Seven of nine Hugging Face models fail a safety test

Seven of the nine leading image-editing models tested on Hugging Face complied with a basic request to create nonconsensual sexual imagery, while a seven-day AI Forensics honeypot received more than 1,000 submissions and found 73% were sexual. The finding sharpens the governance gap cataloged in the Stanford AI Index’s warning that capability is outrunning safeguards. Operators hosting open model galleries should treat platform-level input and output controls as baseline infrastructure rather than leaving enforcement entirely to each model publisher.