Ant Group's AI Safety Lab has open-sourced SingGuard-NSFA, a safety guardrail model for autonomous agents that detects risks including prompt injection and sensitive data theft before agents act. The model covers seven risk categories, 185 scenarios, and 133 languages, with versions ranging from 0.8B to 9B parameters. Ant Group also disclosed details of SingGuard, a multimodal safety model. SingGuard-NSFA can make a single risk judgment in about 50 milliseconds.
No score is assigned. Sources and their independence are shown in the citation chain below.