SGuard-v1: Safety Guardrail for Large Language Models

Two-layer defense system: SGuard-v1 uses two specialized components—ContentFilter to identify harmful content in prompts and responses, and JailbreakFilter to detect adversarial attacks—providing multiple checkpoints to catch safety issues rather than relying on a single detection method.

Comprehensive threat coverage: The system was trained to recognize 60 major types of jailbreak attacks (techniques used to bypass AI safety measures) while being designed to reduce false alarms that incorrectly flag safe content as harmful, addressing a key challenge in responsible AI deployment.

Lightweight and practical: Built on a 2-billion parameter model rather than larger systems, SGuard-v1 reduces the computational resources needed to deploy safety guardrails, making responsible AI safety more accessible and cost-effective for organizations of different sizes.

Multilingual safety across 12 languages: The system supports safety monitoring in 12 languages, addressing the global nature of AI deployment and ensuring that safety protections aren't limited to English-speaking contexts.

Transparency through explainability: SGuard-v1 provides detailed safety classifications and confidence scores rather than simple yes/no decisions, helping developers and users understand why content was flagged and make more informed decisions about AI system outputs.