Ant Releases SingProbe: An Embedded Safety Guardrail for Large Language Models

During China Cybersecurity Week, Ant AI Safety Lab released SingProbe, an endogenous safety guardrail for large language models that integrates security detection into the generation process. The module reuses internal information produced during model inference and outputs risk scores as the model generates, allowing applications to interrupt, regenerate, or issue alerts when risky content is detected.

Many LLM applications currently rely on external guardrail models to review inputs and outputs, which can increase computing and deployment costs. SingProbe aims to run safety assessment in parallel with generation. In production tests on the Ling-3.0-flash model, the extra decoding overhead was below 0.5%. The flash version outperformed selected public baselines on answer safety classification and streaming safety detection, while remaining roughly on par with a reference baseline on hallucination detection.

The team also released SingStreamBench, a benchmark for streaming generation scenarios that focuses on the moment a model shifts from normal to risky content, checking whether guardrails trigger too early and whether detection is timely. In the medical domain, the team introduced SingProbe-Med, which monitors medical risks during generation and selectively applies decoding intervention to high-risk local content. On the AntAngelMed-100B medical evaluation, full intervention corrected 25.03% of originally incorrect answers from the baseline model.

SingProbe currently supports 29 mainstream open-source models, including Ling-3.0 series, GLM-5.2/5.3, Qwen, and DeepSeekV4, and is integrated with SGLang and vLLM inference frameworks. The code, models, and benchmark have been open-sourced.