Researchers investigating Hugging Face found the majority of its models would generate non-consensual deepfakes.