AprielGuard

2025-12-23

AI Summary

Existing LLM safety frameworks typically model content harms (e.g., toxicity, bias) and adversarial threats (e.g., prompt injection, jailbreaking) in isolation, limiting holistic risk mitigation in dialogue and agentic settings. Method: We propose the first unified 8B-parameter safety guardian, introducing a fused classification taxonomy and a multi-stage supervised fine-tuning framework to enable end-to-end joint detection of both risk categories. We incorporate structured reasoning traces for enhanced interpretability and design an input encoding mechanism adaptable to single-turn/multi-turn dialogues and agentic workflows. Contribution/Results: Trained on a hybrid dataset of open-source and synthetically generated examples, our model significantly outperforms baselines—including Llama-Guard and Granite Guardian—across multiple public and internal benchmarks. Notably, it achieves marked improvements in detection accuracy for multi-step adversarial scenarios, demonstrating superior robustness and generalization in complex, realistic deployment settings.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Multiagent Systems: Adversarial Agents

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Safeguarding large language models (LLMs) against unsafe or adversarial behavior is critical as they are increasingly deployed in conversational and agentic settings. Existing moderation tools often treat safety risks (e.g. toxicity, bias) and adversarial threats (e.g. prompt injections, jailbreaks) as separate problems, limiting their robustness and generalizability. We introduce AprielGuard, an 8B parameter safeguard model that unify these dimensions within a single taxonomy and learning framework. AprielGuard is trained on a diverse mix of open and synthetic data covering standalone prompts, multi-turn conversations, and agentic workflows, augmented with structured reasoning traces to improve interpretability. Across multiple public and proprietary benchmarks, AprielGuard achieves strong performance in detecting harmful content and adversarial manipulations, outperforming existing opensource guardrails such as Llama-Guard and Granite Guardian, particularly in multi-step and reasoning intensive scenarios. By releasing the model, we aim to advance transparent and reproducible research on reliable safeguards for LLMs.

Problem

• Unifies safety risks and adversarial threats in a single framework
• Detects harmful content and adversarial manipulations in LLMs
• Improves robustness in multi-turn conversations and agentic workflows

Innovation

• Unified safety taxonomy for diverse threats
• Training with structured reasoning traces
• Strong performance in multi-step scenarios