Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentMalicious model files

Fake OpenAI Privacy Filter repo tops Hugging Face trending and installs an info stealer

HighHiddenLayerHugging Face

HiddenLayer found a Hugging Face repository, Open-OSS/privacy-filter, posing as OpenAI's Privacy Filter model, which detects and redacts personal data in text. It copied the real model card almost word for word and reached number one on the trending list with about 244,000 downloads and 667 likes in 18 hours, most of it from bot accounts. Its setup scripts installed an information stealer on Windows. Hugging Face removed the repository.

Replay the attack step by step in the Learning Center

How it happened

  1. The model card told users to clone the repo and run start.bat on Windows or loader.py on Linux and macOS.
  2. loader.py turned off certificate checks, decoded a hidden URL and fetched PowerShell commands from a paste site, then ran them in a hidden window; this stage only worked on Windows.
  3. A second script added Microsoft Defender exclusions and a scheduled task running as SYSTEM, then dropped a Rust information stealer.
  4. The stealer took browser passwords and cookies, Discord tokens, crypto wallet data, FileZilla and SSH details and screenshots, and sent them to the attacker's server.

Impact

Anyone who ran the setup scripts on Windows lost saved passwords, session cookies, tokens and wallet data; six more repositories with similar loaders were found under another account.

The fix

Hugging Face removed the repository. HiddenLayer advises treating any machine that ran the scripts as fully compromised: reimage it, rotate every stored password, token, SSH key and cloud credential, move crypto to new wallets and block the published indicators.

Lessons

  • Trending rank, downloads and likes can be faked; check that the publisher account is the real organization's.
  • A model that needs you to run a .bat or loader script to start is a program, not a model; do not run it.
  • Pull models through an internal registry that only holds publisher-verified, scanned copies.

Sources