Archive — CN — technode
Ant Group unveils AI safety models for agents and multimodal systems
Ant Group's AI Safety Lab has launched SingGuard-NSFA, an open-source safety model for autonomous agents, which identifies risks like data theft and malicious code execution. Covering 185 scenarios in 133 languages, the model offers various versions and assesses risks in approximately 50 milliseconds.
Key points
- Ant Group opened SingGuard-NSFA, a safety model for detecting risks in AI agents.
- The model addresses risks such as prompt injection, data theft, and permission misuse.
- It categorizes risks into seven main categories and 28 subcategories.
- Includes 185 scenarios and supports assessments in 133 languages.
- Available in versions of different sizes (0.8B, 2B, 4B, 9B).
- Single risk judgments can be made in around 50 milliseconds.
Why it matters
- Related to the 2025 IT trend of increasing emphasis on AI safety and governance.
- Reflects growing concerns over AI-related security risks, including data integrity and security breaches.
- Aligns with global regulatory movements pushing for responsible AI deployment and safety standards.
Ant Group’s launch of AI safety models marks a significant step in risk management for autonomous systems, aligning with broader trends toward responsible AI development and governance.