Pickle Panic: Malicious ML Models Sneak Past Hugging Face’s Security with NullifAI Trickery
Cybersecurity researchers have found two malicious machine learning models on Hugging Face using “broken” pickle files to dodge detection. Dubbed nullifAI, this sneaky approach exploits a gap in Picklescan’s defenses, allowing malicious payloads to execute while causing decompilation errors. Hugging Face has since updated their tools to address this issue.
Hot Take:
Who knew that adding a touch of “pickling” to machine learning models could make them so, well, rotten? Hackers have taken up culinary arts with a side of tech-savvy mischief, and it seems like Hugging Face is in for a sour surprise. Just when you thought your AI was safe, it turns out it might be harboring a secret: a love for pickles. But not the tasty kind.
Key Points:
- Cybersecurity researchers found two malicious ML models on Hugging Face using “broken” pickle files.
- The method has been dubbed “nullifAI” for its clever evasion of existing safeguards.
- The models act as proof-of-concept rather than an active supply chain attack.
- Pickle serialization is a known security risk, allowing arbitrary code execution.
- Hugging Face’s security tool Picklescan failed to detect these sneaky models.
Already a member? Log in here