Five OTX pulses published 2026-11-17/18 converge on a single theme: industrialized credential theft spanning the software supply chain, developer tooling, and consumer cryptocurrency users.
Cluster 1 — Supply-chain and developer targeting. Two pulses (Wiz and Socket reporting, corroborated by four overlapping SHA-256 IOCs) cover the compromise of Microsoft's agent-framework package on PyPI. Malicious versions 1.26.1 and 1.26.2 were live for roughly 50 minutes on 2026-11-17. The payload dropped Python .pth startup-hook files that execute automatically every time the interpreter initializes — a persistence and execution mechanism requiring no user interaction beyond installing the package. The stealer harvested Azure credentials, GitHub Actions secrets, OIDC tokens, managed-identity material, and CI/CD environment variables, exfiltrating to a webhook.site endpoint. A related pulse details 22 malicious GitHub repositories engineered to weaponize AI coding assistants: poisoned AGENTS.md / CLAUDE.md instruction files steer autonomous agents into executing a concealed setup.js credential-stealing payload.
Cluster 2 — Financial fraud. A Romania-focused campaign distributes polymorphic Android WebView trojans (Android.Riskware.SpyAgent.RB) and malicious iOS provisioning profiles via DGA-rotated phishing domains impersonating fake crypto exchanges branded "NACE Exchange" and "TroveX," harvesting banking and exchange credentials. A parallel operation lures hardware-wallet users (Coldcard, Ledger, Pera Algorand, Revolut, Xaman brands) into executing malicious BAT scripts from ZIP archives that silently install ConnectWise ScreenConnect — legitimate RMM software abused as a fully interactive backdoor.
The unifying objective is credential and session-token theft at scale: cloud and CI identities from developers and build pipelines, and financial credentials from end users. All actors are currently unattributed, and the two clusters show no infrastructure overlap at this time.
Threat Actor & Malware Profile
agent-framework PyPI Credential Stealer (Cluster 1)
- Attribution: Unknown; TTPs consistent with prior PyPI/npm token-harvesting operations.
- Distribution: Trojanized package versions 1.26.1/1.26.2 published to PyPI (likely via compromised maintainer credentials or a hijacked publishing token), live approximately 50 minutes before removal.
- Payload behavior: Drops a
.pthfile intosite-packages. Python'ssitemodule executes any line in a.pthfile beginning withimportat interpreter startup, granting code execution on every Python invocation — including CI build steps, test runners, and agent workloads. - Collection: Environment variables (Azure DevOps / GitHub Actions secrets), cloud configuration files, OIDC tokens, and managed-identity tokens from the Azure IMDS endpoint.
- C2 / exfiltration: HTTP(S) POST to
webhook.site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb— a legitimate webhook-testing service abused as a drop endpoint to blend with normal traffic. - Persistence: The
.pthfile itself; it re-executes on every interpreter start and can survive package uninstallation if left behind. - Anti-analysis: A 50-minute exposure window limits sample acquisition, and the payload lives in a file type most security tooling does not inspect.
AI-Agent Repository Lure (setup.js)
- Distribution: 22 malicious GitHub repositories; poisoned
AGENTS.md/CLAUDE.mdfiles cause AI coding assistants to auto-runsetup.jsduring project initialization. - Payload behavior: Node.js script collecting local credentials, tokens, SSH keys, and config files.
- Notable technique: Prompt-instruction-as-malware — the infection vector is trusted documentation, not executable code, evading repository malware scanning and human code review.
Android.Riskware.SpyAgent.RB (NACE Exchange / TroveX)
- Distribution: DGA-rotated phishing domains impersonating crypto exchanges; sideloaded Android APKs and malicious iOS provisioning profiles.
- Payload behavior: Polymorphic WebView trojan wrapping phishing pages in a native-app shell to harvest exchange credentials, PII, and one-time passcodes. Polymorphism mutates the APK hash per download, defeating static hash-based detection.
- C2: DGA-generated domains enabling rapid infrastructure rotation.
- Targeting: Romanian users, finance sector.
ScreenConnect RMM Backdoor (Fake Wallet Updates)
- Distribution: ZIP files masquerading as wallet firmware/app updates (Coldcard, Ledger, Pera, Revolut, Xaman), delivering malicious BAT scripts.
- Payload behavior: The BAT installer silently deploys ConnectWise ScreenConnect, granting operators full interactive remote access to drain wallets and harvest seed phrases.
- C2: ScreenConnect cloud relay (
*.screenconnect.com) — legitimate, signed infrastructure that bypasses most egress controls. - Anti-analysis: Living-off-the-land RMM abuse; there is no custom malware signature to detect.
IOC Analysis
The indicator set across these pulses is hash-heavy with a small but high-value network core:
- File hashes (SHA-256, SHA-1, MD5): The majority of indicators. The agent-framework cluster shares four SHA-256 hashes across both reporting pulses (Wiz and Socket), which corroborates the incident and raises confidence. Note the ScreenConnect pulse lists MD5s annotated with the corresponding SHA-1 of the same file ("MD5 of ...") — load both values into blocklists and do not treat the annotation text as part of the indicator. Caveat: the SpyAgent.RB trojan is polymorphic, so its hashes have a short shelf life — treat them as retro-hunt material rather than forward-looking blocking logic.
- URL —
https://webhook.site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb: The single most actionable indicator. This is the stealer's exfiltration endpoint. Blockwebhook.siteoutright on build agents, servers, and CI segments, and retro-hunt proxy/DNS/firewall logs for any request to this UUID path. Pivot on the UUID in urlscan.io to surface related submissions. - Domain —
exodus-wallet.pro: Fake-wallet lure domain. Block at DNS (RPZ) and proxy; pivot on passive DNS and registration metadata to enumerate the wider lure infrastructure. The Romanian campaign's DGA domains should be handled via threat-feed subscription and DNS-layer blocking rather than static lists.
Operationalization workflow:
- Push network IOCs to DNS RPZ, egress proxy, and firewall within the hour.
- Push file hashes to EDR blocklists and run a retro-hunt across the last 30 days of file and process telemetry, prioritizing build agents and developer workstations.
- Enrich hashes via MalwareBazaar hashlookup and VirusTotal; sandbox any recovered
.pthorsetup.jspayload — the base64 layers typical of these stealers decode cleanly in CyberChef. - Ingest the full pulse IOC sets into your TIP (MISP/OpenCTI) with decay scoring: high confidence on the webhook.site URL and package hashes; aggressive time-decay on the polymorphic Android hashes.
Detection Engineering
Sigma detections (Sysmon + Windows audit telemetry; tune filters to your environment):
---
title: Python .pth Startup Hook Written to site-packages
id: 7c9b2f4a-1d3e-4a5b-9c6d-2e8f0a1b3c4d
status: experimental
description: Detects creation of .pth files in Python site-packages/dist-packages directories. The agent-framework PyPI compromise abused .pth startup hooks to auto-execute a credential stealer on every interpreter start, harvesting Azure, GitHub Actions and OIDC tokens.
references:
- https://www.wiz.io/blog/malicious-version-of-microsofts-agent-framework-published-to-pypi
- https://socket.dev/blog/agent-framework-pypi-package-compromised-with-credential-stealer
author: Security Arsenal
date: 2026/11/18
modified: 2026/11/18
tags:
- attack.persistence
- attack.execution
- attack.t1546
- attack.t1195.002
logsource:
category: file_event
product: windows
detection:
selection:
TargetFilename|endswith: '.pth'
TargetFilename|contains:
- '\site-packages\'
- '\dist-packages\'
filter_package_managers:
Image|endswith:
- '\pip.exe'
- '\pip3.exe'
condition: selection and not 1 of filter_*
falsepositives:
- Legitimate packages that ship .pth files (e.g. distutils-precedence.pth)
level: high
---
title: Outbound Connection to Webhook.site Exfiltration Endpoint
id: 3a1d5e7f-2b4c-4d6e-8f0a-1b2c3d4e5f6a
status: experimental
description: Detects non-browser outbound connections to webhook.site, the exfiltration endpoint used by the agent-framework credential stealer for stolen Azure/CI-CD environment variables and OIDC tokens. High-fidelity on build agents, servers and CI runners where this service has no legitimate use.
references:
- https://www.wiz.io/blog/malicious-version-of-microsofts-agent-framework-published-to-pypi
author: Security Arsenal
date: 2026/11/18
modified: 2026/11/18
tags:
- attack.exfiltration
- attack.t1041
- attack.t1071.001
logsource:
category: network_connection
product: windows
detection:
selection:
DestinationHostname|endswith: '.webhook.site'
filter_browsers:
Image|endswith:
- '\chrome.exe'
- '\firefox.exe'
- '\msedge.exe'
condition: selection and not 1 of filter_*
falsepositives:
- Developers testing webhook integrations on workstations
level: high
---
title: ScreenConnect RMM Install via Batch Script from User-Writable Path
id: 5f6a7b8c-9d0e-4f1a-2b3c-4d5e6f7a8b9c
status: experimental
description: Detects batch scripts in temp/download paths that install or invoke ConnectWise ScreenConnect, matching the fake crypto-wallet update campaign (Coldcard, Ledger, Pera, Revolut, Xaman lures) deploying ScreenConnect as an interactive backdoor.
references:
- https://bartblaze.blogspot.com/2026/11/crypto-hardware-wallet-users-under.html
author: Security Arsenal
date: 2026/11/18
modified: 2026/11/18
tags:
- attack.execution
- attack.command_and_control
- attack.t1059.003
- attack.t1219
logsource:
category: process_creation
product: windows
detection:
selection_bat:
CommandLine|contains: '.bat'
selection_paths:
CommandLine|contains:
- '\AppData\Local\Temp\'
- '\Downloads\'
- '\AppData\Roaming\'
selection_rmm:
CommandLine|contains:
- 'ScreenConnect'
- 'msiexec'
condition: selection_bat and selection_paths and selection_rmm
falsepositives:
- IT-managed RMM deployment (tune to your software-delivery paths and accounts)
level: high
KQL hunt for Microsoft Sentinel / Defender XDR covering hash IOCs, .pth persistence, exfiltration, RMM abuse, and the AI-repo lure:
// Security Arsenal - OTX Hunt: supply-chain credential theft + RMM/WebView fraud (2026-11-18)
let Lookback = 30d;
let IOC_SHA256 = dynamic([
"d27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110",
"3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9",
"54d885b7d7161c9ed96489c66364c4bfbf635929166843bba10d97d46b5a2e1f",
"325e7e9f0ec94e41ad1a1b7a3f18a6b8b5bffece2e2b97b8d3369eb967452f93",
"39fa329f2a93f734b0bb8f5adb203ea3331d1ae479832a5015a54cb17aabd663",
"21f28e8517a10ffe431f8d68f5875c85eca4001ae6d118322b0a41873998eb5e",
"c275ccbe520ef3c0f02dc6a35389e427f9f1b4e0b12e140cd7bf7179ce143f02",
"047d20f53c60708d3dd638c973b6d7d97b8d3f38426466c64a409ee4af949ab7",
"3637e3b4402c935185c62258fa8e3718184470f92467d3cc9fc119fec2752606",
"52a1c0733e4e4306a8f708a95a40c14e6f6c099e4590fbf5c9ad240b206e1439"
]);
union isfuzzy=true
(DeviceProcessEvents
| where TimeGenerated >= ago(Lookback)
| where SHA256 in~ (IOC_SHA256)
| project Hunt="OTX hash executed", TimeGenerated, DeviceName, FileName, FolderPath, SHA256, AccountName=InitiatingProcessAccountName),
(DeviceFileEvents
| where TimeGenerated >= ago(Lookback)
| where SHA256 in~ (IOC_SHA256)
| project Hunt="OTX hash on disk", TimeGenerated, DeviceName, FileName, FolderPath, SHA256, AccountName=InitiatingProcessAccountName),
(DeviceFileEvents
| where TimeGenerated >= ago(Lookback)
| where FileName endswith ".pth"
| where FolderPath has_any ("site-packages", "dist-packages")
| where InitiatingProcessFileName !in~ ("pip.exe","pip3.exe")
| project Hunt="Suspicious .pth startup hook", TimeGenerated, DeviceName, FileName, FolderPath, AccountName=InitiatingProcessAccountName, InitiatingProcessCommandLine),
(DeviceNetworkEvents
| where TimeGenerated >= ago(Lookback)
| where RemoteUrl has "webhook.site"
| where InitiatingProcessFileName !in~ ("chrome.exe","msedge.exe","firefox.exe")
| project Hunt="webhook.site exfil", TimeGenerated, DeviceName, RemoteUrl, RemoteIP, AccountName=InitiatingProcessAccountName, InitiatingProcessCommandLine),
(DeviceNetworkEvents
| where TimeGenerated >= ago(Lookback)
| where RemoteUrl has_any ("exodus-wallet.pro","screenconnect.com")
| project Hunt="Wallet lure / RMM C2", TimeGenerated, DeviceName, RemoteUrl, RemoteIP, AccountName=InitiatingProcessAccountName),
(DeviceProcessEvents
| where TimeGenerated >= ago(Lookback)
| where ProcessCommandLine has "ScreenConnect" and ProcessCommandLine has_any (".bat","Temp","Downloads")
| project Hunt="ScreenConnect via BAT", TimeGenerated, DeviceName, FileName, ProcessCommandLine, AccountName=InitiatingProcessAccountName),
(DeviceProcessEvents
| where TimeGenerated >= ago(Lookback)
| where FileName =~ "node.exe" and ProcessCommandLine has "setup.js"
| project Hunt="AI-repo setup.js lure", TimeGenerated, DeviceName, FileName, ProcessCommandLine, AccountName=InitiatingProcessAccountName)
| order by TimeGenerated desc
PowerShell hunt script for endpoints and build agents without EDR coverage — checks package versions, .pth hooks, IOC hashes, exfil artifacts, and RMM persistence:
#requires -RunAsAdministrator
# Security Arsenal - OTX IOC/Behavior Hunt (2026-11-18)
# Scope: agent-framework PyPI stealer, .pth persistence, webhook.site exfil, ScreenConnect RMM, setup.js AI-repo lure
$ErrorActionPreference = 'SilentlyContinue'
$script:report = @()
function Add-Finding([string]$Category, [string]$Detail) {
$script:report += [pscustomobject]@{ Time=(Get-Date); Host=$env:COMPUTERNAME; Category=$Category; Detail=$Detail }
Write-Host "[!] $Category :: $Detail" -ForegroundColor Red
}
Write-Host "`n=== [1/6] agent-framework malicious versions ===" -ForegroundColor Cyan
$malVersions = @('1.26.1','1.26.2')
$pkgs = pip list --format=json 2>$null | ConvertFrom-Json
foreach ($p in $pkgs) {
if ($p.name -eq 'agent-framework' -and $malVersions -contains $p.version) {
Add-Finding 'MALICIOUS-PACKAGE' "agent-framework $($p.version) installed - treat ALL credentials on this host as compromised"
}
}
Write-Host "`n=== [2/6] Suspicious .pth startup hooks ===" -ForegroundColor Cyan
$pyRoots = @("$env:LOCALAPPDATA\Programs\Python","C:\Program Files\Python*","C:\Python3*","C:\ProgramData\Anaconda3","C:\ProgramData\miniconda3","C:\hostedtoolcache")
Get-ChildItem $pyRoots -Recurse -Filter *.pth | ForEach-Object {
$c = Get-Content $_.FullName -Raw
if ($c -match '^\s*import\s|os\.system|subprocess|urllib|requests|socket|base64') {
Add-Finding 'PTH-HOOK' "$($_.FullName) contains executable import/network logic"
}
}
Write-Host "`n=== [3/6] OTX file-hash IOCs ===" -ForegroundColor Cyan
$badSHA256 = @(
'd27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110',
'3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9',
'54d885b7d7161c9ed96489c66364c4bfbf635929166843bba10d97d46b5a2e1f',
'325e7e9f0ec94e41ad1a1b7a3f18a6b8b5bffece2e2b97b8d3369eb967452f93',
'39fa329f2a93f734b0bb8f5adb203ea3331d1ae479832a5015a54cb17aabd663',
'21f28e8517a10ffe431f8d68f5875c85eca4001ae6d118322b0a41873998eb5e',
'c275ccbe520ef3c0f02dc6a35389e427f9f1b4e0b12e140cd7bf7179ce143f02',
'047d20f53c60708d3dd638c973b6d7d97b8d3f38426466c64a409ee4af949ab7',
'3637e3b4402c935185c62258fa8e3718184470f92467d3cc9fc119fec2752606',
'52a1c0733e4e4306a8f708a95a40c14e6f6c099e4590fbf5c9ad240b206e1439'
)
$scanPaths = @("$env:TEMP","$env:USERPROFILE\Downloads","$env:USERPROFILE\.cache\pip","C:\Windows\Temp")
Get-ChildItem $scanPaths -Recurse -File | ForEach-Object {
$h = (Get-FileHash $_.FullName -Algorithm SHA256).Hash.ToLower()
if ($badSHA256 -contains $h) { Add-Finding 'IOC-HASH' "$($_.FullName) == $h" }
}
Write-Host "`n=== [4/6] Exfil / C2 DNS artifacts ===" -ForegroundColor Cyan
Get-DnsClientCache | Where-Object { $_.Entry -match 'webhook\.site|exodus-wallet\.pro|screenconnect\.com' } |
ForEach-Object { Add-Finding 'DNS-ARTIFACT' $_.Entry }
Write-Host "`n=== [5/6] ScreenConnect RMM persistence ===" -ForegroundColor Cyan
Get-CimInstance Win32_Service | Where-Object { $_.Name -match 'ScreenConnect' -or $_.DisplayName -match 'ScreenConnect' } |
ForEach-Object { Add-Finding 'RMM-SERVICE' "$($_.Name) -> $($_.PathName)" }
'HKLM:\SOFTWARE\Microsoft\Windows\CurrentVersion\Run','HKCU:\SOFTWARE\Microsoft\Windows\CurrentVersion\Run' | ForEach-Object {
$k = Get-ItemProperty $_
$k.PSObject.Properties | Where-Object { $_.Value -match 'ScreenConnect|\.bat' -and $_.Name -notlike 'PS*' } |
ForEach-Object { Add-Finding 'RUN-KEY' "$($_.Name)=$($_.Value)" }
}
Write-Host "`n=== [6/6] setup.js AI-repo lure ===" -ForegroundColor Cyan
Get-ChildItem "$env:USERPROFILE\Downloads","$env:USERPROFILE\source","$env:USERPROFILE\repos","$env:USERPROFILE\projects","$env:TEMP" -Recurse -Filter setup.js | ForEach-Object {
$c = Get-Content $_.FullName -Raw
if ($c -match 'child_process|https\.request|fetch\(|\.env|credentials|homedir') {
Add-Finding 'SETUPJS-LURE' "$($_.FullName) contains credential-access/exfil patterns"
}
}
Write-Host "`n=== Hunt complete: $($script:report.Count) finding(s) ===" -ForegroundColor Cyan
$script:report | Export-Csv -NoTypeInformation -Path ".\otx-hunt-$(Get-Date -Format yyyyMMdd-HHmm).csv"
Response Priorities
Immediate (0–4 hours)
- Block
webhook.siteat the egress proxy/firewall for all server, build-agent, and CI segments; blockexodus-wallet.proand the wallet-lure domains at DNS RPZ. - Push all SHA-256/SHA-1/MD5 hashes to EDR blocklists; run the KQL/PowerShell retro-hunt across 30 days of telemetry, prioritizing build agents, CI runners, and developer workstations.
- Inventory every host (including ephemeral CI images and container base images) that installed
agent-framework1.26.1 or 1.26.2; purge those versions from internal PyPI mirrors and artifact caches. - Isolate any host with a suspicious
.pthfile, matching hash, or webhook.site connection pending credential rotation.
Within 24 hours
- Assume compromise of every secret on any host that ran the malicious package. Rotate Azure service-principal secrets, GitHub Actions and Azure DevOps pipeline secrets, and any cloud keys present in environment variables; revoke outstanding OIDC-issued tokens; audit Entra ID sign-in and token-use logs for replay from unexpected IPs.
- Audit package-proxy and PyPI download logs to scope exactly which pipelines pulled versions 1.26.1/1.26.2 during the 50-minute window.
- Force credential resets for any user who installed the fake wallet apps or Android APKs; for managed mobile fleets, query MDM for the SpyAgent.RB package hashes and unauthorized iOS provisioning profiles.
- Open fraud-monitoring cases for accounts matching the Romanian campaign's victim profile.
Within 1 week
- Enforce package provenance controls: pin dependency versions, verify hashes, and route all installs through an internal mirror with a 24–72 hour quarantine/cooldown before newly published versions become installable.
- Add
.pthfile creation insite-packagesto permanent FIM/EDR watchlists; evaluate running CI interpreters with-S(nositeprocessing) where feasible. - Implement RMM allowlisting via AppLocker/WDAC — block ScreenConnect and other unauthorized remote-access tooling enterprise-wide.
- Establish AI-agent governance: sandbox coding assistants, require human approval before agents execute project scripts (
setup.js, Makefile targets), and treatAGENTS.md/CLAUDE.mdas untrusted input in code review. - Distribute the lure-brand list to fraud and helpdesk teams and subscribe DNS security to DGA-domain feeds covering the Romanian campaign.
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.