AprielGuard

TL;DR

AprielGuard presents a unified safeguard model that simultaneously detects safety risks and adversarial manipulations across standalone prompts, conversations, and agentic AI workflows. By leveraging a diverse synthetic and open data mix with structured reasoning traces, it achieves strong performance on multi-step, reasoning-heavy tasks while maintaining interpretability. The paper contributes a unified taxonomy, a robust data-generation pipeline with rigorous filtering and augmentation, and a two-mode training design that supports reasoning-enabled and reasoning-free deployments. Its evaluation shows robust, multilingual, and long-context resilience, with practical implications for deploying safer LLMs in real-world agentic settings.

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.

Figures (4)

• Figure 1: AprielGuard combines open and synthetic data to cover safety and adversarial risks. Unsafe content is generated with models such as Mixtral-8x7B and Llama-8B, filtered for duplicates (semantic and syntactic similarity), and augmented with typographical and paraphrasing noise. The resulting dataset balances safe/unsafe and adversarial/non-adversarial samples for supervised fine-tuning.
• Figure 2: AprielGuard Prompt Template and Output. Each moderation request embeds the content within a <|content|> tag, accompanied by an instruction provided under the user role. The model produces outputs in a predefined structured format. The special instruction varies depending on the content type—conversational or standalone prompt. For conversational content, the model evaluates safety risks in the assistant’s response and detects adversarial behavior in the user’s message. In contrast, for standalone prompts, both safety risks and adversarial threats are assessed for the entire content.
• Figure 3: Multilingual performance comparison of AprielGuard models relative to English. The figure presents F1-scores across nine languages for AprielGuard on Safety Risk and Adversarial Attack benchmarks, with and without reasoning. AprielGuard achieves consistently high performance across most languages, except Japanese where a slight drop in performance is observed.
• Figure :