AprielGuard: 8B safety-security guard for modern LLM agent ecosystems

AI Impact Summary

AprielGuard introduces an 8B safety-security safeguard model designed to detect 16 safety categories and a broad set of adversarial attacks within modern LLM agent ecosystems. It supports standalone prompts, multi-turn conversations, and agentic workflows, outputting safety categories and optional structured reasoning to explain decisions. By integrating with the Apriel-1.5 Thinker Base and toolchain components (Mixtral-8x7B, NVIDIA NeMo Curator, SyGra), it addresses long-horizon reasoning and memory/tool manipulation threats, enabling production-grade gating and explainability in real-time pipelines.