Ant Group launches SingProbe, an in-generation LLM safety guardrail with under 0.5% overhead
The AMW Read
SingProbe embeds real-time safety scoring into decoding across 29 open models plus a new streaming-safety benchmark, a meaningful shift from bolt-on guardrails without introducing a new top-tier player.
Ant Group launches SingProbe, an in-generation LLM safety guardrail with under 0.5% overhead
Ant Group's AI Security Lab released SingProbe, a safety guardrail built directly into the LLM generation process instead of running as a separate pre- or post-processing check. The system reuses internal signals the model already produces during inference and emits a continuous risk score while text is being generated, adding under 0.5% overhead at the decoding stage as measured in production on Ling-3.0-flash. Ant also released a companion benchmark, SingStreamBench, built to test how quickly and accurately a guardrail catches the exact moment a response turns risky mid-generation, and open-sourced the code, models, and benchmark. SingProbe is already adapted to 29 open-source models, including Ant's own Ling-3.0 line, Zhipu's GLM-5.2/5.3, Qwen, and DeepSeek V4, and plugs into the vLLM and SGLang serving frameworks.
Most production guardrails today are bolt-on classifiers that review a prompt or a finished response, which either delays detection until risky content has already reached the user or adds cost from re-running output through a second model. By folding detection into the decoding pass itself, Ant is pitching inline safety as a property of the serving stack rather than an external audit step. The release also lands alongside Ant's own build-out of the Ling-3.0 family, following last month's finance-tuned Ling-3.0-flash-Fin, suggesting Ant wants safety instrumentation baked across its model line rather than treated as a one-off compliance add-on.
For teams serving open-weight models in regulated settings, the open release lowers the cost of inline moderation without standing up a separate reviewer model — Ant also tested a medical variant, SingProbe-Med, that intervened only on high-risk content and corrected just over a quarter of originally wrong answers on an internal medical benchmark, though both figures still need validation outside Ant's own test harness.
