In April, the in-house DeepGuard LLM completed its major annual upgrade. Retrained on tens of millions of operations logs, security events, and ticket cases, threat-triage accuracy rose from 89% to 95% with false positives falling further.
The new version introduces attack-chain reasoning: the model simulates a penetration tester’s chain of thought, reconstructing the full attack path from a single alert and generating defense-policy recommendations automatically. It also accepts voice commands and multimodal input such as topology diagrams and log screenshots, with voice recognition accuracy of at least 98% in noisy server rooms.
The model upgrades, customers feel nothing — every existing customer gets the full improvement with zero action required. That is the value of an AI-native architecture.
The DeepGuard upgrade has already been rolled out across the XianJue line and ZhiKeXing.