Back to Intelligence

ScreenConnect Spray Campaign, agent-framework PyPI Credential Stealer & Trojanized AI-Agent Repos: OTX Pulse Analysis — Enterprise Detection Pack

SA
Security Arsenal Team
November 18, 2026
11 min read

The five pulses pulled from AlienVault OTX on 2026-11-19 do not describe five isolated incidents — they describe a single convergent trend: adversaries are industrializing credential and session theft by abusing trusted tooling and trusted supply chains, then monetizing access against financial and developer targets.

Three distinct attack chains stand out:

  1. Supply-chain credential harvesting against developers and cloud CI/CD. The compromise of Microsoft's agent-framework PyPI package (versions 1.26.1 and 1.26.2) used .pth startup hooks that execute silently on every Python interpreter initialization, scraping Azure tokens, GitHub Actions secrets, and CI/CD environment variables. In parallel, a separate campaign seeded 22 malicious GitHub repositories with a concealed setup.js payload hidden inside AI-agent instruction files (AGENTS.md, CLAUDE.md) — weaponizing the trust developers place in autonomous coding assistants.

  2. Remote access tooling weaponized against MSPs and crypto holders. CVE-2026-84869 (CVSS 9.9) in ConnectWise ScreenConnect was sprayed across 42 MSP servers with 688,000 malicious file-transfer requests, letting malicious guests push and execute files on technician workstations with elevated privileges — dropping the SCKit malware family. A companion campaign lured hardware-wallet users (Coldcard, Ledger, Pera, Revolut, Xaman) into running trojanized ZIP/BAT "wallet updates" that silently installed ScreenConnect as a persistent remote-access beachhead.

  3. DGA-driven mobile financial fraud. The NACE Exchange / TroveX operation targeted Romanian users with polymorphic WebView trojans (Android.Riskware.SpyAgent.RB), rotating phishing domains via domain generation algorithms and distributing malicious APKs and iOS provisioning profiles impersonating crypto exchanges.

The collective objective is unambiguous: credential, session-token, and wallet-access theft at scale, using legitimate software (ScreenConnect, PyPI packaging, GitHub) as the delivery vehicle to defeat reputation-based controls. Expect harvested developer credentials to be brokered or used for follow-on cloud compromise.

Threat Actor / Malware Profile

SCKit (ScreenConnect exploitation toolset)

  • Distribution: Delivered via CVE-2026-84869 file-transfer abuse during active ScreenConnect sessions; also installed by malicious BAT scripts inside fake crypto-wallet update ZIPs.
  • Payload behavior: Establishes unauthorized remote access on technician or victim workstations under the guise of legitimate remote support, enabling hands-on-keyboard theft of credentials, wallet seeds, and session data.
  • C2 communication: Rides ScreenConnect's own relay infrastructure, blended with attacker-controlled staging at masterdarky.com and IPv4 infrastructure including 194.15.36.33, 38.18.228.173, 38.18.228.165, 172.81.61.149, and 154.30.4.248.
  • Persistence: ScreenConnect installs as a service (ScreenConnect Client / ConnectWiseControl), providing durable SYSTEM-level access that survives reboots.
  • Anti-analysis: Uses a signed, legitimate RMM binary — evading most signature and reputation engines; detection must be behavioral.

agent-framework PyPI credential stealer (versions 1.26.1, 1.26.2)

  • Distribution: Poisoned package published to PyPI on 2026-11-17; pulled transitively into developer workstations and CI/CD runners.
  • Payload behavior: Python .pth startup hook executes on every interpreter launch — no import required. Harvests Azure CLI tokens, GitHub Actions runner secrets, environment variables, cloud metadata identity files, and config files, then exfiltrates.
  • Persistence: The .pth file in site-packages IS the persistence mechanism — it survives as long as the environment exists and fires on every Python invocation.
  • Anti-analysis: Execution is indistinguishable from normal interpreter startup; payload only triggers in interpreter context, defeating static sandbox detonation of the wheel.

AI-agent repository malware (setup.js / AGENTS.md poisoning)

  • Distribution: 22 malicious GitHub repositories crafted to be ingested by AI coding assistants; instruction files (AGENTS.md, CLAUDE.md) steer the agent toward executing a concealed setup.js.
  • Payload behavior: Node.js credential stealer targeting developer secrets, SSH keys, tokens, and wallet data.
  • Anti-analysis: The malicious instruction is semantic — invisible to file-signature scanning and only "executed" by the AI agent acting as a confused deputy.

Android.Riskware.SpyAgent.RB (NACE Exchange / TroveX)

  • Distribution: DGA-rotated phishing domains impersonating crypto exchanges; sideloaded APKs and iOS provisioning profiles pushed to Romanian users.
  • Payload behavior: Polymorphic WebView trojan that renders cloned exchange login flows, harvesting credentials and PII.
  • Anti-analysis: Polymorphic repackaging per-download defeats hash-based mobile AV lookups; DGA defeats static domain blocklists.

IOC Analysis

The indicator set spans four operational classes:

  • IPv4 (ScreenConnect/SCKit infrastructure): 194.15.36.33, 38.18.228.173, 38.18.228.165, 172.81.61.149, 154.30.4.248. These are block-and-hunt indicators — push to egress firewall, proxy, and EDR network-block lists, then retro-hunt NetFlow/proxy logs for 90 days. Note the 38.18.228.0/24 clustering — consider a monitored (not blanket-blocked) watch on the /24 after validating no business use.
  • Domains: masterdarky.com (SCKit staging), exodus-wallet.pro (wallet-phish lure). Sinkhole at DNS resolvers and add to secure-web-gateway block categories. For the DGA campaign, static domain indicators decay within hours — rely on the file hashes and behavioral rules instead.
  • File hashes (MD5/SHA1/SHA256): The PyPI stealer wheels, SCKit payloads, SpyAgent.RB APKs, and trojanized wallet-update archives. Operationalize SHA256 values into EDR blocklists (block-on-write), and sweep MD5/SHA1 into your threat-intel platform for retro matching against historical file-creation telemetry. Note the IOC convention "MD5 of " in the wallet campaign — these MD5s correspond to the SHA1-named artifacts; load both values.
  • CVE-2026-84869: Not a traditional IOC — operationalize as an exposure query: enumerate all ScreenConnect/ConnectWise instances (internal and MSP-managed) and confirm patched builds; treat any unpatched instance as an active incident given the observed spray volume.

Tooling: normalize indicators through your TIP (MISP/OpenCTI) with OTX as the feed source; use yara/thor for endpoint sweeps with the hash sets; decode and triage the .pth persistence artifact by enumerating site-packages for recently modified .pth files; validate mobile APK hashes against VirusTotal/Koodous for sandbox detonation.

Detection Engineering

YAML
---
title: Python Startup Hook (.pth) Creation in Site-Packages - PyPI Supply Chain Persistence
id: 7d3f1a2e-4b5c-4d6e-9f0a-1b2c3d4e5f01
status: experimental
description: Detects creation or modification of .pth startup-hook files in Python site-packages, the persistence and execution mechanism used by the compromised agent-framework 1.26.1/1.26.2 PyPI packages to silently harvest Azure, GitHub Actions, and CI/CD credentials on interpreter initialization.
author: Security Arsenal
references:
    - https://socket.dev/blog/agent-framework-pypi-package-compromised-with-credential-stealer
date: 2026/11/19
logsource:
    category: file_event
    product: windows
    definition: Linux/macOS equivalents via auditd or osquery file_events on site-packages directories
detection:
    selection_path:
        TargetFilename|contains:
            - '\site-packages\'
            - '/site-packages/'
            - '\dist-packages\'
            - '/dist-packages/'
    selection_ext:
        TargetFilename|endswith: '.pth'
    condition: selection_path and selection_ext
falsepositives:
    - Legitimate packages shipping .pth files (e.g., distutils-precedence.pth) during pip installs
level: high
tags:
    - attack.persistence
    - attack.t1546
    - attack.t1195.002
---
title: Unauthorized ScreenConnect Client Installation or Service Execution
id: 8e4a2b3f-5c6d-4e7f-0a1b-2c3d4e5f6a02
status: experimental
description: Detects installation or execution of ConnectWise ScreenConnect client components, associated with CVE-2026-84869 technician-workstation exploitation (SCKit) and trojanized crypto-wallet update BAT scripts deploying ScreenConnect as a covert RMM beachhead. Alert where ScreenConnect is not an approved RMM tool.
author: Security Arsenal
references:
    - https://www.huntress.com/blog/cve-2026-84869-screenconnect-technician-exploitation
    - https://bartblaze.blogspot.com/2026/11/crypto-hardware-wallet-users-under.html
date: 2026/11/19
logsource:
    category: process_creation
    product: windows
detection:
    selection_img:
        Image|contains:
            - '\ScreenConnect'
            - '\ConnectWiseControl'
    selection_parent_script:
        ParentImage|endswith:
            - '\cmd.exe'
            - '\powershell.exe'
            - '\wscript.exe'
            - '\mshta.exe'
    selection_service:
        CommandLine|contains:
            - 'ScreenConnect.ClientService'
            - 'ScreenConnect Windows Client'
    condition: selection_img and (selection_parent_script or selection_service)
falsepositives:
    - Approved MSP/IT remote support deployments - baseline authorized ScreenConnect instances before enabling
level: high
tags:
    - attack.command_and_control
    - attack.t1219
    - attack.t1059
---
title: Node.js Script Execution Referencing AI Agent Instruction Files or Concealed setup.js
id: 9f5b3c4a-6d7e-4f8a-1b2c-3d4e5f6a7b03
status: experimental
description: Detects Node.js execution of setup.js or scripts staged alongside AGENTS.md/CLAUDE.md instruction files, matching the campaign targeting developers via 22 malicious GitHub repositories engineered to have AI coding assistants execute a concealed credential-stealing payload.
author: Security Arsenal
references:
    - https://www.bitdefender.com/en-us/blog/labs/how-ai-agents-become-new-targets-in-github-malware-campaign
date: 2026/11/19
logsource:
    category: process_creation
    product: windows
detection:
    selection_img:
        Image|endswith:
            - '\node.exe'
            - '\nodejs.exe'
    selection_cli:
        CommandLine|contains:
            - 'setup.js'
            - 'AGENTS.md'
            - 'CLAUDE.md'
    filter_devtools:
        ParentImage|endswith:
            - '\code.exe'
            - '\npm.exe'
    condition: selection_img and selection_cli and not filter_devtools
falsepositives:
    - Legitimate project setup scripts - tune with known-good repository paths
level: medium
tags:
    - attack.execution
    - attack.t1059.007
    - attack.t1195
KQL — Microsoft Sentinel / Defender
// Hunt: ScreenConnect/SCKit C2 + agent-framework PyPI stealer artifact exposure
// Microsoft Sentinel / Defender XDR - last 30 days
let C2_IPs = dynamic(["194.15.36.33","38.18.228.173","38.18.228.165","172.81.61.149","154.30.4.248"]);
let MalDomains = dynamic(["masterdarky.com","exodus-wallet.pro"]);
let NetworkHits =
    DeviceNetworkEvents
    | where TimeGenerated > ago(30d)
    | where RemoteIP in (C2_IPs) or RemoteUrl has_any (MalDomains)
    | project TimeGenerated, DeviceName, InitiatingProcessFileName, InitiatingProcessCommandLine, RemoteIP, RemoteUrl, RemotePort, ActionType
    | extend Hunt = "ScreenConnect/SCKit C2";
let PthArtifacts =
    DeviceFileEvents
    | where TimeGenerated > ago(30d)
    | where FolderPath has_any ("site-packages", "dist-packages") and FileName endswith ".pth"
    | where not(FileName in~ ("distutils-precedence.pth"))  // known-good baseline; extend as needed
    | project TimeGenerated, DeviceName, FolderPath, FileName, SHA256, InitiatingProcessFileName
    | extend Hunt = "PyPI .pth startup hook persistence";
let ScreenConnectInstall =
    DeviceProcessEvents
    | where TimeGenerated > ago(30d)
    | where (FileName has_any ("ScreenConnect", "ConnectWiseControl")
        or ProcessCommandLine has_any ("ScreenConnect.ClientService", "ScreenConnect Windows Client"))
    | where InitiatingProcessFileName has_any ("cmd.exe","powershell.exe","wscript.exe","mshta.exe","node.exe")
    | project TimeGenerated, DeviceName, FileName, ProcessCommandLine, InitiatingProcessFileName, InitiatingProcessCommandLine, SHA256
    | extend Hunt = "Script-spawned ScreenConnect install";
union NetworkHits, PthArtifacts, ScreenConnectInstall
| sort by TimeGenerated desc
PowerShell
# Security Arsenal - OTX Pulse IOC Sweep (2026-11-19)
# Hunts: PyPI .pth persistence, ScreenConnect/SCKit artifacts, AI-agent repo payload, C2 connections
# Run elevated on suspect endpoints. Read-only; no remediation performed.

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

# 1) Suspicious .pth startup hooks (agent-framework 1.26.1/1.26.2 technique)
$knownGoodPth = @('distutils-precedence.pth')
$pthHits = Get-ChildItem -Path "$env:LOCALAPPDATA\Programs\Python","C:\Python*","C:\Program Files\Python*" -Recurse -Filter *.pth |
    Where-Object { $_.Name -notin $knownGoodPth -and $_.LastWriteTime -gt (Get-Date).AddDays(-7) } |
    Select-Object FullName, LastWriteTime, @{N='Content';E={(Get-Content $_.FullName -Raw).Trim()}}
foreach ($h in $pthHits) { $report += [pscustomobject]@{Check='PthStartupHook'; Severity='HIGH'; Detail="$($h.FullName) | $($h.Content)"} }

# Also flag the malicious package versions directly
$pipList = (pip list 2>$null) | Select-String 'agent-framework'
if ($pipList -match '1\.26\.[12]') { $report += [pscustomobject]@{Check='MaliciousPyPIPackage'; Severity='CRITICAL'; Detail="$pipList - REMOVE IMMEDIATELY and rotate Azure/GitHub/CI-CD credentials"} }

# 2) ScreenConnect / SCKit persistence (services, run keys, install dirs)
$scSvc = Get-CimInstance Win32_Service | Where-Object { $_.Name -match 'ScreenConnect|ConnectWise' -or $_.PathName -match 'ScreenConnect|ConnectWiseControl' }
foreach ($s in $scSvc) { $report += [pscustomobject]@{Check='ScreenConnectService'; Severity='HIGH'; Detail="$($s.Name) | $($s.State) | $($s.PathName)"} }
$runKeys = 'HKLM:\SOFTWARE\Microsoft\Windows\CurrentVersion\Run','HKCU:\SOFTWARE\Microsoft\Windows\CurrentVersion\Run'
foreach ($k in $runKeys) {
    (Get-ItemProperty $k).PSObject.Properties | Where-Object { $_.Value -match 'ScreenConnect|ConnectWise' } |
        ForEach-Object { $report += [pscustomobject]@{Check='RunKeyPersistence'; Severity='HIGH'; Detail="$k\$($_.Name) = $($_.Value)"} }
}
foreach ($p in "$env:ProgramFiles\ScreenConnect*","${env:ProgramFiles(x86)}\ScreenConnect*","$env:TEMP\ScreenConnect*") {
    if (Test-Path $p) { $report += [pscustomobject]@{Check='ScreenConnectPath'; Severity='HIGH'; Detail=$p} }
}

# 3) AI-agent repo payload artifacts (concealed setup.js near instruction files)
$agentHits = Get-ChildItem -Path "$env:USERPROFILE\source","$env:USERPROFILE\repos","$env:USERPROFILE\dev","$env:USERPROFILE\Documents" -Recurse -Filter 'setup.js' -Depth 4 |
    Where-Object { Test-Path (Join-Path $_.DirectoryName 'AGENTS.md') -or Test-Path (Join-Path $_.DirectoryName 'CLAUDE.md') } |
    Select-Object FullName, LastWriteTime
foreach ($a in $agentHits) { $report += [pscustomobject]@{Check='AIAgentRepoPayload'; Severity='MEDIUM'; Detail=$a.FullName} }

# 4) Active/historical connections to SCKit C2 infrastructure
$c2 = '194.15.36.33','38.18.228.173','38.18.228.165','172.81.61.149','154.30.4.248'
$conns = Get-NetTCPConnection | Where-Object { $_.RemoteAddress -in $c2 } |
    Select-Object RemoteAddress, RemotePort, State, OwningProcess, @{N='Process';E={(Get-Process -Id $_.OwningProcess).ProcessName}}
foreach ($c in $conns) { $report += [pscustomobject]@{Check='C2Connection'; Severity='CRITICAL'; Detail="$($c.RemoteAddress):$($c.RemotePort) $($c.State) by $($c.Process)"} }
$dns = Get-DnsClientCache | Where-Object { $_.Entry -match 'masterdarky\.com|exodus-wallet\.pro' }
foreach ($d in $dns) { $report += [pscustomobject]@{Check='MaliciousDomainCache'; Severity='CRITICAL'; Detail=$d.Entry} }

if ($report.Count -eq 0) { Write-Output '[+] No indicators found.' } else { $report | Sort-Object Severity | Format-Table -AutoSize; $report | Export-Csv ".\otx_hunt_$(Get-Date -f yyyyMMdd_HHmm).csv" -NoTypeInformation }

Response Priorities

Immediate (0-4 hours)

  • Block the five IPv4 C2 addresses and both domains (masterdarky.com, exodus-wallet.pro) at egress firewall, proxy, and DNS resolver; load all SHA256/MD5/SHA1 hashes into EDR block-on-write lists.
  • If agent-framework 1.26.1 or 1.26.2 exists anywhere in your estate (workstations, CI/CD runners, containers): treat as confirmed credential compromise. Kill affected environments, rotate ALL Azure tokens, GitHub Actions secrets, and CI/CD variables those environments could access — do not wait for exfiltration confirmation.
  • Pin/audit all ScreenConnect instances; if CVE-2026-84869 is unpatched, take the instance offline — 688,000 observed exploit requests means scanning is opportunistic and automated.
  • Sweep for .pth startup hooks and script-spawned ScreenConnect installs using the KQL and PowerShell above.

24 Hours

  • Force credential rotation for any identity present on endpoints positive for ScreenConnect/SCKit or the PyPI stealer: revoke Azure AD refresh tokens, reset passwords, re-issue SSH keys and API tokens. Credential-stealing malware means session tokens are already in adversary hands — password resets alone are insufficient.
  • Audit GitHub for recently cloned repos containing AGENTS.md/CLAUDE.md plus setup.js; review Actions logs for unexpected secret access. Alert developers that AI coding assistants can be weaponized as execution proxies.
  • Review MSP-facing remote sessions for anomalous file transfers (guest-initiated pushes to technician machines) across the October-November 2026 window.

1 Week

  • Enforce a package-provenance policy: private PyPI mirror with allow-listing, hash-pinned dependencies, and mandatory SBOM diffing on version bumps — the 1.26.1/1.26.2 poison releases would have been caught by a block-on-new-version gate.
  • Establish an RMM allow-list: ScreenConnect (and any remote-access binary) blocked by default via WDAC/AppLocker unless explicitly approved; alert on any non-allow-listed RMM installation.
  • Deploy DGA-resistant controls (DNS firewalling with ML-based domain scoring) for the mobile-phishing vector, and push user awareness on sideloaded APKs/iOS provisioning profiles for crypto apps — targeted at any Romania-based staff or customers.
  • Require human approval before AI coding agents execute any script discovered in a repository; sandbox agent execution environments with no access to credential stores.

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.

ScreenConnect Spray Campaign, agent-framework PyPI Credential Stealer & Trojanized AI-Agent Repos: OTX Pulse Analysis — Enterprise Detection Pack | Security Arsenal | Security Arsenal