요새를 넘어오는 이미지 공격에 대비하기: FENCE | 신뢰와 안전
kakao tech · 22:12 · Watch on YouTube
Kakao Bank's existing text-only guardrail LabFortress can't catch attacks hidden in images, so the team built FENCE — a 10,000-sample Korean/English finance-specific VLM prompt-attack dataset — and trained a 3B guardr...