Back to Intelligence

CyberStrikeAI LLM-Orchestrated C2 + VOLTA AI-Powered Infostealer: OTX Pulse Analysis — Credential Theft & LLMjacking Detection Pack

SA
Security Arsenal Team
November 2, 2026
12 min read

Intelligence Category: Infostealer & Credential Theft Campaigns Pulse Count Analyzed: 5 TLP: WHITE Threat Level: HIGH


Threat Summary

Five concurrent OTX pulses reveal a singular, alarming trend: threat actors have operationalized large language models across every phase of the intrusion lifecycle — from attack orchestration and victim prioritization to social engineering lure creation and monetization of stolen AI credentials themselves.

The collective picture:

  1. CyberStrikeAI — an open-source, LLM-orchestrated attack framework — was observed in live intrusions against educational institutions across Africa and Latin America. An AI model resident on the C2 server plans attack actions and generates commands in real time, while implants (NodeRabbit, PollCat) execute them on compromised hosts. Initial access was gained through compromised SQL servers, with Mimikatz deployed for credential dumping.

  2. VOLTA infostealer represents the commoditization of AI in cybercrime: after harvesting credentials, browser data, crypto wallets, and session tokens, VOLTA compiles a structured "evidence bundle" and submits it to OpenRouter using models like Claude Opus to score and prioritize victims by monetization potential. A companion module, VOLTA-AUTODOXER, automates doxxing of high-value victims.

  3. A fake ChatGPT Plus billing phishing campaign harvests credit card details and OTP codes via pages hosted on compromised WordPress instances under Nexcess's nxcli.io infrastructure — with live attacker involvement, where operators monitor submissions in real time and dynamically interact with victims to defeat OTP challenges.

  4. Microsoft Teams help desk impersonation campaigns use Quick Assist remote sessions to gain footholds, then escalate to full Active Directory compromise via DCSync, PetitPotam coercion, and AD CS abuse (ESC1/ESC8), exfiltrating data with Rclone.

  5. "Black October" LLMjacking wave — at least nine AI Developer Platforms (Fireworks AI, Zeabur, Lightning AI, Railway, Warp, and others) breached for stored AI API credentials, which were then brokered or abused for model distillation and unauthorized LLM access.

Objective: Credential theft at scale, monetization through fraud, resale of AI platform access, and data extortion — with AI used both as a weapon and as a target.


Threat Actor / Malware Profile

CyberStrikeAI (with NodeRabbit / PollCat implants)

  • Distribution: Post-exploitation deployment following SQL server compromise; targets education sector infrastructure.
  • Payload Behavior: Implant-based architecture where the heavy lifting (decision-making, command generation) occurs server-side via LLM orchestration. Implants execute AI-generated commands, making host-side signatures minimal and behavior adaptive per-victim.
  • C2 Communication: Traditional C2 channel to attacker infrastructure at 104.234.26.163 and media.fastcdn.bar, with the LLM layer planning actions upstream. Traffic may blend with legitimate CDN-style patterns given the fastcdn naming convention.
  • Persistence: Credential dumping via Mimikatz enables lateral movement and account-based persistence rather than relying solely on implant survivability.
  • Anti-Analysis: LLM-driven command generation means attack patterns are non-deterministic — no two intrusions look identical, defeating static playbooks and naive behavioral baselines.

VOLTA / VOLTA-AUTODOXER

  • Distribution: Windows-targeted infostealer (typical stealer-as-a-service channels: malvertising, cracked software, phishing loaders).
  • Payload Behavior: Harvests browser credentials/cookies, cryptocurrency wallets, platform session tokens, and PII. Uniquely, it assembles a structured evidence bundle and queries OpenRouter (Claude Opus with fallbacks) to appraise victim value — effectively automating triage that human operators used to perform.
  • C2 Communication: Exfiltration of stolen data plus outbound API calls to OpenRouter infrastructure — meaning egress traffic to LLM API endpoints from non-developer endpoints is a high-fidelity detection signal.
  • Anti-Analysis: AI-based victim scoring lets operators ignore low-value infections quickly, reducing their exposure window during analysis.

Teams Help Desk Impersonation Intrusion Set

  • Distribution: Pure social engineering — attacker-initiated Teams chats impersonating IT, followed by Quick Assist sessions using victim-shared connection codes.
  • Payload Behavior: Deployment of portable Python, Bun, and Java runtimes (living-off-the-land interpreter abuse), AD enumeration, PetitPotam NTLM coercion, AD CS exploitation (ESC1/ESC8), DCSync for full domain credential theft, and Rclone for cloud exfiltration.

LLMjacking Intrusion Set ("Black October")

  • Distribution: Exploitation of vulnerabilities in AI developer platforms to access stored environment variables containing AI API keys.
  • Objective/MO: Theft of LLM credentials for resale ("LLM brokers"), unauthorized model access, and model distillation attacks. Known infrastructure includes 31.57.201.201, 38.246.114.118, artex.nhhh.de, and direct.465415.xyz.

IOC Analysis

The indicator set spans four types, each requiring distinct operational handling:

IOC TypeExamplesOperationalization
IPv4 (C2/Infra)104.234.26.163, 31.57.201.201, 38.246.114.118Block at egress firewall/proxy; add to EDR network watchlists; retro-hunt NetFlow for 90 days
Domains/Hostnamesfastcdn.bar, authgpt.info, ai-billing.info, url-y.uk, *.nxcli.io (campaign hosts), artex.nhhh.de, direct.465415.xyzDNS sinkhole/block; alert on resolution; note nxcli.io subdomains indicate compromised WordPress hosting — full-domain blocking may cause collateral damage, so block the specific hostnames and alert on new *.nxcli.io resolutions
File Hashes (MD5/SHA1/SHA256)61c0810a2358...d8a1 (CyberStrikeAI), 57ef4242666d...2279 (VOLTA), Teams-campaign hash setPush to EDR block lists; hash reputation lookups via VirusTotal/OTX; SHA256 values are the highest-fidelity — prioritize them over MD5/SHA1 due to collision concerns
Behavioral (no hash)Quick Assist usage, portable runtime execution, OpenRouter API egressSignature-less detection required — see Detection Engineering

Tooling guidance: Enrich hashes with OTX, VirusTotal, and MalwareBazaar pivots. Decode and validate domains with passive DNS (e.g., SecurityTrails, CIRCL pDNS) to catch infrastructure rotation — fastcdn.bar-style names suggest the actor rotates CDN-themed domains. Import the full indicator set into your TIP with appropriate confidence aging (IPs decay fast; file hashes persist).


Detection Engineering

YAML
---
title: CyberStrikeAI / NodeRabbit C2 Communication to Known Infrastructure
id: 8f3a2c11-9e4b-4d7a-a1c2-cyberstrikeai001
status: experimental
description: Detects network connections to CyberStrikeAI LLM-orchestrated C2 infrastructure identified in OTX pulses targeting education sector in Africa and Latin America
author: Security Arsenal Threat Intelligence
date: 2026/11/03
references:
    - https://securelist.com/tr/cyberstrikeai-attacks/121715/
logsource:
    category: network_connection
product: windows
detection:
    selection_ip:
        DestinationIp:
            - '104.234.26.163'
    selection_domain:
        DestinationHostname|contains:
            - 'fastcdn.bar'
    condition: selection_ip or selection_domain
falsepositives:
    - Unlikely; fastcdn.bar is not associated with legitimate CDN providers
level: high
tags:
    - attack.command_and_control
    - attack.t1071
---
title: VOLTA Infostealer LLM-Based Victim Profiling Egress
id: 7b2e9d44-3f1a-4c8e-b2d3-voltaai0002
status: experimental
description: Detects outbound connections to OpenRouter or LLM API endpoints from non-developer workstation processes, consistent with VOLTA infostealer submitting victim evidence bundles for AI-based monetization scoring
author: Security Arsenal Threat Intelligence
date: 2026/11/03
references:
    - https://blog.talosintelligence.com/appraisal-by-ai-model-driven-victim-prioritization-in-the-volta-infostealer/
logsource:
    category: network_connection
product: windows
detection:
    selection_llm_egress:
        DestinationHostname|contains:
            - 'openrouter.ai'
            - 'api.anthropic.com'
            - 'api.openai.com'
    filter_legitimate_dev_tools:
        Image|endswith:
            - '\code.exe'
            - '\cursor.exe'
            - '\python.exe'
            - '\node.exe'
    filter_known_users:
        User|contains: 'dev_'
    condition: selection_llm_egress and not 1 of filter_*
falsepositives:
    - Developers testing LLM integrations from workstations; tune filters per environment
    - Legitimate AI-enabled applications; maintain an allowlist of approved AI tooling
level: medium
tags:
    - attack.exfiltration
    - attack.t1041
    - attack.credential_access
    - attack.t1555
---
title: Quick Assist Execution Followed by AD Credential Theft Tooling
id: 4c1f8a77-2d9b-4e3a-c5f6-teamsad00003
status: experimental
description: Detects Quick Assist remote support execution followed by Rclone, portable runtime, or DCSync-related activity consistent with Microsoft Teams help desk impersonation intrusions
author: Security Arsenal Threat Intelligence
date: 2026/11/03
references:
    - https://cyberpress.org/hackers-impersonate-it-help-desk-microsoft-teams/
logsource:
    category: process_creation
product: windows
detection:
    selection_quickassist:
        Image|endswith: '\quickassist.exe'
    selection_followup_tools:
        Image|endswith:
            - '\rclone.exe'
            - '\bun.exe'
        CommandLine|contains:
            - 'lsadump::dcsync'
            - 'PetitPotam'
            - 'syncall'
    timeframe: 30m
    condition: selection_quickassist | near selection_followup_tools
falsepositives:
    - Legitimate help desk use of Quick Assist; correlate with Teams chat origin and verify with IT ticketing system
    - Approved Rclone backup jobs; allowlist known service accounts and paths
level: critical
tags:
    - attack.remote_access
    - attack.t1219
    - attack.credential_access
    - attack.t1003.006
    - attack.exfiltration
    - attack.t1567.002
KQL — Microsoft Sentinel / Defender
// Security Arsenal Hunt: AI-Orchestrated C2, VOLTA Egress & LLMjacking Infrastructure
// Scope: 30 days | Tables: DeviceNetworkEvents, DeviceProcessEvents
let Lookback = 30d;
let MaliciousIPs = dynamic(["104.234.26.163", "31.57.201.201", "38.246.114.118"]);
let MaliciousDomains = dynamic(["fastcdn.bar", "authgpt.info", "ai-billing.info", "url-y.uk", "artex.nhhh.de", "direct.465415.xyz"]);
let SuspiciousHostPatterns = dynamic([".nxcli.io", "465415.xyz"]);
// Stage 1: Network hits to known campaign infrastructure
let NetworkHits = DeviceNetworkEvents
    | where Timestamp > ago(Lookback)
    | where RemoteIP in (MaliciousIPs)
        or RemoteUrl has_any (MaliciousDomains)
        or RemoteUrl has_any (SuspiciousHostPatterns)
    | project Timestamp, DeviceName, InitiatingProcessFileName, InitiatingProcessCommandLine, RemoteIP, RemoteUrl, ActionType;
// Stage 2: Suspicious LLM API egress from unexpected processes (VOLTA victim-profiling behavior)
let LLMEgress = DeviceNetworkEvents
    | where Timestamp > ago(Lookback)
    | where RemoteUrl has_any ("openrouter.ai", "api.anthropic.com", "api.openai.com")
    | where InitiatingProcessFileName !in~ ("code.exe", "cursor.exe", "python.exe", "node.exe", "msedge.exe", "chrome.exe", "firefox.exe")
    | project Timestamp, DeviceName, InitiatingProcessFileName, InitiatingProcessCommandLine, RemoteUrl;
// Stage 3: Quick Assist followed by credential theft / exfil tooling (Teams impersonation chain)
let QuickAssistChain = DeviceProcessEvents
    | where Timestamp > ago(Lookback)
    | where FileName =~ "quickassist.exe"
    | join kind=inner (
        DeviceProcessEvents
        | where Timestamp > ago(Lookback)
        | where FileName has_any ("rclone.exe", "bun.exe")
            or ProcessCommandLine has_any ("dcsync", "PetitPotam", "lsadump")
    ) on DeviceName
    | where Timestamp1 between (Timestamp .. Timestamp + 30m)
    | project QuickAssistTime=Timestamp, ToolExecutionTime=Timestamp1, DeviceName, QuickAssistCmd=ProcessCommandLine, ToolCmd=ProcessCommandLine1;
union NetworkHits, LLMEgress, QuickAssistChain
| sort by Timestamp desc
PowerShell
# Security Arsenal IOC Hunt: CyberStrikeAI / VOLTA / LLMjacking / Teams-Impersonation Artifacts
# Run elevated. Checks network connections, DNS cache, suspicious runtimes, and persistence artifacts.

$ErrorActionPreference = 'SilentlyContinue'
$report = @()

# --- 1. Active & historical network connections to known C2 / phishing infra ---
$badIPs   = @('104.234.26.163','31.57.201.201','38.246.114.118')
$badNames = @('fastcdn.bar','authgpt.info','ai-billing.info','url-y.uk','artex.nhhh.de','direct.465415.xyz','nxcli.io')

Write-Host '[*] Checking active TCP connections...' -ForegroundColor Cyan
Get-NetTCPConnection | Where-Object { $badIPs -contains $_.RemoteAddress } | ForEach-Object {
    $report += [pscustomobject]@{ Check='ActiveC2Connection'; Detail="$($_.LocalAddress):$($_.LocalPort) -> $($_.RemoteAddress):$($_.RemotePort) [$($_.State)] PID=$($_.OwningProcess)" }
}

Write-Host '[*] Checking DNS client cache for malicious domain resolutions...' -ForegroundColor Cyan
Get-DnsClientCache | Where-Object { $n = $_.Name; $badNames | Where-Object { $n -like "*$_*" } } | ForEach-Object {
    $report += [pscustomobject]@{ Check='DNSCacheHit'; Detail="$($_.Name) -> $($_.Data)" }
}

# --- 2. VOLTA stealer artifacts: unexpected LLM API usage + known hash presence ---
Write-Host '[*] Scanning for VOLTA infostealer sample by hash...' -ForegroundColor Cyan
$voltaHash = '57ef4242666d78c50df145b34023df564f0fad38812d6a580e66433d427e2279'
$scanPaths = @("$env:TEMP","$env:APPDATA","$env:LOCALAPPDATA","$env:USERPROFILE\Downloads")
foreach ($p in $scanPaths) {
    Get-ChildItem -Path $p -Recurse -File -ErrorAction SilentlyContinue | ForEach-Object {
        if ((Get-FileHash $_.FullName -Algorithm SHA256).Hash -eq $voltaHash) {
            $report += [pscustomobject]@{ Check='VOLTAHashMatch'; Detail=$_.FullName }
        }
    }
}

# --- 3. Teams-impersonation artifacts: Quick Assist logs, portable runtimes, Rclone ---
Write-Host '[*] Checking Quick Assist execution artifacts...' -ForegroundColor Cyan
$qaLogs = "$env:LOCALAPPDATA\Microsoft\QuickAssist"
if (Test-Path $qaLogs) {
    $report += [pscustomobject]@{ Check='QuickAssistArtifacts'; Detail="Quick Assist profile present: $qaLogs — review for unauthorized sessions" }
}

Write-Host '[*] Hunting for portable runtimes (python/bun/java) in user-writable paths...' -ForegroundColor Cyan
Get-ChildItem -Path "$env:APPDATA","$env:LOCALAPPDATA","$env:TEMP" -Recurse -Include 'bun.exe','rclone.exe' -ErrorAction SilentlyContinue | ForEach-Object {
    $report += [pscustomobject]@{ Check='SuspiciousPortableTool'; Detail=$_.FullName }
}

# --- 4. Persistence check: run keys & scheduled tasks referencing temp/appdata binaries ---
Write-Host '[*] Auditing Run keys and scheduled tasks for stealer persistence...' -ForegroundColor Cyan
$runKeys = @('HKLM:\SOFTWARE\Microsoft\Windows\CurrentVersion\Run','HKCU:\SOFTWARE\Microsoft\Windows\CurrentVersion\Run')
foreach ($k in $runKeys) {
    (Get-ItemProperty $k).PSObject.Properties | Where-Object { $_.Value -match 'AppData|Temp|Public' } | ForEach-Object {
        $report += [pscustomobject]@{ Check='SuspiciousRunKey'; Detail="$k :: $($_.Name) = $($_.Value)" }
    }
}
Get-ScheduledTask | Where-Object { $_.Actions.Execute -match 'AppData|Temp' -and $_.TaskPath -notlike '\Microsoft*' } | ForEach-Object {
    $report += [pscustomobject]@{ Check='SuspiciousScheduledTask'; Detail="$($_.TaskName) -> $($_.Actions.Execute)" }
}

# --- 5. AI credential exposure: scan env vars of running processes for LLM API keys (LLMjacking risk) ---
Write-Host '[*] Checking for AI API keys exposed in common config locations...' -ForegroundColor Cyan
Get-ChildItem "$env:USERPROFILE\.env","$env:USERPROFILE\.config" -Recurse -ErrorAction SilentlyContinue |
    Select-String -Pattern 'OPENAI_API_KEY|ANTHROPIC_API_KEY|OPENROUTER' -List | ForEach-Object {
        $report += [pscustomobject]@{ Check='ExposedAICredential'; Detail="$($_.Path) — rotate key immediately if unauthorized egress is observed" }
    }

if ($report.Count -eq 0) { Write-Host '[+] No IOC artifacts found on this host.' -ForegroundColor Green }
else { $report | Format-Table -AutoSize; $report | Export-Csv -Path ".\ioc_hunt_$(Get-Date -Format yyyyMMdd_HHmm).csv" -NoTypeInformation; Write-Host "[!] $($report.Count) findings exported." -ForegroundColor Yellow }

Response Priorities

Immediate (0–4 hours)

  • Block all listed IOCs at egress: 104.234.26.163, 31.57.201.201, 38.246.114.118, fastcdn.bar, authgpt.info, ai-billing.info, url-y.uk, artex.nhhh.de, direct.465415.xyz, and the two identified *.nxcli.io phishing hostnames.
  • Push the full SHA256 hash set to EDR prevention lists; retro-hunt 90 days for execution and network hits.
  • Alert on any egress to OpenRouter/LLM API endpoints from non-approved processes — this is the highest-fidelity VOLTA indicator available.
  • Run the PowerShell hunt script across endpoints prioritized by education-sector relevance and recent Teams/Quick Assist activity.

24 Hours

  • Credential exposure triage: All five pulses involve credential theft. Assume compromise of any host with IOC hits: force resets for affected users, revoke session tokens (browser cookies are actively harvested by VOLTA), and invalidate OAuth grants.
  • Active Directory integrity check: If Quick Assist artifacts are found, immediately audit for DCSync replication events (Event ID 4662 with replication GUIDs), unexpected certificate enrollments (ESC1/ESC8 indicators), and new privileged account creation. Rotate krbtgt twice if DCSync is confirmed.
  • AI API key rotation: Inventory all AI platform credentials (OpenAI, Anthropic, OpenRouter, Fireworks, etc.) stored in environment variables, CI/CD secrets, and developer workstations. Rotate any key with unexplained usage spikes — LLMjacking victims often discover theft via billing anomalies.
  • Verify whether any corporate payment cards were used on the phishing domains; coordinate with finance for card reissue and fraud monitoring.

1 Week

  • Architectural hardening against AI-orchestrated attacks: CyberStrikeAI's adaptive, LLM-planned behavior means static detection rules will age poorly. Deploy behavioral analytics (UEBA) focused on command cadence anomalies and novel process-lineage patterns rather than fixed signatures.
  • AD CS remediation: Audit certificate templates for ESC1 (enrollee-supplied subject) and ESC8 (HTTP enrollment without EPA) misconfigurations; the Teams-impersonation actors specifically exploited these.
  • Remote support governance: Restrict Quick Assist via policy to an approved IT group, require ticket-number validation for all remote sessions, and train users that real help desk staff never initiate unsolicited Teams chats requesting connection codes.
  • Secrets management uplift: Migrate AI API keys out of .env files and environment variables into a managed vault with usage-based anomaly alerting — directly countering the LLMjacking wave.
  • Education-sector outreach: If operating in or adjacent to education in Africa or Latin America, share this briefing with peer institutions — CyberStrikeAI targeting is sector-specific and ongoing.

This briefing reflects live OTX pulse data as of 2026-11-03. Indicators should be aged appropriately: network IOCs decay within weeks; behavioral detections (LLM egress profiling, Quick Assist chains) retain long-term value.

Related Resources

Security Arsenal Incident Response Managed SOC & MDR Services AlertMonitor Threat Detection From The Dark Side Intel Hub

Is your security operations ready?

Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.