最新资讯
reddit.com OpenAI 称 Astra 仅满足 10% 安全任务标准 dev.to FreeToken:在自有 GPU 上运行前沿 MoE 模型及其内存需求 dev.to 当 AI Agent 说“完成”时,究竟完成了什么? dev.to 部署 Underdog Bark 3.0 Pro:逐字转写、自定义后端与分块设置 dev.to 为艰难对话构建 AI:五个 Agent、一条安全正则与一位盲评监督者 techcrunch.com Palo Alto Networks 洽谈以 5 亿美元收购 AI Agent 安全初创公司 Clutch theinformation.com 黄仁勋被要求就涉华间谍指控向国会作简报 www.infoq.com Edera 推出 Object Capability Bounded Authority,试图将凭证与 AI Agent 工作负载隔离 wechat 这波杀疯了!旗舰级模型不限量、1个月视频额度免费领 bloomberg.com AI 抛售蔓延至信用市场,软银领跌亚洲美元债 dev.to 如何在专用 GPU 服务器上用 OpenAI 兼容 API 部署 LLM dev.to 审计 153 份 Claude Code 与 Codex 会话记录:User-Agent 与 URL 中的邮箱泄露
多元动态
登录 注册
大模型 www.aibase.com ·

蚂蚁开源 SingProbe:轻量探针实现大模型生成风险实时拦截,额外算力开销低于 0.5%

原文 我的收藏

蚂蚁集团 AI 安全实验室开源大模型内生安全护栏技术 SingProbe,与主流外挂式安全审核模型不同,它复用基座模型推理过程中的隐藏状态,在 token 级别做实时风险评估,宣称额外计算开销低于 0.5%,可同时完成意图分类、安全检测和幻觉识别,在医疗等高风险场景中可在内容输出前实现毫秒级拦截。目前兼容 29 个主流大模型,开源代码与预训练模型已同步发布。上述性能与'首个完整内生安全方案'等说法来自来源方转述,尚无论文或第三方复现证据。

AI推荐理由 蚂蚁集团 AI 安全实验室开源 SingProbe,复用基座模型推理时的隐藏状态做 token 级实时风险评估,宣称额外算力开销低于 0.5%,可同时做意图分类、安全检测与幻觉识别,兼容 29 个主流模型,并同步放出代码与预训练模型。信息点具体、对安全护栏与推理侧工程有实际参考价值;但来源为 AIbase 转述,未给出论文、基准数据或第三方复现,'首个完整内生安全方案'等表述属来源方说法,需保留不确定性。

72

正文

AntGroup Open Sources SingProbe: Real-Time Risk Blocking for Large Model Generation with Lightweight Probe, Computing Overhead Below 0.5% AIbase基地 Published inAI News · 2 min read · Sep 15, 2026 2 The AI Safety Lab of Ant Group has recently officially open-sourced the intrinsic safety barrier technology for large models, SingProbe. This technology differs from the mainstream external safety review model by reusing the hidden states during the reasoning process of the base model, achieving real-time risk assessment at the token level with less than 0.5% additional computing overhead. It can simultaneously perform intent classification, safety detection, and hallucination identification tasks. It shows significant advantages in high-risk scenarios such as healthcare, enabling millisecond-level blocking before content output. Currently, SingProbe is compatible with 29 mainstream large models. It is the first complete intrinsic safety technology solution open-sourced by a major tech company. Its open-source code and pre-trained models have been released simultaneously. AI New TermsSecurity GuardrailLarge ModelAnt Group This article is from AIbase Daily Scan to view Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications. —— Created by the AIbase Daily Team © Copyright AIbase Base 2024, Click to View Source -

来源报道 · 1 篇 · 以下链接均前往原始媒体

← 返回精选动态 原文

更多内容