What threats does SingGuard-NSFA detect?

The model identifies prompt injections, sensitive data theft, malicious code execution, resource abuse, and unauthorized access. It covers 28 subcategories of risks across seven primary threat vectors.

Which languages does SingGuard-NSFA support?

The model supports 133 languages, enabling global deployment. It includes nearly 100,000 test cases for performance validation.

What versions of SingGuard-NSFA are available to developers?

Four variants are offered: 0.8B, 2B, 4B, and 9B parameters. Each processes requests in approximately 50 milliseconds, ensuring high performance.

← All news

Artificial intelligence

Ant Group Unveils Open-Source AI Safety Models for Agents and Multimodal Systems

July 13, 2026

Photo: TechNode

Quick answer

Ant Group has open-sourced the SingGuard-NSFA model to safeguard AI agents from cyber threats, including prompt injections and data leaks. The model covers 185 scenarios in 133 languages with a latency of just 50ms.

Chinese tech giant Ant Group, renowned for its fintech innovations, has taken a significant step in advancing AI security by open-sourcing the SingGuard-NSFA model. This model is specifically designed to protect autonomous AI agents from a wide array of cyber threats.

SingGuard-NSFA identifies seven major risk categories, encompassing 28 subcategories and 185 distinct scenarios. Key threats addressed include prompt injections, sensitive data theft, malicious code execution, and abuse of computational resources and access rights. The model supports 133 languages and features an evaluation system with nearly 100,000 test cases.

Developers can access four versions of the model, varying in parameter size: 0.8B, 2B, 4B, and 9B. According to Ant Group, the model processes single requests in about 50 milliseconds, making it suitable for high-performance systems. The second model, SingGuard, is tailored for multimodal AI systems, providing robust protection in complex scenarios.

Open-sourcing SingGuard-NSFA is expected to accelerate the adoption of secure AI solutions across industries, including financial services, cloud platforms, and enterprise applications. The company emphasizes that its developments align with modern cybersecurity standards and are suitable for both commercial and research initiatives.

Common questions

What threats does SingGuard-NSFA detect?: The model identifies prompt injections, sensitive data theft, malicious code execution, resource abuse, and unauthorized access. It covers 28 subcategories of risks across seven primary threat vectors.
Which languages does SingGuard-NSFA support?: The model supports 133 languages, enabling global deployment. It includes nearly 100,000 test cases for performance validation.
What versions of SingGuard-NSFA are available to developers?: Four variants are offered: 0.8B, 2B, 4B, and 9B parameters. Each processes requests in approximately 50 milliseconds, ensuring high performance.

Dzen feed: /feed/dzen.xml · RSS: /feed.xml