Pickle Predicament: Hugging Face Faces Malware Mayhem with Malicious ML Models
Malicious machine learning models on Hugging Face have found a sneaky way to dodge security scans by cleverly exploiting Pickle file serialization. This “malicious Pickling” has exposed gaps in Hugging Face’s Picklescan tool, raising eyebrows and highlighting the need for tighter security measures on open ML platforms.
Hot Take:
Looks like Hugging Face has been caught with its digital pants down! It seems even AI models need to watch out for creepy Pickles lurking in the shadows. Time to step up the game, Hugging Face, and make sure your Pickle jar isn’t full of worms!
Key Points:
- Researchers discovered two malicious ML models on Hugging Face, exploiting Pickle files.
- Pickling in Python is risky as it allows code execution during deserialization.
- Malicious models used a novel technique bypassing Hugging Face’s Picklescan detection.
- Reversing Labs criticized Hugging Face’s reliance on basic security features.
- Hugging Face removed the malicious models and updated its detection tools.
Already a member? Log in here