| 本文作者： Nemo | 2026-09-14 21:22 |
大模型正在进入日常问答、办公辅助、客户服务等真实业务情形。使用者期待AI快速给出有用的回答，也需要这些回答安全、可靠。如何及时发现不安全内容和编造信息，同时减少安全审核对响应速度和使用者感受的影响，成为大模型应用落地需要应对的难题。
国家网络安全宣传周期间，蚂蚁AI安全实验室正式发布大模型内生式安全护栏SingProbe，将安全检测融入模型生成过程。它如同一个内置的“安全探头”，让AI一边生成回答，一边输出风险提示，以较低的额外计算开销，为大模型应用给出更及时的安全防护。
例如，使用者向AI咨询健康难题时，会关心它是否给出不恰当的用药建议；使用AI查找资料时，也需要警惕它编造不存在的文献或事实依据。面对这些风险，SingProbe可以在回答生成过程中持续给出风险信号，协助应用平台及时采取中止回答、重新生成等措施。
当下，大模型应用通常借助独立的外置护栏模型审核输入或输出。这种方式通用、直接，但需要额外处理待审核内容，增加计算和部署成本。如果等完整回答生成后再检查，风险内容可能已经展现给使用者；如果提高检查频率，又会进一步增加运行负担。
SingProbe尝试让安全判断与模型生成同步开展。它复用大模型推理过程中已经产生的内部信息，借助轻量化的安全检测模块，持续输出风险分数。这意味着，模型在生成回答的同时，就能给出当前内容是否存在安全或事实可靠性风险的信号，供平台及时采取告警、终止生成、重试等措施。
据研发团队在Ling-3.0-flash模型生产环境中的实测，SingProbe在解码阶段引入的额外开销低于0.5%。团队实测表明，flash版本在回答安全分类和流式安全检测任务上超过所选公开基线，在幻觉检测任务上与参考基线基本持平，为兼顾检测能力与运行效率给出了新的工艺路径。
针对流式生成情形，此次还同步发布了实测基准SingStreamBench。该基准关注模型从正常回答转向风险内容的具体时刻，不仅检验护栏能否发现风险，还考察是否过早触发、发现是否及时，以更贴近实际应用的方式衡量安全防护成效。
在医疗生成情形中，研发团队进一步探索了从风险识别到按需纠偏的工艺应用，提出SingProbe-Med：持续监测生成过程中的医疗风险，仅在达到触发条件后，对局部高风险内容开展解码干预。据团队在AntAngelMed-100B上的医疗实测，完整干预可以纠正基线模型中25.03%原本出错的回答，展示了内生风险信号用于生成过程安全控制的潜力。
当下，SingProbe已适配Ling-3.0系列、GLM-5.2/5.3、Qwen、DeepSeekV4等29个主流开源大模型，并接入SGLang、vLLM推理框架，相关代码、模型和实测基准同步开源。后续，团队将逐步扩展对更多开源模型的兼容，推动安全防护更紧密地融入大模型推理与部署流程。
雷峰网(公众号：雷峰网) 雷峰网
-
蚂蚁百宝箱推出全新商圈智能体模版，万达率先发布逛 ... -
蚂蚁数科在外滩大会达成31项AI合作，推动智能体走进 ... -
大模型牌桌上，蚂蚁如何出牌？ -
从“GPASS”到“灵影”，蚂蚁为AI终端铺设Agent基础 ...
编辑
本文来源：https://www.leiphone.com/category/industrynews/pvwJSKM5MEEd1H5N.html
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Author: Nemo 2026-09-14 21:22
Large language models are increasingly integrated into real-world scenarios such as daily Q&A, office assistance, and customer service. Users expect quick, useful responses, but equally value safety and reliability. Detecting unsafe content and hallucinations promptly while minimizing the impact on response speed and user experience is a key challenge for deploying large models.
During National Cybersecurity Week, Ant AI Safety Lab officially released SingProbe, a built-in safety guardrail for large models. It embeds safety detection into the generation process, acting as a real-time "safety probe" that outputs risk signals alongside the response. This approach provides timely protection with minimal additional computational overhead.
For instance, when users consult AI about health issues, they need to know if the advice is medically sound. When searching for literature, they need to verify the existence of cited references. SingProbe continuously monitors these risks during generation, enabling the platform to intervene if necessary — by halting the response or triggering a retry — before the content reaches the user.
Current practice often relies on external, standalone safety models to review inputs or outputs. While effective, this adds latency and increases deployment costs. If checked after full generation, risks may already be exposed; more frequent checks slow down the service.
SingProbe addresses this by performing safety assessment in parallel with model generation. It leverages intermediate information from the inference process and uses lightweight modules to compute risk scores in real-time. Thus, potential issues are flagged as they arise, supporting proactive intervention.
Initial experiments on Ling-3.0-flash show that SingProbe adds less than 0.5% overhead to the decoding phase. In benchmarking, it outperforms baseline models in safety classification and streaming detection tasks, while maintaining comparable performance in hallucination detection — proving it can be both effective and efficient.
To better capture real-world dynamics, the team also introduced SingStreamBench, a benchmark for streaming generation scenarios. This evaluates not only whether risks are detected, but also how quickly and appropriately they are flagged, ensuring the guardrail behaves as intended under operational conditions.
In medical applications, the concept extends further: SingProbe-Med identifies risks during text generation and can trigger localized corrections if thresholds are exceeded. On AntAngelMed-100B, this intervention successfully corrected 25.03% of erroneous outputs, demonstrating the potential for context-aware self-correction.
SingProbe currently supports 29 major开源大模型 (open-source models), including Ling-3.0, GLM-5.2/5.3, Qwen, and DeepSeekV4, and integrates with SGLang and vLLM frameworks. The code, models, and benchmarks are open-sourced, with plans to extend compatibility to more models and embed safety deeper into the deployment pipeline.
Lei Feng Network
0 collections
Related articles
Ant Toolbox launches new commercial district intelligent body template, Wanda is the first to release shopping guide
Ant Group Digital Technology reaches 31 AI cooperation agreements at the Bund Conference, promoting intelligent body
How does Ant Group play its cards at the large model table?
From "GPASS" to "Ling Ying", Ant Group lays the foundation for AI terminal agent
Nemo
Edit
Send private message
Popular searches
DiDi Chinese Academy of Sciences OpenAI Jia Yueting Appearance Haier Innovation Mobile advertising TikTok Dahua Hikvision