AprielGuard: 8B safety-security guardrail for modern LLM agent workflows

AI Impact Summary

AprielGuard provides an 8B parameter guardrail model that detects 16 safety categories and flags adversarial attack patterns across standalone prompts, multi-turn conversations, and agentic workflows (tool calls, memory, and reasoning traces). Trained on synthetic data generated with Mixtral-8x7B, NVIDIA NeMo Curator, and SyGra, it targets long-context, multi-step interactions typical of modern LLM agent ecosystems. It outputs safety classifications and violated categories, with optional structured reasoning for explainability, enabling a unified safety layer in production pipelines. This reduces exposure to jailbreaks, prompt injections, and memory manipulation, but teams should plan for integration with existing tool orchestration and potential latency impacts.