Back to Intelligence

PHP2Shell AI-Agent Exploitation, agent-framework PyPI Hijack & CyberLeek Extortion: OTX Credential-Theft Campaign Analysis — Enterprise Detection Pack

SA
Security Arsenal Team
November 19, 2026
12 min read

Five fresh OTX pulses published November 18–19, 2026 paint a coherent and alarming picture: credential theft operations are now being executed at machine speed, and the very AI tooling developers rely on has become both the weapon and the target.

Three distinct but thematically linked campaigns dominate this reporting window:

  1. PHP2Shell Credential-Harvesting Campaign — Within days of public disclosure (October 21, 2026) of CVE-2026-109618 and CVE-2026-109619, an unknown adversary deployed an AI-agent-driven exploitation pipeline against PHP2Shell-vulnerable infrastructure. The operation used a custom job dispatcher backed by DeepSeek V4.1 endpoints, ingesting target lists from Shodan and autonomously executing exploitation, webshell deployment, and credential harvesting. All activity funneled through a single consolidated infrastructure node — 206.245.153[.]56 — serving simultaneously as C2, payload host, and exfiltration collector. Targeted verticals: Finance, Education, and Technology. The 'day-one exploitation by AI agents' pattern marks a structural shift in the exploit-to-weaponization timeline: defenders no longer have a patch grace period measured in weeks.

  2. Microsoft agent-framework PyPI Supply-Chain Compromise — Two corroborating pulses (JFrog and Socket research) confirm that versions 1.26.1 and 1.26.2 of Microsoft's agent-framework package on PyPI, released November 17, 2026, were hijacked. The attacker repacked legitimate wheels and injected a .pth startup hook — a particularly insidious technique that executes the credential stealer every time the Python interpreter starts, regardless of whether the package is ever imported. The payload harvests Azure tokens, GitHub Actions secrets, CI/CD environment variables, configuration files, and cloud identity material. Any developer workstation or build pipeline that installed either version in the Nov 17–18 window must be treated as compromised.

  3. Malicious GitHub Repositories Targeting AI Coding Assistants — Bitdefender-tracked campaign abusing the instruction files AI agents consume (AGENTS.md, CLAUDE.md). Twenty-two malicious repositories embed a concealed setup.js payload that executes when an AI coding assistant processes the repo — effectively prompt-injecting the agent into running a credential stealer. This is a new initial-access surface: the human never runs the code; their AI assistant does.

The fourth pulse profiles CyberLeek, a newly emerged extortion group notable for abandoning traditional ransom economics in favor of engagement-driven monetization via a Solana token ($CYBERLEEK), following its leak of proprietary Rockstar Games content. While not a credential-stealing operation per se, its model — leak first, monetize attention — increases the incentive for upstream credential-harvesting campaigns like the three above, which supply initial access and data to such groups.

Collective objective: bulk harvesting of cloud, CI/CD, and developer credentials for resale, extortion leverage, and downstream intrusions. The unifying thread is the developer toolchain and AI-agent ecosystem as the attack surface.

Threat Actor / Malware Profile

PHP2Shell AI-Agent Campaign (Unknown Actor)

  • Distribution/Access: Automated scanning via Shodan-derived target lists; exploitation of CVE-2026-109618 / CVE-2026-109619 in PHP2Shell within days of disclosure. Additional CVEs in the indicator set (CVE-2025-55182, CVE-2025-66478, CVE-2026-31431, and legacy CVE-2016-5195 'Dirty COW' for Linux privilege escalation) indicate a broad exploitation toolkit.
  • Orchestration: Custom job dispatcher calling DeepSeek V4.1 endpoints to drive exploitation decisions — automated target triage, payload selection, and likely error recovery.
  • Payload behavior: Webshell deployment for persistence; credential harvesting from compromised hosts; data staged back to the collector node.
  • C2: Consolidated single-node infrastructure at 206.245.153[.]56 handling command-and-control, payload delivery, and exfiltration — a detection gift, but expect rotation.
  • Persistence: Webshells on compromised PHP-facing servers; potential local privilege escalation via Dirty COW on older Linux kernels.

agent-framework Malicious Wheels (Unknown Actor)

  • Distribution: Trojanized PyPI releases 1.26.1 and 1.26.2 (Nov 17, 2026). Repacked wheels preserving legitimate functionality to evade detection.
  • Payload behavior: Credential stealer harvesting Azure tokens, GitHub Actions secrets, CI/CD environment variables, cloud identity configs.
  • Persistence/Execution mechanism: .pth file dropped into site-packages — Python executes lines in .pth files beginning with import at every interpreter startup. No import of the package required; any python invocation triggers the stealer.
  • Anti-analysis: Blends into legitimate package; no obviously malicious module imports; startup-hook execution evades import-time monitoring tools.

Malicious AI-Agent Repos (Unknown Actor)

  • Distribution: 22 GitHub repositories seeded with poisoned AGENTS.md / CLAUDE.md instruction files.
  • Payload: Concealed setup.js (Node.js) credential stealer executed when AI coding assistants process the repository.
  • Anti-analysis: Payload hidden behind agent-instruction context — human code reviewers focus on source files; the trap is in the agent-facing metadata.

CyberLeek (Extortion Group)

  • Model: Data-leak extortion with engagement-tokenized monetization ($CYBERLEEK on Solana) rather than ransom negotiation. Targets major software publishers and gaming industry. Watch for this group as a buyer/amplifier of credential-derived access.

IOC Analysis

The indicator set across these pulses comprises:

  • IPv4 (1): 206.245.153.56 — PHP2Shell campaign C2/payload/exfil node. Immediate egress block. Because this node serves three roles, any observed connection — inbound fetch or outbound POST — is high-fidelity. Add to firewall, proxy, and DNS sinkhole; hunt retroactively for the full window since Oct 21, 2026.
  • File hashes (SHA256/SHA1/MD5, ~30 total): Trojanized agent-framework wheels (7 SHA256 + 1 SHA1 shared across both JFrog/Socket pulses — strong corroboration), PHP2Shell campaign payloads, CyberLeek tooling, and 8+ MD5s for the malicious-repo setup.js payloads. Deploy SHA256 values to EDR blocklists immediately. Treat MD5/SHA1 values as hunt-only indicators (collision-weakened) — use them in retro-hunts, not sole blocking logic.
  • CVEs (6): Map to exposure management. CVE-2026-109618/109619 are the active exploitation drivers; CVE-2016-5195 signals post-compromise Linux privilege escalation attempts on unpatched legacy hosts.

Operationalization guidance:

  • Enrich hashes via OTX, VirusTotal, and your sandbox of record before blocking to scope affected systems.
  • The PyPI hashes should be matched against package caches, artifact repositories (Artifactory/Nexus/devpi mirrors), CI build images, and developer venvs — not just endpoint file systems.
  • For the GitHub-repo campaign, hash-based detection is secondary; behavioral detection (agents spawning node/setup.js execution, unexpected network egress from AI-assistant child processes) is primary since repos rotate.
  • Sigma-compatible pipelines (pySigma → Sentinel/Splunk/Elastic) will handle the rules below; ensure your .pth and site-packages paths are covered by file-integrity monitoring.

Detection Engineering

YAML
---
title: Python .pth Startup Hook Creation in Site-Packages
description: Detects creation or modification of .pth files in Python site-packages, the persistence/execution mechanism used in the agent-framework PyPI supply-chain compromise (Nov 2026). .pth files execute code on every Python interpreter startup.
status: experimental
author: Security Arsenal Threat Intel
logsource:
  category: file_event
  product: windows
  definition: 'Sysmon EventID 11 or equivalent EDR telemetry; also applicable to Linux/macOS file audit'
detection:
  selection_path:
    TargetFilename|contains:
      - '\site-packages\'
      - '\dist-packages\'
      - '/site-packages/'
      - '/dist-packages/'
  selection_ext:
    TargetFilename|endswith: '.pth'
  filter_known_installers:
    Image|endswith:
      - '\pip.exe'
      - '\pip3.exe'
      - '\python.exe'
  condition: selection_path and selection_ext and not filter_known_installers
falsepositives:
  - Legitimate package installations outside pip (poetry, conda, uv) - tune filters per environment
level: high
tags:
  - attack.persistence
  - attack.t1546
  - attack.t1195.002
  - supply-chain
---
title: AI Coding Assistant Spawning Suspicious Script Execution
description: Detects AI coding assistant processes (Claude, Copilot agents, Cursor, Aider) spawning node/python script execution consistent with the malicious GitHub repository campaign concealing setup.js payloads in AGENTS.md/CLAUDE.md instruction files.
status: experimental
author: Security Arsenal Threat Intel
logsource:
  category: process_creation
  product: windows
detection:
  selection_parent:
    ParentImage|contains:
      - '\claude'
      - '\cursor'
      - '\copilot'
      - '\aider'
      - '\codeium'
      - '\windsurf'
  selection_child:
    Image|endswith:
      - '\node.exe'
      - '\python.exe'
      - '\python3.exe'
      - '\powershell.exe'
      - '\cmd.exe'
  selection_script:
    CommandLine|contains:
      - 'setup.js'
      - 'AGENTS.md'
      - 'CLAUDE.md'
  condition: selection_parent and selection_child and selection_script
falsepositives:
  - Legitimate project scaffolding by AI assistants - investigate command line and network egress of child process
level: high
tags:
  - attack.execution
  - attack.t1059
  - attack.t1195
---
title: Outbound Connection to PHP2Shell Campaign C2
description: Detects network connections to the consolidated C2/payload/exfiltration node 206.245.153.56 used in the AI-agent-orchestrated PHP2Shell credential-harvesting campaign (CVE-2026-109618/109619).
status: experimental
author: Security Arsenal Threat Intel
logsource:
  category: network_connection
detection:
  selection:
    DestinationIp: '206.245.153.56'
  condition: selection
falsepositives:
  - Threat research and sandbox detonation activity
level: critical
tags:
  - attack.command_and_control
  - attack.t1071
  - attack.exfiltration
  - attack.t1041
KQL — Microsoft Sentinel / Defender
// Security Arsenal — agent-framework PyPI compromise + PHP2Shell C2 hunt
// Window: covers activity since PHP2Shell disclosure (Oct 21, 2026) and PyPI hijack (Nov 17, 2026)
let HuntStart = datetime(2026-10-21);
let C2IP = "206.245.153.56";
let MaliciousWheelHashes = dynamic([
  "d27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110",
  "3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9",
  "54d885b7d7161c9ed96489c66364c4bfbf635929166843bba10d97d46b5a2e1f",
  "325e7e9f0ec94e41ad1a1b7a3f18a6b8b5bffece2e2b97b8d3369eb967452f93",
  "39fa329f2a93f734b0bb8f5adb203ea3331d1ae479832a5015a54cb17aabd663",
  "21f28e8517a10ffe431f8d68f5875c85eca4001ae6d118322b0a41873998eb5e",
  "c275ccbe520ef3c0f02dc6a35389e427f9f1b4e0b12e140cd7bf7179ce143f02",
  "d01948e930848dd139915f646760e6dba61caf3eec222d5479518c19b5076ad2"
]);
union isfuzzy=true
(
  DeviceNetworkEvents
  | where TimeGenerated >= HuntStart
  | where RemoteIP == C2IP
  | project Hunt="PHP2Shell C2 Contact", TimeGenerated, DeviceName, RemoteIP, RemotePort, InitiatingProcessFileName, InitiatingProcessCommandLine
),
(
  DeviceFileEvents
  | where TimeGenerated >= HuntStart
  | where SHA256 in~ (MaliciousWheelHashes)
     or (FolderPath has_any ("site-packages", "dist-packages") and FileName endswith ".pth")
     or (FolderPath has "agent_framework" or FolderPath has "agent-framework")
  | project Hunt="Trojanized Package / .pth Hook", TimeGenerated, DeviceName, FolderPath, FileName, SHA256, InitiatingProcessFileName
),
(
  DeviceProcessEvents
  | where TimeGenerated >= datetime(2026-11-17)
  | where (InitiatingProcessFileName has_any ("claude", "cursor", "copilot", "aider", "windsurf")
           and FileName in~ ("node.exe", "node", "python.exe", "python", "python3")
           and ProcessCommandLine has_any ("setup.js", "AGENTS.md", "CLAUDE.md"))
  | project Hunt="AI Agent Malicious Repo Exec", TimeGenerated, DeviceName, InitiatingProcessFileName, FileName, ProcessCommandLine, AccountName
)
| sort by TimeGenerated desc
PowerShell
# Security Arsenal — agent-framework / AI-agent credential-theft hunt script
# Run on developer workstations, build agents, and CI runners. Requires elevated rights for full coverage.

$Report = @()

# 1) Check installed agent-framework versions (compromised: 1.26.1, 1.26.2)
Write-Host "[*] Checking for compromised agent-framework versions..." -ForegroundColor Cyan
$pipList = & python -m pip list --format=json 2>$null | ConvertFrom-Json
if ($pipList) {
    $bad = $pipList | Where-Object { $_.name -match 'agent-framework' -and $_.version -in @('1.26.1','1.26.2') }
    if ($bad) { $Report += [pscustomobject]@{Check='Compromised Package'; Finding="agent-framework $($bad.version) INSTALLED — TREAT HOST AS COMPROMISED"; Severity='CRITICAL'} }
}

# 2) Scan all site-packages for recently modified or suspicious .pth files
Write-Host "[*] Scanning site-packages for suspicious .pth startup hooks..." -ForegroundColor Cyan
$sitePaths = @()
$sitePaths += (& python -c "import site; print('\n'.join(site.getsitepackages()))" 2>$null)
foreach ($sp in $sitePaths) {
    if (Test-Path $sp) {
        Get-ChildItem -Path $sp -Filter '*.pth' -ErrorAction SilentlyContinue | ForEach-Object {
            $content = Get-Content $_.FullName -Raw -ErrorAction SilentlyContinue
            if ($content -match 'import ' -or $_.LastWriteTime -gt (Get-Date '2026-11-17')) {
                $Report += [pscustomobject]@{Check='Suspicious .pth Hook'; Finding="$($_.FullName) — contains import/exec or modified after Nov 17"; Severity='HIGH'}
            }
        }
    }
}

# 3) Check network connections to PHP2Shell C2
Write-Host "[*] Checking for connections to 206.245.153.56..." -ForegroundColor Cyan
$c2Conns = Get-NetTCPConnection -ErrorAction SilentlyContinue | Where-Object { $_.RemoteAddress -eq '206.245.153.56' }
foreach ($c in $c2Conns) {
    $proc = Get-Process -Id $c.OwningProcess -ErrorAction SilentlyContinue
    $Report += [pscustomobject]@{Check='C2 Connection'; Finding="Active connection to PHP2Shell C2 by $($proc.ProcessName) (PID $($c.OwningProcess))"; Severity='CRITICAL'}
}

# 4) Hunt known malicious hashes across likely payload locations
Write-Host "[*] Hash-hunting in temp/cache/venv locations..." -ForegroundColor Cyan
$badHashes = @(
  'd27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110',
  '3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9',
  '54d885b7d7161c9ed96489c66364c4bfbf635929166843bba10d97d46b5a2e1f',
  '325e7e9f0ec94e41ad1a1b7a3f18a6b8b5bffece2e2b97b8d3369eb967452f93',
  '39fa329f2a93f734b0bb8f5adb203ea3331d1ae479832a5015a54cb17aabd663',
  '21f28e8517a10ffe431f8d68f5875c85eca4001ae6d118322b0a41873998eb5e',
  'c275ccbe520ef3c0f02dc6a35389e427f9f1b4e0b12e140cd7bf7179ce143f02',
  'd01948e930848dd139915f646760e6dba61caf3eec222d5479518c19b5076ad2'
)
$searchDirs = @("$env:TEMP", "$env:LOCALAPPDATA\pip\cache", "$env:USERPROFILE\Downloads", "$env:USERPROFILE\.cache")
foreach ($dir in $searchDirs) {
    if (Test-Path $dir) {
        Get-ChildItem -Path $dir -Recurse -Include '*.whl','*.pth','setup.js' -ErrorAction SilentlyContinue | ForEach-Object {
            $h = (Get-FileHash $_.FullName -Algorithm SHA256 -ErrorAction SilentlyContinue).Hash
            if ($badHashes -contains $h.ToLower()) {
                $Report += [pscustomobject]@{Check='Known Malicious Hash'; Finding="$($_.FullName) matches OTX IOC"; Severity='CRITICAL'}
            }
        }
    }
}

# 5) Check for webshell-era artifacts: suspicious recent PHP file writes in web roots (IIS default shown)
$webRoot = 'C:\inetpub\wwwroot'
if (Test-Path $webRoot) {
    Get-ChildItem -Path $webRoot -Recurse -Filter '*.php' -ErrorAction SilentlyContinue |
        Where-Object { $_.LastWriteTime -gt (Get-Date '2026-10-21') } | ForEach-Object {
            $Report += [pscustomobject]@{Check='Recent PHP Write'; Finding="$($_.FullName) modified $($_.LastWriteTime) — verify legitimacy (webshell risk)"; Severity='MEDIUM'}
        }
}

if ($Report.Count -eq 0) { Write-Host "[+] No indicators found." -ForegroundColor Green }
else { $Report | Sort-Object Severity | Format-Table -AutoSize; $Report | Export-Csv -Path ".\arsenal_hunt_$(Get-Date -Format 'yyyyMMdd_HHmm').csv" -NoTypeInformation }

Response Priorities

Immediate (0–4 hours):

  • Block 206.245.153[.]56 at egress firewall, proxy, and DNS; retro-hunt all netflow/proxy logs back to October 21, 2026.
  • Push all SHA256 indicators to EDR blocklists; search artifact repositories and pip caches for the trojanized agent-framework wheels.
  • Inventory every system with agent-framework 1.26.1 or 1.26.2 installed — developer workstations, CI runners, container images, notebook environments — and isolate pending triage.
  • Emergency-patch PHP2Shell instances (CVE-2026-109618/109619) or take exposed instances offline; audit web roots for webshells written since Oct 21.

Within 24 hours:

  • Credential rotation at blast radius scale: any host with a compromised package version or confirmed C2 contact must have ALL credentials rotated — Azure service principals and tokens, GitHub PATs and Actions secrets, CI/CD variables, SSH keys, cloud IAM session material. Assume the .pth hook exfiltrated everything in the environment block every time Python started.
  • Revoke and re-issue GitHub Actions secrets org-wide if any runner is implicated; audit Actions logs for anomalous workflow runs and unexpected outbound calls during builds.
  • Identity-layer verification: hunt Azure AD/Entra sign-in logs for token replay from anomalous ASNs and impossible-travel patterns tied to affected service principals.
  • Freeze AI coding assistant usage on unvetted repositories; enumerate which repos developers' agents have ingested since Nov 17 and check for AGENTS.md/CLAUDE.md-driven setup.js execution.

Within 1 week:

  • Pipeline hardening: enforce hash-pinned, allow-listed package installs via an internal PyPI proxy (no direct pypi.org pulls from build agents); require attestation/provenance (Sigstore) for critical dependencies.
  • Deploy file-integrity monitoring over site-packages/dist-packages across the fleet; alert on any .pth write.
  • Sandbox or restrict AI coding assistants: deny autonomous execution of repo-discovered scripts, gate agent file execution behind human approval, and EDR-monitor agent child processes.
  • Accelerate edge patching SLA to 24–48 hours for internet-facing PHP/application servers — the AI-orchestrated exploitation window in this campaign was measured in days, not weeks.
  • Brief engineering leadership: the developer toolchain is now a Tier-1 attack surface; fold supply-chain and agentic-AI risk into the threat model and procurement review.

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.