Back to Intelligence

PyPI agent-framework .pth Startup Hook Hijack, GCP LLMjacking Phishing & CyberPower PPB SYSTEM Abuse: OTX Detection Pack — 2026-11-20

SA
Security Arsenal Team
November 19, 2026
12 min read

Intelligence Category: Infostealer & Credential Theft Campaigns | Pulses Analyzed: 3 | TLP: WHITE


Threat Summary

Three pulses published through AlienVault OTX between 2026-11-19 and 2026-11-20 converge on a single thematic objective: wholesale theft and abuse of cloud credentials to monetize enterprise compute and identity infrastructure. While no single named adversary links all three campaigns, the tradecraft overlap is striking — each attack chain targets the credential material that gates cloud control planes (Azure, GCP, Active Directory), and each abuses trusted execution channels rather than exploiting memory-corruption vulnerabilities.

Campaign 1 — PyPI Supply Chain Poisoning. Microsoft's agent-framework package was hijacked on PyPI. Versions 1.26.1 and 1.26.2 (published 2026-11-17) were repacked locally with a malicious .pth startup hook. Because .pth files execute at every Python interpreter startup — not merely on import — any developer workstation, CI/CD runner, or container that installed the poisoned wheel silently executes a credential stealer harvesting environment variables containing Azure secrets, tokens, and service principal material. This is a developer-pipeline attack: the blast radius is every pipeline secret the interpreter can see.

Campaign 2 — Phishing-to-LLMjacking. Actors are purchasing sponsored search ads that imitate Google Cloud login portals (typosquat domains such as groosl.com, groonl.com, gboogl.com). Stolen privileged credentials are used to weaken GCP organizational policies, mint attacker-controlled service accounts with Gemini API access, and invite external identities — a cross-tenant exploitation pattern that turns the victim's billing account into free LLM compute for the adversary (LLMjacking). Secondary monetization paths include reselling API access on criminal marketplaces.

Campaign 3 — UPS Management Software as an Execution Channel. Intruders abused CyberPower PowerPanel Business (PPB) on a Windows domain controller, uploading malicious .cmd files via the PPB web console on TCP 3052. The PowerPanel service (ppbed.exe) executed them with SYSTEM privileges, enabling Active Directory extraction and deploying MSIL.Killav!gen1, an AV/EDR-killing payload delivered via DLL side-loading. A legitimate power-management feature became a remote code execution primitive on the most sensitive host in the enterprise.

Collective assessment: All three campaigns bypass perimeter controls by riding trusted software and identity flows. The common denominators are (1) credential theft as the primary objective, (2) abuse of legitimate administrative automation (.pth hooks, PPB power-event scripts, GCP service accounts), and (3) cloud monetization (Azure secrets, Gemini compute, AD dominance). Security teams should treat developer workstations, SaaS/IdP login paths, and infrastructure management tools as a unified identity attack surface.

Threat Actor / Malware Profile

PyPI agent-framework Credential Stealer

  • Attribution: Unknown. Reference research published by JFrog.
  • Distribution: Repacked legitimate wheels uploaded to PyPI as versions 1.26.1/1.26.2. Victims are poisoned through normal pip install / dependency resolution — no social engineering required.
  • Payload behavior: A .pth file dropped into site-packages runs an embedded one-liner on every Python startup. It enumerates environment variables, filtering for Azure-related keys (AZURE_*, client secrets, tenant IDs, connection strings, tokens) and exfiltrates them.
  • Persistence: The .pth startup hook itself is the persistence — it survives until the package is removed or the environment rebuilt, and triggers regardless of whether agent_framework is ever imported.
  • Anti-analysis: Blends into a legitimate signed-by-behavior package ecosystem; the malicious logic is a small hook inside an otherwise functional library, defeating diff-based review unless file hashes are compared against known-good releases.

LLMjacking Operation (GCP / Gemini API Abuse)

  • Attribution: Unknown. Reference research published by Wiz.
  • Distribution: Malvertising — sponsored search ads pointing to typosquatted GCP login clones (e.g., groosl.com, gcslom.com, gbrokl.space).
  • Payload behavior: No host malware required. Post-compromise activity is pure cloud API abuse: org-policy manipulation to disable guardrails, service-account creation with Gemini API scopes, external account invitation for cross-tenant persistence, and high-volume LLM inference billed to the victim.
  • C2 / access channel: Legitimate Google Cloud APIs from attacker infrastructure; detection must rely on cloud audit telemetry, not network signatures.
  • Persistence: Attacker-controlled service accounts, API keys, and external IAM bindings survive password resets unless explicitly revoked.

MSIL.Killav!gen1 via CyberPower PPB

  • Attribution: Unknown. Family signature: MSIL.Killav!gen1 (a .NET-based AV/EDR terminator).
  • Distribution / execution: Malicious .cmd scripts uploaded through the PPB web console (TCP 3052) and executed by ppbed.exe as SYSTEM during configured power events.
  • Payload behavior: DLL side-loading through trusted binaries; termination of security tooling; Active Directory database (NTDS.dit) extraction from the domain controller.
  • Persistence: PPB power-event script configuration acts as a durable, low-scrutiny execution mechanism; additional persistence likely via standard Windows autostart artifacts once SYSTEM is achieved.
  • Anti-analysis: DLL side-loading into signed processes, generic .NET obfuscation typical of Killav variants, and execution parented to a legitimate infrastructure service to defeat parent-child heuristics.

IOC Analysis

The pulse indicator sets fall into two operational classes:

File hashes (SHA-256 / SHA-1, 41+ indicators). Campaigns 1 and 3 are dominated by hash indicators: the poisoned agent-framework wheels/.pth payloads and the Killav!gen1 payloads plus side-loaded DLLs. Hashes are high-precision but brittle — they detect these exact builds. Operationalize them by: (1) pushing SHA-256 values into EDR blocklists and threat-intel platforms within the hour; (2) retro-hunting package caches, site-packages directories, artifact repositories (Artifactory/Nexus), and CI build logs for matching files; (3) correlating hash sightings with pip install telemetry to establish install timeline and blast radius. Do not rely on hashes alone for the PyPI campaign — repacked wheels rehash trivially, so pair with behavioral detections below.

Typosquat domains (8+ indicators). Campaign 2's domains (groosl.com, groonl.com, gboogl.com, gbrokl.space, gbrokl.xyz, gcoalg.com, gcslom.com, gpoosl.com) are phishing infrastructure. These should be blocked at DNS sinkhole, secure web gateway, and email/URL filtering layers, and hunted in proxy, DNS, and identity-provider sign-in logs. Because typosquats rotate rapidly, supplement exact-match blocking with fuzzy matching (Levenshtein distance ≤2 against google.com, gcloud, console.cloud.google.com) on newly-observed DNS queries from privileged users.

Tooling guidance: Hash and domain indicators ingest natively into MISP, OpenCTI, Microsoft Sentinel threat intelligence, and CrowdStrike/SentinelOne custom IOC feeds. For .pth file discovery, osquery or a simple filesystem sweep (below) is sufficient. For GCP, the authoritative telemetry is Cloud Audit Logs (Admin Activity) — specifically SetIamPolicy, CreateServiceAccount, CreateServiceAccountKey, and org-policy changes — streamed into your SIEM.

Detection Engineering

YAML
---
title: Malicious .pth Startup Hook in Python site-packages
id: 7f3a9c21-4d5e-4b8a-9c1f-agentpth001
status: experimental
description: Detects creation or modification of .pth files in Python site-packages, a technique used by the hijacked agent-framework PyPI package (v1.26.1/1.26.2) to execute a credential stealer on every interpreter startup.
author: Security Arsenal Threat Intelligence
date: 2026/11/20
references:
    - https://research.jfrog.com/post/agent-framework-hijack/
logsource:
    category: file_event
    product: windows
    definition: Linux auditd or Sysmon file creation telemetry equivalent applies on Linux/macOS
detection:
    selection_path:
        TargetFilename|endswith: '.pth'
    selection_dirs:
        TargetFilename|contains:
            - '\\site-packages\\'
            - '\\dist-packages\\'
            - '/site-packages/'
            - '/dist-packages/'
    filter_legit_installers:
        Image|endswith:
            - '\\pip.exe'
            - '\\python.exe'
            - '\\python3.exe'
    condition: selection_path and selection_dirs and not filter_legit_installers
falsepositives:
    - Legitimate package installation or upgrade outside pip (conda, poetry, manual builds)
level: high
tags:
    - attack.persistence
    - attack.t1546
    - attack.t1195.002
---
title: CyberPower PowerPanel Business Script Execution via ppbed.exe
id: 8b4c2d17-9e6f-4a7b-8d2e-ppbexec0002
status: experimental
description: Detects cmd.exe or script interpreters spawned by the CyberPower PowerPanel Business service (ppbed.exe), which threat actors abused to execute uploaded .cmd files with SYSTEM privileges on a domain controller.
author: Security Arsenal Threat Intelligence
date: 2026/11/20
references:
    - https://www.security.com/blog-post/power-management-software-attack-vector
logsource:
    category: process_creation
    product: windows
detection:
    selection_parent:
        ParentImage|endswith: '\\ppbed.exe'
    selection_child:
        Image|endswith:
            - '\\cmd.exe'
            - '\\powershell.exe'
            - '\\pwsh.exe'
            - '\\wscript.exe'
            - '\\cscript.exe'
            - '\\rundll32.exe'
    condition: selection_parent and selection_child
falsepositives:
    - Legitimate administrator-configured power-event shutdown or notification scripts (expected to be rare and static — baseline them)
level: critical
tags:
    - attack.privilege_escalation
    - attack.execution
    - attack.t1059.003
    - attack.t1574.002
---
title: Suspicious Network Connection from Python Interpreter to Uncommon Host
id: 2e9f1a54-7c3b-4d8a-b6f1-pyexfil0003
status: experimental
description: Detects python.exe / python3 processes making outbound network connections shortly after interpreter start without a script argument context consistent with automation, consistent with the agent-framework .pth credential stealer exfiltrating Azure environment variables.
author: Security Arsenal Threat Intelligence
date: 2026/11/20
logsource:
    category: network_connection
    product: windows
detection:
    selection_image:
        Image|endswith:
            - '\\python.exe'
            - '\\python3.exe'
            - '\\pythonw.exe'
    selection_dest_port:
        DestinationPort:
            - 443
            - 80
            - 8443
    filter_known_hosts:
        DestinationHostname|endswith:
            - '.pypi.org'
            - '.pythonhosted.org'
            - '.microsoft.com'
            - '.azure.com'
            - '.azureedge.net'
            - '.github.com'
            - '.githubusercontent.com'
    condition: selection_image and selection_dest_port and not filter_known_hosts
falsepositives:
    - Legitimate Python applications calling third-party APIs; tune allowlist per environment
level: medium
tags:
    - attack.exfiltration
    - attack.t1041
    - attack.t1552.001
KQL — Microsoft Sentinel / Defender
// Hunt: LLMjacking phishing infrastructure + agent-framework IOCs + PPB execution artifacts
// Scope: last 14 days across endpoints, network, and identity telemetry
let TyposquatDomains = dynamic(["groosl.com","groonl.com","gboogl.com","gbrokl.space","gbrokl.xyz","gcoalg.com","gcslom.com","gpoosl.com"]);
let MalHashes = dynamic([
  "d27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110",
  "3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9",
  "54d885b7d7161c9ed96489c66364c4bfbf635929166843bba10d97d46b5a2e1f",
  "325e7e9f0ec94e41ad1a1b7a3f18a6b8b5bffece2e2b97b8d3369eb967452f93",
  "39fa329f2a93f734b0bb8f5adb203ea3331d1ae479832a5015a54cb17aabd663",
  "21f28e8517a10ffe431f8d68f5875c85eca4001ae6d118322b0a41873998eb5e",
  "c275ccbe520ef3c0f02dc6a35389e427f9f1b4e0b12e140cd7bf7179ce143f02",
  "04206431c83d1d81726e56fffb5166c0a7f138eb366e5aa918dee54af1ef22b5",
  "12e716f48875befa978ff4a78a0b99a18b1809fc5bcc34c761158875da611aa8",
  "1949894c26e03d6f15ed7f4393a89dbd52f8acc62959d3d5e4f695d6a19c6d9c",
  "22b8f2776f1c941d8240ee4243da537147f8d2ee4d859a68038335da7b0d7688",
  "2776cfcdd46264043b823adaf37e0c85d0efd3b46d5fdbb72219535061d44b86",
  "3d73d5342a11dce35081e70a4ea75d39ad65bc6d480337bd33c1d2e0ba4090ee",
  "5b2285f788eec4d17bf18d85795d457f40f7598bbab2697919f5810f54a9c3fe",
  "76670eaadda9806c9aefa91c2ee4f2a33dcf8124574721a6ff072f19e7c03514"]);
let DomainHits = DeviceNetworkEvents
    | where TimeGenerated > ago(14d)
    | where RemoteUrl has_any (TyposquatDomains)
    | project TimeGenerated, DeviceName, InitiatingProcessAccountName, RemoteUrl, RemoteIP, InitiatingProcessFileName
    | extend HitType = "Phishing Domain Contact";
let HashHits = DeviceProcessEvents
    | where TimeGenerated > ago(14d)
    | where SHA256 has_any (MalHashes)
    | project TimeGenerated, DeviceName, AccountName, FileName, FolderPath, SHA256, ProcessCommandLine
    | extend HitType = "Malicious Hash Execution";
let PthHits = DeviceFileEvents
    | where TimeGenerated > ago(14d)
    | where FileName endswith ".pth" and FolderPath has_any ("site-packages","dist-packages")
    | where InitiatingProcessFileName !in~ ("pip.exe","python.exe","python3.exe")
    | project TimeGenerated, DeviceName, FileName, FolderPath, InitiatingProcessFileName, SHA256
    | extend HitType = "Suspicious .pth Write";
let PpbHits = DeviceProcessEvents
    | where TimeGenerated > ago(14d)
    | where InitiatingProcessFileName =~ "ppbed.exe"
    | where FileName in~ ("cmd.exe","powershell.exe","pwsh.exe","rundll32.exe","wscript.exe","cscript.exe")
    | project TimeGenerated, DeviceName, AccountName, FileName, ProcessCommandLine, InitiatingProcessCommandLine
    | extend HitType = "PPB SYSTEM Script Execution";
union DomainHits, HashHits, PthHits, PpbHits
| sort by TimeGenerated desc
PowerShell
# Security Arsenal IOC Hunt - agent-framework .pth hook / PPB abuse / typosquat DNS
# Run elevated on endpoints, build servers, and domain controllers.
$ErrorActionPreference = 'SilentlyContinue'
$report = @()

# 1) Enumerate .pth files in all Python site-packages / dist-packages locations
$pthRoots = @(
    "$env:LOCALAPPDATA\Programs\Python",
    "$env:ProgramFiles\Python*",
    "C:\Python*",
    "C:\ProgramData\Anaconda3",
    "C:\ProgramData\miniconda3",
    "$env:USERPROFILE\.venv*", "$env:USERPROFILE\venv*"
)
foreach ($root in $pthRoots) {
    Get-ChildItem -Path $root -Recurse -Filter "*.pth" -ErrorAction SilentlyContinue |
        Where-Object { $_.FullName -match 'site-packages|dist-packages' } |
        ForEach-Object {
            $content = Get-Content $_.FullName -Raw
            $suspicious = $content -match 'import |exec\(|base64|subprocess|urllib|requests|os\.environ'
            $report += [pscustomobject]@{
                Check   = 'PTH-StartupHook'
                Path    = $_.FullName
                Hash    = (Get-FileHash $_.FullName -Algorithm SHA256).Hash
                Flagged = $suspicious
                Snippet = ($content -replace "`r?`n",' ').Substring(0, [Math]::Min(200, $content.Length))
            }
        }
}

# 2) Detect installed agent-framework malicious versions
$pipList = & python -m pip list --format=json 2>$null | ConvertFrom-Json
$af = $pipList | Where-Object { $_.name -match 'agent-framework' -and $_.version -in @('1.26.1','1.26.2') }
if ($af) { $report += [pscustomobject]@{ Check='Poisoned-Package'; Path="agent-framework $($af.version)"; Hash=''; Flagged=$true; Snippet='Remove immediately and rotate all env-visible secrets' } }

# 3) CyberPower PPB: check service, listening port 3052, and staged .cmd files
$ppb = Get-Service -Name '*PowerPanel*','*ppbe*'
if ($ppb) {
    $report += [pscustomobject]@{ Check='PPB-Service'; Path=($ppb | Out-String).Trim(); Hash=''; Flagged=($env:COMPUTERNAME -match 'DC|AD'); Snippet='PPB present - verify necessity; DC presence is CRITICAL' }
}
$net3052 = Get-NetTCPConnection -LocalPort 3052 -State Listen
if ($net3052) { $report += [pscustomobject]@{ Check='PPB-Port3052'; Path='TCP/3052 LISTENING'; Hash=''; Flagged=$true; Snippet='Restrict console access; audit web console uploads' } }
Get-ChildItem -Path 'C:\Program Files*\CyberPower*','C:\ProgramData\CyberPower*' -Recurse -Include '*.cmd','*.bat','*.ps1' -ErrorAction SilentlyContinue |
    ForEach-Object { $report += [pscustomobject]@{ Check='PPB-Script'; Path=$_.FullName; Hash=(Get-FileHash $_.FullName -Algorithm SHA256).Hash; Flagged=$true; Snippet='Review for unauthorized power-event script' } }

# 4) Recent DNS cache hits for typosquat domains
$squats = 'groosl.com','groonl.com','gboogl.com','gbrokl.space','gbrokl.xyz','gcoalg.com','gcslom.com','gpoosl.com'
$dns = Get-DnsClientCache | Where-Object { $e=$_.Entry; $squats | Where-Object { $e -like "*$_*" } }
foreach ($d in $dns) { $report += [pscustomobject]@{ Check='Typosquat-DNS'; Path=$d.Entry; Hash=''; Flagged=$true; Snippet=$d.Data } }

$report | Format-List
$report | Export-Csv -Path ".\ioc_hunt_$(Get-Date -Format yyyyMMdd_HHmmss).csv" -NoTypeInformation
Write-Host "[+] Hunt complete. Flagged items: $(($report | Where-Object Flagged).Count)"

Response Priorities

Immediate (0–4 hours):

  • Push all 41+ file hashes and the 8 typosquat domains into EDR blocklists, DNS sinkholes, and secure web gateways.
  • Freeze agent-framework installs; pin or yank versions 1.26.1/1.26.2 in internal package indexes and artifact repositories; purge poisoned wheels from caches.
  • Run the .pth sweep on developer workstations, CI runners, and container base images; flag any .pth containing executable code (import, exec, base64).
  • On domain controllers and critical servers, inventory CyberPower PPB installations; block or ACL TCP 3052 immediately.

24 hours:

  • Treat every secret visible to any Python environment that ran the poisoned package as compromised: rotate Azure service principal secrets, connection strings, storage keys, and CI/CD pipeline variables. Prioritize secrets present in environment variables on affected hosts.
  • For GCP: force re-authentication of privileged users who may have clicked sponsored ads; audit Cloud Audit Logs for SetIamPolicy changes, new service accounts/keys, external IAM bindings, and org-policy modifications; revoke all suspect keys. Review Gemini API usage and billing anomalies.
  • If PPB artifacts are found on a DC: assume AD compromise — initiate krbtgt double-reset, audit for DCSync-style replication events (4662 with GUID 1131f6aa...), and review all privileged group memberships.

1 week:

  • Enforce package provenance controls: require hash-pinned dependencies (pip install --require-hashes), private PyPI mirrors with allowlisting, and Sigstore/SLSA attestation verification in CI.
  • Deploy FIDO2/phishing-resistant MFA for all cloud console access; block sponsored-ad click-through to identity portals via browser policy; add typosquat fuzzy-matching to DNS analytics.
  • Remove or harden infrastructure management tools (UPS, PDU, BMC consoles) from domain controllers and tier-0 assets; network-segment management interfaces; alert on any script-interpreter child process of infrastructure services.
  • Enable and centralize GCP Admin Activity and Azure Activity logs in the SIEM with alerting on service-account key creation and org-policy drift.

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.