{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/7c98af563d82479cb6b84ef6cc8975c3\" frameborder=\"0\" width=\"1114\" height=\"835\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":835,"width":1114,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":835,"thumbnail_width":1114,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/7c98af563d82479cb6b84ef6cc8975c3-9ee09a0661feddac.gif","duration":300.133,"title":"Blind Auditing Vietnamese AI Safety Gaps","description":"This Loom presents ApplyFault, a method for blindly auditing Vietnamese AI safety blind spots using secret enclaves. It explains that LLM safety is carefully evaluated mainly for English and not for Vietnamese or Global South languages, leaving issues like scams and medical misinformation under-tested, and describes a trust deadlock where model weights and evaluation benchmarks are not shared. The speaker then demonstrates Plyfold, a Zero Trust LLM evaluation on a Vietnamese harm dataset using secure enclave so parties do not need to share private data. They report experiments on open models including GPT from Vinai Research in 2023, Qwen 0.5B and Qwen 3B (CSEA-focused Lambda), finding that the most Vietnamese fluent model can be least safe in certain directions."}