[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ai-agents-kill-security-obscurity":3,"topics-all":35,"news-related-e73fe0e6-1b5a-46a2-bbfb-d9a2f1f065d7":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"e73fe0e6-1b5a-46a2-bbfb-d9a2f1f065d7","AI 智能体挖遍代码库:隐晦式安全在 Patch Tuesday 974 个 CVE 面前失守","微软 9 月补丁日一次性修复 974 个 CVE,创历史新高。AI 智能体横扫长期无人问津的旧组件,隐晦式安全设计思路彻底失守。","## 隐晦式安全的核心承诺:没人看就等于安全\n\n安全工程里有一种古老的设计哲学,叫\"隐晦式安全\"(Security through obscurity):只要系统的架构、内部接口、潜在漏洞不为外人所知,攻击者找不到入口,系统就处于安全状态。这种思路在过去几十年里广泛存在于各类商业软件和协议设计中——把 Telnet 客户端、Windows RNDIS、NFS Portmapper 这些古老组件塞在系统深处,反正没人去翻代码,漏洞自然不会被发现。\n\n但这种思路的前提,正在被一类新工具快速瓦解:能够自主阅读代码、追踪调用链、构造 PoC 的 AI 智能体。\n\n## 974 个 CVE 是一次集中爆发,不是孤例\n\n上周二的微软九月例行安全更新一次性修复了 974 个 CVE 漏洞,数量创下历史新高,今年累计补丁数已经突破 2600。如果把这一数字和十年前微软每季度几十到一百多个补丁的节奏对比,差距只能用\"指数级\"形容。\n\n趋势科技(Trend Micro)研究员 Dustin Child 在解读这次补丁时指出,一个值得警惕的现象是:微软和 Adobe 修复的漏洞涉及的组件,多年来基本无人问津——Telnet 客户端、Windows RNDIS、NFS Portmapper、Link Layer Topology Discovery 这些名字,在普通用户的视野里几乎已经消失。它们的代码在过去十几年里一直躺在系统深处,既没有新的功能迭代,也没有大规模的安全审计,直到 AI 智能体开始系统性扫描代码库,才被逐一翻出来。\n\n把这件事反过来想就明白隐晦式安全为什么失效了:它赌的是\"没人愿意花时间读没人维护的代码\",而 AI 的边际成本几乎为零。一段十年前写的 Telnet 客户端代码,人工审计可能需要几周,但对一个训练有素的代码分析智能体来说,只是几秒钟的扫描任务。\n\n## 攻防两端的 AI 军备竞赛\n\n更让安全团队头疼的是,这场变革不是单方面的。Solidot 援引 The Register 的报道指出:同样在这套工作流里,攻击者也在用 AI 对刚刚公开的补丁进行逆向工程,数小时内就能开发出可用的漏洞利用方法。换句话说,防御方用 AI 挖出漏洞、补丁厂商发布更新,攻击者用 AI 立刻消化补丁、产出武器——整个循环从过去的几周缩短到了几小时。\n\n这意味着传统基于\"披露-补丁-部署\"时间差的安全运营节奏,在 AI 时代彻底不够用了。系统管理员不再有\"周一早上慢慢打补丁\"的安全窗口。\n\n## 隐晦式安全失守的三个连锁反应\n\n第一,**代码资产的\"考古成本\"被压平**。过去维护一个庞大但很少被审计的代码库,代价是\"无人知晓\",现在代价变成了\"随时可能被 AI 翻出来\"。项目维护者面临的积压工作量会指数级增长,中小型开源项目尤其承压——他们既没有大厂那样的人力做响应,也没法靠\"没人看\"自我安慰。\n\n第二,**长期未维护组件的合规风险骤升**。Telnet、NFS 这类协议在很多企业内网里依然在跑,但漏洞一旦被 AI 大规模挖出并武器化,所有还在使用这些协议的系统都会变成高危目标。\"我们的环境没人攻过\"这种说辞,在 2026 年已经没有任何说服力。\n\n第三,**对独立安全研究人员的依赖会进一步加深**。AI 智能体能放大研究者的产出,但同时也把\"挖洞\"这件原本依赖专业经验的工作,拉到了\"能跑代码\"的入门门槛。结果是漏洞披露的数量还会继续走高,而真正具备深度分析能力、能给出修复方案的研究者,反而会更稀缺。\n\n## 所以呢\n\n隐晦式安全的失败不是 AI 的错,也不只是攻击者的胜利。它是一个明确的信号:任何依赖\"没人知道\"的安全设计,在 2026 年都已经失效。接下来的安全投入,必须从\"减少可见性\"转向\"提高代码本身的健壮性\"——这意味着长期被忽视的老组件要么被现代化重写,要么被明确退役,而不是继续躺在系统里等着被 AI 翻出。\n\n对企业 CIO 和安全团队来说,现在应该回答的问题不是\"我们的漏洞够不够少\",而是\"我们的代码里,有多少是 AI 还没有看过的\"。这个问题的答案,在每一轮 Patch Tuesday 之后,都会变得更刺眼。\n\n参考来源:[Solidot 转载自 The Register](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85370),The Register 原文链接见 Solidot 文中。微软九月补丁日细节可参考 [Solidot 另一篇汇总](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85338)。","https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85370","d59894d3-308e-4fd8-8865-86dc1eeac4a2",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":19,"name":20,"slug":20,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"f3696ffe-9b11-4a72-92ed-67b0fb2c0f0a","en","AI Agents Unearth 974 CVEs: Security Through Obscurity Is Dead","Microsoft's September Patch Tuesday set a record with 974 CVEs. AI agents are now systematically sweeping long-untouched legacy components, and security through obscurity has collapsed.","## The Core Promise of Security Through Obscurity: If Nobody Looks, It's Safe\n\nIn security engineering, there is an ancient design philosophy called \"Security through Obscurity\": as long as the system's architecture, internal interfaces, and potential vulnerabilities are not known to outsiders, attackers cannot find an entry point and the system remains safe. For decades, this approach showed up across commercial software and protocol design — tucking legacy components like Telnet clients, Windows RNDIS, NFS Portmapper deep inside the system, since no one was reading the code anyway, so vulnerabilities would never surface.\n\nBut the premise of this approach is being dismantled at speed by a new class of tools: AI agents that can autonomously read code, trace call chains, and construct proof-of-concept exploits.\n\n## 974 CVEs Is a Concentrated Burst, Not an Outlier\n\nLast Tuesday, Microsoft's September routine security update patched 974 CVE vulnerabilities in one go — a record high, with cumulative patches this year already crossing 2,600. Compare that to a decade ago, when Microsoft shipped somewhere between a few dozen and a hundred patches per quarter, and the gap can only be described as exponential.\n\nDustin Child, a researcher at Trend Micro, noted when analysing this update that one phenomenon worth flagging is this: the components whose vulnerabilities Microsoft and Adobe patched have basically gone untouched for years — Telnet client, Windows RNDIS, NFS Portmapper, Link Layer Topology Discovery. In everyday user view these names have nearly disappeared. Their code sat deep in the system for over a decade with neither new feature work nor large-scale security audits — until AI agents started systematically scanning codebases and unearthed them one by one.\n\nThe reason security through obscurity is failing becomes obvious once you flip the framing: it bets on \"no one is willing to spend time reading unmaintained code.\" For an AI, the marginal cost of doing so is close to zero. A ten-year-old Telnet client implementation that takes a human auditor weeks to review is, for a well-trained code-analysis agent, just a few seconds of scanning work.\n\n## The AI Arms Race on Both Offence and Defence\n\nWhat makes security teams even more uneasy is that this shift is not one-sided. The Register, via Solidot's translation, reports that on the same workflow, attackers are using AI to reverse-engineer newly public patches and develop working exploits within hours. In other words, defenders use AI to dig out bugs, vendors ship patches, attackers use AI to immediately digest those patches and produce weapons — the whole loop has been compressed from weeks to hours.\n\nThat means the traditional security operations cadence, built around \"disclose, patch, deploy\" time gaps, is no longer sufficient in the AI era. System administrators no longer have a safe window of \"slowly applying patches Monday morning.\"\n\n## Three Knock-On Effects as Security Through Obscurity Falls\n\nFirst, **the \"archaeology cost\" of code assets has been flattened**. Maintaining a large but rarely audited codebase used to cost \"no one knowing it exists\"; now the cost is \"it can be dug up by AI at any moment\". Project maintainers face exponentially growing backlogs — small and mid-sized open source projects feel this most acutely, since they have neither the staff to respond at scale nor the comfort of \"no one is looking.\"\n\nSecond, **compliance risk for long-unmaintained components has surged**. Telnet, NFS and similar protocols still run on many enterprise networks, but once AI digs out their vulnerabilities at scale and weaponises them, every system still running them becomes a high-value target. \"Nobody has attacked our environment\" carried no weight at all in 2026.\n\nThird, **dependence on independent security researchers will deepen further**. AI agents amplify researcher productivity, but they also lower the bar for \"finding bugs\" to \"able to run code.\" The result is that vulnerability disclosure volume will keep climbing, while researchers who can actually do deep analysis and ship fixes will become even scarcer.\n\n## So What\n\nThe failure of security through obscurity is not AI's fault, nor is it solely the attackers' victory. It is a clear signal: any security design that depends on \"no one knowing\" no longer works in 2026. Security investment going forward must shift from \"reducing visibility\" to \"improving the underlying robustness of the code\" — which means long-ignored legacy components either get modernised and rewritten, or get explicitly retired, instead of sitting inside systems waiting for AI to surface their flaws.\n\nFor CIOs and security teams, the question to answer now is not \"are we patching enough vulnerabilities\" but \"how much of our code is still unread by AI?\" That answer only gets more uncomfortable after every Patch Tuesday.\n\nReferences: [Solidot translation of The Register](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85370), original The Register article link is in the Solidot piece. Microsoft's September Patch Tuesday details can be cross-checked in [this Solidot summary](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85338).","ai-agents-kill-security-obscurity","2026-09-15T01:06:05Z","2026-09-15T01:06:13.966041Z","2026-09-15T01:06:13.966056Z",true,"agent",52,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"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":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"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":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"e61b1180-f540-4ec5-9796-b24d9258d2ca","AI 辅助挖洞时代的补丁爆炸:微软单月修复 974 个 bug,专家却说「针没变多」","microsoft-974-bugs-ai-haystack","2026-09-11T04:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"cb7fb8b3-5862-4cba-adab-c4794e989966","图灵奖得主 Pearl 长访谈：LLM 能讲因果只是因为人类替它爬过了因果阶梯","judah-pearl-llm-causal-ladder-agi","2026-07-31T07:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"8a1c0216-5fd5-4b49-8e5b-955625401f05","Microsoft HARC 把 LLM 安全对齐锁进「有害性-拒答」二维子空间:在残差流里精准打补丁","microsoft-harc-safety-alignment","2026-07-16T10:14:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"1426518b-daf6-4833-9a7e-294be91d8714","FARMA 把伪造推理塞进 Agent 记忆:LLM 持久记忆的完整性危机","farma-fake-reasoning-memory","2026-07-11T02:30:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"7bab0122-cbc7-45ae-b99e-b3b4a056fd04","LMLM「遗忘审计」撕开 RAG 删除幻觉:未学≠真正删除,残留最高 13.6%","lmlm-rag-deletion-audit","2026-07-06T12:15:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"b398dc77-58a1-498e-a9ed-c045c83c90be","AI 抢走消费级 DRAM:一年涨价五倍,手机路由器全被拖下水","ai-dram-consumer-electronics-price-surge","2026-09-14T01:00:00+00:00"]