On September 14, 2026, Ant Group's AI Security Lab proudly announced the release and open-sourcing of SingProbe, an innovative, built-in safety guardrail specifically designed for large models. Leveraging the inference hidden states of the foundational model, this technology performs real-time risk assessments at the token level, incurring less than a 0.5% increase in computational overhead. SingProbe is compatible with 29 leading open-source large models, such as the Ling-3.0 series, GLM-5.2/5.3, Qwen, and DeepSeekV4. Additionally, it seamlessly integrates with inference frameworks like SGLang and vLLM. During the content generation process, SingProbe employs lightweight probes to output real-time risk signals on a token-by-token basis, enabling the identification of potential risks as they arise and effectively curbing the dissemination of non-compliant information. To facilitate global development, the relevant code, models, and evaluation benchmarks have been made openly available for developers worldwide to reference and utilize.
