Hugging Face platform continues to be plagued by vulnerable ‘pickles’

ReversingLabs identified two machine-learning models hosted on Hugging Face that used malicious Python pickle files to deploy web shells pointing to a hardcoded IP, evading the platform's Picklescan scanner by using alternative compression and broken-pickle techniques; the packages appeared to be proof-of-concept rather than widespread active exploitation, were removed after disclosure on Jan 20, and Picklescan received updates to better detect malicious or corrupted pickle payloads.