SingGuard-NSFA Offers Extensible Guardrails for Agentic AI Systems.

This paper introduces SingGuard-NSFA, a novel guardrail framework engineered to enhance the security of agentic AI systems. It aims to protect against a range of operational threats, including prompt injection, the extraction of sensitive information, malicious code requests, dangerous tool misuse, and resource exhaustion. The framework is built upon the NSFA taxonomy, a comprehensive hierarchy of 185 risk variants grounded in the CIA triad and cross-validated against OWASP guidelines. To evaluate its effectiveness, a benchmark suite was constructed, featuring over 93,000 purpose-built samples across 133 languages, targeting both user queries and agent responses, alongside 3,435 samples adapted from existing public agent-security datasets. SingGuard-NSFA employs a dual-mode detection strategy: SFT-based generative reasoning for interpretable offline auditing and discriminative classification heads on a frozen backbone for real-time threat detection (approximately 50ms). The released models, ranging from 0.8B to 9B parameters, consistently achieve high F1 scores (>=94%) on purpose-built benchmarks, significantly outperforming competing guardrails. The extensibility of the approach is demonstrated by classification heads detecting risks beyond their original scope, establishing state-of-the-art performance.