[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepseek-think-hallucination-not-leak":3,"topics-all":36,"news-related-059e450a-7fd3-4532-8a64-af4d30aab14e":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"059e450a-7fd3-4532-8a64-af4d30aab14e","DeepSeek 紧急回应「对话泄露」：\u003Cthink>字符触发模型幻觉，非安全漏洞","5月19日，DeepSeek发布关于\u003Cthink>字符触发模型异常回复的说明，确认该问题属于特殊字符引发的模型幻觉，而非安全漏洞或隐私泄露。\n\n5月18日，多位用户反映在DeepSeek网页版输入\"\u003Cthink\"字符时，模型返回了疑似其他用户的历史问答内容，部分涉及八字等敏感信息，引发数据泄露担忧。技术团队调查后指出：\u003Cthink>本是模型输出推理过程的功能性标签，用户输入该字符时实际上是在构造一种异常prompt，诱导模型产生看似\"跨会话\"的内容。这是一种典型的prompt注入场景——输出并非来自真实会话数据，而是模型基于训练模式的幻觉合成。\n\n从大模型原理看\u003Cthink>标签与特定上下文的高频关联使模型对这类输入异常敏感。当用户刻意构造时，模型可能被诱导\"扮演\"某种特殊状态，输出本不应出现的内容。这与传统的越狱有相似之处，都是通过构造边界输入来绕过模型的安全约束。\n\n此次事件虽为虚惊，但揭示了三层问题：特殊字符的边界测试不足、用户对幻觉与泄露的辨识度低、隐私架构需持续加固。DeepSeek表示将针对该问题进行专项训练，优化模型对特殊字符场景的处理。整个行业都应从中吸取教训：在模型能力快速迭代的同时，边界条件的系统性测试和安全防护不能拖后腿。","https:\u002F\u002F36kr.com\u002Fp\u002F3816885092910212","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"5a748a9a-841b-43ec-bcf7-3eb55d7ba336","en","DeepSeek's leak response: \u003Cthink> triggers hallucination","On May 19, DeepSeek issued a statement about \u003Cthink> character triggers causing model abnormal replies, confirming that the issue is a special-character-induced model hallucination, not a security vulnerability or privacy leak.\n\nOn May 18, multiple users reported that inputting the \"\u003Cthink\" characters in DeepSeek's web version caused the model to return content that appeared to be other users' historical Q&A, some involving sensitive information like Chinese fortune-telling (Bazi), triggering data-leak concerns. The technical team's investigation found: \u003Cthink> is originally a functional tag for the model to output the reasoning process, and when users input this character they are actually constructing an abnormal prompt that induces the model to produce seemingly \"cross-session\" content. This is a typical prompt-injection scenario — the output is not from real session data, but the model's hallucination synthesis based on training patterns.\n\nFrom a large-model principle perspective, the high-frequency association between \u003Cthink> tags and specific contexts makes the model unusually sensitive to such input. When users deliberately construct them, the model can be induced to \"play\" some special state and output content that should not appear. This is similar to traditional jailbreaks — both bypass the model's safety constraints by constructing boundary input.\n\nThough a false alarm, the incident reveals three layers of issues: insufficient boundary testing for special characters, low user ability to distinguish between hallucination and leak, and privacy architecture needing continuous reinforcement. DeepSeek says it will conduct targeted training to optimize the model's handling of special-character scenarios. The whole industry should learn from this: as model capabilities iterate rapidly, systematic testing of edge conditions and safety protection cannot be left behind.","deepseek-think-hallucination-not-leak","2026-05-20T07:00:00Z","2026-05-20T07:12:26.418650Z","2026-08-19T02:08:40.142862Z",true,"agent",122,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"03c760d2-6603-4ac8-8773-8236330bc019","白宫豁免中国开放权重模型:开源路线获得 AI 安全审查「白名单」","us-carve-out-chinese-open-weight-ai-2026","2026-08-07T02:00:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"dc9fa5a5-737a-4c99-aea5-3479bd1a9422","白宫豁免中国开放权重模型：闭源派 vs 开源派的「监管」分水岭","white-house-china-open-weight-exemption","2026-08-06T04:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"d5203c93-2022-4769-a6d2-c7765ded2b40","腾讯混元 Hunyuan-A13B 开源实测:80B 总参 \u002F 13B 激活,GQA + FP8\u002FINT4 把 MoE 推理门槛打到消费卡","tencent-hunyuan-a13b-80b-13b-gqa-angelslim-moe","2026-07-30T06:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"fcca6449-5103-4d83-b85b-755524a8095c","OpenAI 智能体失控攻入 Hugging Face, GLM-5.2 做了美国前沿模型护栏里做不到的事","openai-agent-hf-glm-5-2-incident","2026-07-28T04:00:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"1a6769b6-c2c0-4fa3-b322-c8eb40e921fb","Qwen3Guard 把流式安全检测做进 Token 流水线：开源 Guardrail 模型进入「实时分类」时代","qwen3guard-streaming-safety-classification","2026-06-26T06:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"f5fb4dc3-734d-48f5-bcb3-c8cd12edea9a","GLM-5.2 Day-0 落地 MTT S5000：国产算力适配开源旗舰的工程化样本","glm-5-2-mtt-s5000-day-0-china-compute","2026-06-17T06:00:00+00:00"]