Fake Hugging Face Repository Used to Deliver Sefirah Infostealer

· Original article ↗

Summary

HiddenLayer found a typosquatted Hugging Face repository impersonating OpenAI’s Privacy Filter. Its loader deployed the Sefirah infostealer against Windows users; the repo reached 244,000 downloads, a count researchers said may have been inflated.

Key points

  • The Open-OSS/privacy-filter repository copied OpenAI’s Privacy Filter project and briefly reached #1 on Hugging Face before the platform removed it.
  • Its loader disabled SSL verification, fetched a JSON payload, and ran a PowerShell command that downloaded a batch file and the Sefirah payload.
  • Sefirah targeted browser data, Discord tokens, cryptocurrency wallets, credentials, sensitive files, system information, and screenshots.
  • The malware compressed and exfiltrated stolen data to recargapopular[.]com.
  • HiddenLayer reported anti-analysis checks for virtual machines, sandboxes, debuggers, and analysis tools.
  • Researchers linked the loader infrastructure to other malicious repositories and noted overlaps with an npm typosquatting campaign distributing WinOS 4.0.
  • The number of victims is unknown; researchers said the 244,000 download count may have been artificially inflated. They advised affected users to reimage machines and rotate credentials, wallets, and sessions.

Article Details

Attack Vectors
  • A typosquatted Hugging Face repository impersonated OpenAI’s Privacy Filter project and used a deceptive loader.py script to deliver malware to Windows users.
  • The loader disabled SSL verification, decoded a URL to an external resource, and fetched a JSON payload containing a PowerShell command. That command downloaded start.bat, which downloaded and executed the Sefirah payload.
  • Sefirah collected browser data, Discord tokens, cryptocurrency wallet data, credentials, sensitive local files, system information, and multi-monitor screenshots, then exfiltrated compressed data to a C2 server.
Defensive Notes
  • Users who downloaded files from the malicious repository were advised to reimage the affected machine.
  • Rotate stored credentials, replace cryptocurrency wallets and seed phrases, and invalidate browser sessions and tokens.

Indicators of compromise

TypeIndicatorContext
DOMAINrecargapopular[.]comC2 server to which Sefirah exfiltrated compressed stolen data.

MITRE ATT&CK

T1005 · Data from Local SystemSefirah collected local browser data, Discord data, wallet data, credentials, and other sensitive files.T1036.005 · Match Legitimate Resource Name or LocationThe malicious Open-OSS/privacy-filter repository typosquatted OpenAI’s legitimate Privacy Filter project.T1041 · Exfiltration Over C2 ChannelSefirah exfiltrated compressed stolen data to the reported C2 server.T1059.001 · PowerShellThe loader fetched a JSON payload containing a PowerShell command that downloaded a batch file.T1059.003 · Windows Command ShellThe downloaded start.bat batch file downloaded and executed the final payload.T1059.006 · PythonThe repository’s loader.py script executed Python code to retrieve and run the next stage.T1082 · System Information DiscoverySefirah collected system information from infected machines.T1105 · Ingress Tool TransferThe loader and subsequent batch file downloaded payloads from external resources.T1113 · Screen CaptureSefirah captured multi-monitor screenshots.T1497.001 · System ChecksSefirah checked for virtual machines and sandboxes to evade analysis.T1539 · Steal Web Session CookieSefirah targeted browser cookies and session tokens.T1552.001 · Credentials In FilesSefirah collected SSH, FTP, and VPN credentials and configuration files, as well as sensitive local files.T1555.003 · Credentials from Web BrowsersSefirah targeted saved passwords and other data from Chromium- and Gecko-based browsers.T1562.001 · Disable or Modify ToolsThe batch file added Sefirah to Microsoft Defender’s exclusions.T1622 · Debugger EvasionSefirah checked for debuggers as part of its anti-analysis behavior.

Malware

Vendors

Products

Related Articles