SingGuard-NSFA-9B-GGUF

SingGuard-NSFA-9B-GGUF

SingGuard-NSFA-9B is the largest of four model sizes (0.8B, 2B, 4B, 9B) in a dual-mode guardrail framework developed by the SingGuard Team at Ant Group's AI Security Lab, fine-tuned from Qwen3.5-9B Base to secure agentic AI systems against operational threats like prompt injection, sensitive information extraction, malicious code requests, dangerous tool misuse, and resource exhaustion. It is built on the NSFA (Not-Secure-For-Agents) taxonomy — a CIA-triad-grounded hierarchical classification of 185 risk variants spanning 5 query-side domains (e.g., Prompt Injection & Jailbreak, Malicious Code & Cyberattack) and 2 response-side domains (Hazardous Action Generation, Sensitive Information Leakage) — and operates as a single-turn, text-based guardrail supporting 133 languages through two complementary inference modes: lightweight discriminative classification heads on a frozen backbone for real-time detection (~50ms per sample on an A100 via vLLM embedding mode), and full generative chain-of-thought reasoning for interpretable offline auditing and compliance workflows. Across three purpose-built multilingual benchmarks (including a cross-source benchmark adapted from AgentDojo, InjecAgent, AgentHarm, and others), all SingGuard-NSFA models achieve above 94% F1, surpassing the strongest competing guardrails by 6–12 absolute F1 points, with the architecture also natively extensible to new risk domains — including content safety — by training only additional lightweight classification heads without retraining the backbone; it is explicitly scoped to single-turn operational security rather than multi-turn trajectory analysis, multimodal threats, or textual content moderation, and is intended strictly as a defensive tool.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp