Ant Group Security simultaneously open-sourced two AI safety frameworks, SingGuard-NSFA and SingGuard: the former targets Agents, providing 0.8B-9B four-tier dual-mode inference guardrails, building NSFA risk classification based on CIA + OWASP, with the 9B model reaching a generalized F1 SOTA of 91.29%, and new risks only require training additional classifier heads; the latter targets multimodal large models, treating safety rules as runtime input, supporting fast/slow thinking switching and RI-Mask parallel 5× speedup.