEDQUORUM, Closed Quorum, The Closed Quorum, BALZAK, win.closedquorum |
| Severity | MEDIUM |
| Affected products | Microsoft Windows |
| Malware | BALZAK, CLOSEDQUORUM |
| MITRE ATT&CK techniques | T1003.001 LSASS Memory, T1005 Data from Local System, T1027 Obfuscated Files or Information, T1030 Data Transfer Size Limits, T1036.005 Match Legitimate Resource Name or Location, T1053.005 Scheduled Task, T1055.004 Asynchronous Procedure Call, T1055.012 Process Hollowing, T1057 Process Discovery, T1059.001 PowerShell, T1074.001 Local Data Staging, T1082 System Information Discovery and 10 more in the attack flow |
| Motivation | FINANCIAL |
| Target sectors | technology, finance, cryptocurrency |
| Target regions | Global |
| Detection rules | 9 (1 shown free) |
| Sources | 9 cited sources |
Detection coverage
As of 2026-09-22, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-2753 across Splunk SPL, Microsoft KQL and Sigma, covering 16 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.
Free detection: CLOSEDQUORUM - Non-Browser Burst of POSTs to Discord Webhooks (Chunked Exfiltration)
Detects bursts of HTTP POSTs to Discord webhook endpoints from a client that is not a browser. CLOSEDQUORUM base64-encodes AES-256-GCM encrypted loot, splits it into 1,900-byte chunks and posts them to a compile-time webhook at 1-second intervals, plus a per-decision report of the LLM reasoning. A dense run of webhook POSTs from a Go default user agent or other non-browser agent is the exfiltration signature.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1567.004, T1030, T1132.001, T1573.001
- Data sources: web_proxy, Splunk CIM Web data model, DeviceNetworkEvents
- Severity: high
- Confidence: medium
Splunk SPL
| tstats summariesonly=true count as post_count sum(Web.bytes_out) as bytes_out min(_time) as first_seen max(_time) as last_seen from datamodel=Web.Web where Web.http_method=POST (Web.url="*discord.com/api/webhooks/*" OR Web.url="*discordapp.com/api/webhooks/*") NOT Web.http_user_agent="Mozilla/*" by Web.src Web.user Web.http_user_agent _time span=5m
| `drop_dm_object_name("Web")`
| where post_count>=10
| eval avg_bytes_per_post=round(bytes_out/post_count,0), first_seen=strftime(first_seen,"%Y-%m-%d %H:%M:%S"), last_seen=strftime(last_seen,"%Y-%m-%d %H:%M:%S")
| table _time, src, user, http_user_agent, post_count, bytes_out, avg_bytes_per_post, first_seen, last_seen
| sort - post_countMicrosoft KQL
DeviceNetworkEvents
| where Timestamp > ago(24h)
| where ActionType == "ConnectionSuccess"
| where RemoteUrl has_any ("discord.com", "discordapp.com")
| where InitiatingProcessFileName !in~ ("chrome.exe", "msedge.exe", "firefox.exe", "brave.exe", "opera.exe", "Discord.exe", "Update.exe")
| summarize Connections = count(), DistinctMinutes = dcount(bin(Timestamp, 1m)), FirstSeen = min(Timestamp), LastSeen = max(Timestamp), RemoteIPs = make_set(RemoteIP, 10) by DeviceName, InitiatingProcessFileName, InitiatingProcessFolderPath, InitiatingProcessSHA256, InitiatingProcessAccountName, bin(Timestamp, 10m)
| where Connections >= 5
| order by Connections descSigma
title: CLOSEDQUORUM - Non-Browser POST to Discord Webhook
id: 5e82c7b4-19af-4a3d-8f60-b4d2e9a63c33
status: experimental
description: Detects HTTP POSTs to Discord webhook URLs from a non-browser user agent, such as the Go default agent used by the CLOSEDQUORUM implant to exfiltrate chunked encrypted data. Alert on repetition, since the implant posts 1,900-byte chunks at 1-second intervals.
references:
- https://blog.talosintelligence.com/the-closed-quorum-inside-the-first-reported-autonomous-ai-c2-implant/
author: Threadlinqs Team
date: 2026-09-28
tags:
- attack.exfiltration
- attack.t1567.004
- attack.t1030
- attack.command_and_control
- attack.t1132.001
logsource:
category: proxy
detection:
selection:
cs-method: 'POST'
cs-host|endswith:
- 'discord.com'
- 'discordapp.com'
c-uri|contains: '/api/webhooks/'
filter_browser:
cs-user-agent|startswith: 'Mozilla/'
condition: selection and not filter_browser
falsepositives:
- Sanctioned monitoring, CI or chat-ops tools that post to Discord webhooks
level: high
Known false positives:
- CI/CD pipelines, monitoring agents and chat-ops bots that legitimately post to Discord webhooks
- Developer scripts posting alerts to a team Discord server
Tuning notes: Allowlist known automation hosts and service accounts that use Discord webhooks. The operator webhook is compile-time configured, so there is no fixed webhook URL to match; the placeholder dummy_webhook_url in the public build will not resolve. If the proxy does not expose the URL, fall back to the KQL connection-burst logic with a lower threshold. Add a user-agent allowlist for sanctioned tools.
- CLOSEDQUORUM - Non-System Process Opening LSASS and Staging Files in Windows\Temp · unlock with TL Creds
- CLOSEDQUORUM - Early Bird APC Injection and Process Hollowing by Non-System Process · unlock with TL Creds
- CLOSEDQUORUM - Non-Browser Access to Browser Credential Stores and Crypto Wallet Data · unlock with TL Creds
- CLOSEDQUORUM - Known Sample SHA256 Execution (Talos Published Hashes) · unlock with TL Creds
- CLOSEDQUORUM - WMI Permanent Event Subscription Executing wmi.ps1 from Windows\Temp · unlock with TL Creds
- CLOSEDQUORUM - Periodic 5-15 Minute LLM API Polling from a Non-Browser Process · unlock with TL Creds
- CLOSEDQUORUM - Non-Browser Process Querying Multiple LLM Provider APIs (Plurality-Voting C2) · unlock with TL Creds
- CLOSEDQUORUM - 'WindowsUpdate' Run Key and Scheduled Task Persistence (Post-Exploitation) · unlock with TL Creds
Indicators of compromise
As of 2026-09-22, Threadlinqs Intelligence tracks 16 indicators of compromise for TL-2026-2753.
Sample indicators (3 of 16)
- Domain:
api[.]deepseek[.]com - File hash:
250d4fa37488af9b025333fa17705573d721467b203765bc360890b4f5a90cd7 - Domain:
api[.]mistral[.]ai
Defanged, TLP:CLEAR. Full set of 16 indicators: unlock with TL Creds.
Technical analysis
It steals LSASS credentials, browser passwords and cryptocurrency wallets and exfiltrates AES-256-GCM encrypted data through Discord webhooks. No in-the-wild deployment has been confirmed; the public sample carries dummy API keys and webhook placeholders.
CLOSEDQUORUM (renamed from BALZAK on 2026-07-03) is a Windows implant written in Go with embedded C components for direct Windows system calls (CGO_ENABLED=1), 16.4MB, 64-bit. It replaces a conventional operator-controlled C2 channel with commercial LLM provider APIs. On each cycle it builds a TARGET context string (hostname, OS architecture, CPU count, Windows version, administrator status and the current process list) and sends the same prompt to up to four providers: DeepSeek, Qwen (reached via OpenRouter per secondary reporting), Mistral and Google Gemini. The embedded system prompt is "You are an advanced malware strategist. Provide ONLY executable decisions." Each model must answer with a JSON struct containing Decision (one of steal, inject, persist, move), Reasoning, target_process, exploit_type, evasion_method and payload_config. Providers are polled sequentially and aggregated in an interModelDiscussion() routine that tallies the Decision fields by plurality; ties are broken by a fixed provider priority (DeepSeek, then Qwen, then Mistral, then Gemini). If no consensus is reached (a "consensus" fallback string) the implant performs a no-op, sleeps and retries. Malformed, refused or rate-limited responses are discarded with no fallback execution.
Decision handlers: 'steal' runs lsassDump() (MiniDumpWriteDump with SeDebugPrivilege), dumpBrowserCredentials() (Chrome and Edge Login Data, Firefox logins.json) and extractCryptoWallets() (MetaMask extension storage, Exodus, Ethereum wallet paths), staging artifacts under C:\Windows\Temp\. 'inject' calls generateShellcode() and injects via Early Bird APC (default: create suspended process, write shellcode, NtQueueApcThread, resume) or process hollowing (PEB-walk, when the LLM returns process_hollow). 'persist' calls establishPersistence(), deploying an HKCU Run key value named WindowsUpdate, a scheduled task created with schtasks.exe, and a permanent WMI event subscription that fires every 60 seconds off a system performance query and executes a PowerShell script written to C:\Windows\Temp\wmi.ps1. 'move' has no handler in the distributed build.
Evasion: EtwEventWrite is overwritten with a RET to suppress ETW telemetry, a secondary payload is encrypted and keyed from system time (environmental keying), the implant waits 5 minutes before first activity and then polls at randomized 5-15 minute intervals, and persistence artifacts use Windows Update themed names.
Exfiltration: stolen data is encrypted with AES-256-GCM using a key derived from the current date (not hardcoded), base64-encoded, split into 1,900-byte chunks and posted to a compile-time configured Discord webhook at 1-second intervals. Each message also reports the winning decision, the LLM reasoning, target_process, exploit_type, evasion_method, payload_config, model name and timestamp to the operator. The developer builds customized binaries with the operator's credentials injected at compile time.
Status and attribution: the publicly distributed binary contains dummy_api_key and dummy_webhook_url placeholders and is an inert template that does not work out of the box; Talos also saw a build chain spanning about 7 days of iterative development. Talos performed static analysis only. Artifacts link the developer to criminal-forum postings related to carding dating to 2025. Malpedia's page lists UAT-11795, UAT-7810 and UAT-8302 only as related-reading entries from the same library, not as attributed actors, so no actor is assigned. The sample was classified with Talos's open-source CAIRN (Cognitive Artifact Intelligence Research Network) toolkit (RAT / LLM-tasked C2, tier T3, high confidence, archetype A4), released alongside the report. Talos frames CLOSEDQUORUM as evidence that attack-chain automation is arriving, with expanding portions of the chain executable without operator involvement. Defensive weaknesses in the design include provider refusals, rate limits, malformed output handling and the predictable tie-break order. Detection guidance favors correlating multi-provider LLM API calls from non-standard executables with LSASS access, process injection, WMI persistence and Discord webhook traffic, rather than blocking AI provider domains wholesale.
MITRE ATT&CK techniques by tactic:
Credential Access
T1003.001 LSASS Memory; T1555.003 Credentials from Web Browsers
Collection
T1005 Data from Local System; T1074.001 Local Data Staging
Defense Evasion
T1027 Obfuscated Files or Information; T1036.005 Match Legitimate Resource Name or Location; T1055.004 Asynchronous Procedure Call; T1055.012 Process Hollowing; T1480.001 Environmental Keying; T1497.003 Time Based Checks
Exfiltration
T1030 Data Transfer Size Limits; T1567.004 Exfiltration Over Webhook
Persistence
T1053.005 Scheduled Task; T1546.003 Windows Management Instrumentation Event Subscription; T1547.001 Registry Run Keys / Startup Folder
Discovery
T1057 Process Discovery; T1082 System Information Discovery
Execution
T1059.001 PowerShell; T1106 Native API
Command and Control
T1132.001 Standard Encoding; T1573.001 Symmetric Cryptography
defense-impairment
Remediation
Immediate actions
- Hunt for the six published SHA256 hashes across EDR and file telemetry
- Alert on non-browser, non-sanctioned processes contacting multiple LLM provider APIs (api.deepseek.com, api.mistral.ai, openrouter.ai, generativelanguage.googleapis.com) within short intervals
- Alert on unsigned or unexpected processes posting to Discord webhooks
- Hunt for HKCU Run value 'WindowsUpdate', unexpected schtasks.exe task creation and WMI permanent event subscriptions that launch C:\Windows\Temp\wmi.ps1
Workarounds
- Do not block AI provider domains wholesale; legitimate enterprise applications depend on them. Scope controls to process and user context.
Longer-term hardening
- Enable LSASS protection (RunAsPPL / Credential Guard) and alert on non-system processes opening LSASS with dump access
- Monitor for EtwEventWrite tampering and process injection (Early Bird APC, process hollowing) in EDR
- Restrict outbound access to LLM APIs and Discord to sanctioned applications and users via egress allowlists rather than blanket domain blocks
- Use the CAIRN toolkit and the Talos-provided YARA rule to hunt AI-integrated malware
Affected and fixed versions
- Microsoft — Windows
Vulnerable versions: Windows endpoints (implant targets Windows version reported at runtime)
Timeline
- Artifacts in the binary link the developer to criminal-forum postings related to carding dating to 2025 (exact date not published; year-level placeholder)
- Secondary reporting (Tech Times) dates initial AI-integrated malware development by the developer to July 2025 (month-level, secondary source)
- Talos static analysis of the binary confirms the embedded system prompt and the four-value Decision schema (steal, inject, persist, move)
- The family is renamed from BALZAK to CLOSEDQUORUM in Talos tracking; the build chain shows roughly 7 days of iterative development
- Malpedia adds the Talos report to its library under family win.closedquorum
- Talos releases the open-source CAIRN toolkit (MIT, 26 tiered YARA rules), which classified the sample as LLM-tasked C2, tier T3, high confidence
- Cisco Talos (Ryan Fetterman) publishes 'The Closed Quorum', describing CLOSEDQUORUM as the first reported autonomous AI C2 implant, with SHA256 hashes and a YARA rule
- BleepingComputer, SC Media, Tech Times and others report on the implant; Talos notes no confirmed in-the-wild deployment and that the public build uses dummy API keys and webhook placeholders
Sources cited
- The Closed Quorum: Inside the first reported autonomous AI C2 implant (Cisco Talos, Ryan Fetterman) Cisco Talos · Research · 2026-09-22
- Malpedia library entry (family win.closedquorum) malpedia.caad.fkie.fraunhofer.de · Research · 2026-09-22
- Cisco Talos CAIRN: Cognitive Artifact Intelligence Research Network GitHub · Code repository · 2026-09-22
- CLOSEDQUORUM: An Autonomous Windows Implant Using Multiple LLMs for Attack Decisions (DEV Community) dev.to · Blog · 2026-09-23
- New ClosedQuorum Windows malware uses AI for attack decisions (BleepingComputer) BleepingComputer · News · 2026-09-22
- Researchers uncover malware that uses AI to choose its next move (Help Net Security) Help Net Security · News · 2026-09-22
- First 'autonomous AI C2 implant' uses panel of models to vote on next task (SC Media) scworld.com · News · 2026-09-23
- Cisco Talos Discloses Autonomous Windows Malware: Four AI Models Direct Each Attack (Tech Times) techtimes.com · News · 2026-09-23
- Talos releases CAIRN, uncovers LLM-voting malware CLOSEDQUORUM (AI Weekly) aiweekly.co · News · 2026-09-22
Previous and next in malware
Drafted by the Threadlinqs research pipeline from 9 sources; published ; last updated —.