[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-microsoft-mdash-finds-16-windows-flaws-zh":3,"article-related-microsoft-mdash-finds-16-windows-flaws-zh":37,"series-research-aefdd28e-fccb-46ca-a78b-ad6ad718058d":89},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":20,"translated_content":10,"views":21,"is_premium":22,"created_at":23,"updated_at":23,"cover_image":11,"published_at":24,"rewrite_status":25,"rewrite_error":10,"rewritten_from_id":26,"slug":27,"category":28,"related_article_id":29,"status":30,"google_indexed_at":10,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":31,"topic_cluster_id":35,"embedding":36,"is_canonical_seed":22},"aefdd28e-fccb-46ca-a78b-ad6ad718058d","Microsoft MDASH 找出 16 個 Windows 漏洞","\u003Cp data-speakable=\"summary\">\u003Ca href=\"\u002Ftag\u002Fmicrosoft\">Microsoft\u003C\u002Fa> 的 MDASH AI 找出 16 個 Windows 漏洞，包含 4 個重大遠端執行漏洞。\u003C\u002Fp>\u003Cp>說真的，這篇重點不是「AI 很會找洞」。重點是 Microsoft 把它做成一套流程。MDASH 已經在 Windows 裡抓到 16 個新漏洞，還幫忙進了 5 月 12 日的 Patch Tuesday。\u003C\u002Fp>\u003Cp>其中 4 個被列為 Critical。這代表它不是只會吐報告。它真的碰到企業會怕的那種 RCE。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>指標\u003C\u002Fth>\u003Cth>數值\u003C\u002Fth>\u003Cth>意義\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>Windows 漏洞數\u003C\u002Ftd>\u003Ctd>16\u003C\u002Ftd>\u003Ctd>已找到實際新問題\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Critical 漏洞\u003C\u002Ftd>\u003Ctd>4\u003C\u002Ftd>\u003Ctd>包含遠端程式碼執行\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>私測時間\u003C\u002Ftd>\u003Ctd>2026 年 6 月\u003C\u002Ftd>\u003Ctd>企業可先接觸\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>CyberGym 分數\u003C\u002Ftd>\u003Ctd>88.45%\u003C\u002Ftd>\u003Ctd>公開榜單表現很高\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>內部測試漏洞\u003C\u002Ftd>\u003Ctd>21\u003C\u002Ftd>\u003Ctd>全部都被找出來\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>MDASH 到底找到了什麼\u003C\u002Fh2>\u003Cp>Microsoft 說，這 4 個 Critical 都碰到核心 Windows 元件。這些元件在企業環境裡超常見。只要其中一個出事，影響範圍就不小。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779041037625-66oq.png\" alt=\"Microsoft MDASH 找出 16 個 Windows 漏洞\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>其中一個是 CVE-2026-33827。它是 Windows IPv4 stack 的 remote unauthenticated use-after-free。簡單講，攻擊者可能透過特製封包去打。\u003C\u002Fp>\u003Cp>另一個是 CVE-2026-33824。它是 IKEEXT service 的 pre-auth double-free。這會影響 RRAS VPN、DirectAccess 和 Always-On VPN。\u003C\u002Fp>\u003Cp>另外兩個 Critical 打到 Netlogon 和 Windows DNS \u003Ca href=\"\u002Fnews\u002Fwhy-claudes-real-advantage-is-not-raw-intelligence-zh\">Cl\u003C\u002Fa>ient。兩個都拿到 CVSS 9.8。剩下 12 個是 Important，類型包含 DoS、權限提升、資訊洩漏和安全功能繞過。\u003C\u002Fp>\u003Cul>\u003Cli>4 個 Critical 都是高風險類型\u003C\u002Fli>\u003Cli>2 個 Critical 的 CVSS 是 9.8\u003C\u002Fli>\u003Cli>12 個漏洞列為 Important\u003C\u002Fli>\u003Cli>涵蓋 tcpip.sys、http.sys、ikeext.dll、telnet.exe\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>真正猛的是它的 agentic 架構\u003C\u002Fh2>\u003Cp>MDASH 不是單一模型在跑。Microsoft 說，它會串起 100 多個 \u003Ca href=\"\u002Ftag\u002Fai-agents\">AI agents\u003C\u002Fa>。每個 agent 負責一小段工作。\u003C\u002Fp>\u003Cp>有的 agent 掃 source code。有的 agent 判斷漏洞是不是假的。還有一段流程會嘗試重現問題。這種拆法，比單純掃描器更像一條自動化研究線。\u003C\u002Fp>\u003Cp>Microsoft 主管 Taesoo Kim 在部落格寫了一句很直白的話。\u003C\u002Fp>\u003Cblockquote>“The model is one input. The system is the product.” — Taesoo Kim, Microsoft vice president for agentic security\u003C\u002Fblockquote>\u003Cp>這句話我覺得很準。現在大家都愛談模型分數。但真正有用的是整套系統。包含驗證、重現、人工審核，還有後面的修補流程。\u003C\u002Fp>\u003Cp>MDASH 也強調 model-agnostic。意思是底層模型可以換。這點很實際。因為模型更新速度太快了。今天很強的模型，明天可能就被別家超車。\u003C\u002Fp>\u003Cp>而且 Microsoft 不是只押一條路。它前陣子才公布 \u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fsecurity\u002Fblog\u002F\" target=\"_blank\" rel=\"noopener\">Project Glasswing\u003C\u002Fa>。那是另一個 AI 漏洞研究計畫。裡面還有 \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa> 和 \u003Ca href=\"\u002Fnews\u002F7-milestones-in-the-claude-timeline-2026-zh\">Clau\u003C\u002Fa>de \u003Ca href=\"\u002Ftag\u002Fmythos\">Mythos\u003C\u002Fa> Preview。\u003C\u002Fp>\u003Cul>\u003Cli>100+ AI agents 協作\u003C\u002Fli>\u003Cli>模型可替換，不綁死單一供應商\u003C\u002Fli>\u003Cli>包含驗證與重現步驟\u003C\u002Fli>\u003Cli>Microsoft 同時推多條 AI 安全路線\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>數字好看，但還不能直接下結論\u003C\u002Fh2>\u003Cp>Microsoft 也丟了一些 benchmark。這些數字確實不差，但你不能直接拿來當企業安全成效。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779041036428-q1tj.png\" alt=\"Microsoft MDASH 找出 16 個 Windows 漏洞\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>它說 MDASH 在內部 Windows test driver 找到 21 個刻意植入的漏洞，而且 0 false positives。這表示它不是亂槍打鳥。\u003C\u002Fp>\u003Cp>它還說，在 \u003Ca href=\"https:\u002F\u002Fwww.cybergym.org\" target=\"_blank\" rel=\"noopener\">CyberGym\u003C\u002Fa> 的 vulnerability reproduction 任務上，分數是 88.45%。Microsoft 還說這是公開榜單第一。\u003C\u002Fp>\u003Cp>但 benchmark 就是 benchmark。真實世界裡，企業有舊系統、有客製化程式、還有一堆沒人敢動的流程。分數高，不代表落地就順。\u003C\u002Fp>\u003Cul>\u003Cli>21 個內部測試漏洞全數找出\u003C\u002Fli>\u003Cli>0 false positives\u003C\u002Fli>\u003Cli>CyberGym 分數 88.45%\u003C\u002Fli>\u003Cli>歷史 MSRC 案例也能回收不少\u003C\u002Fli>\u003C\u002Ful>\u003Cp>Greyhound Research 的首席分析師 \u003Ca href=\"https:\u002F\u002Fgreyhoundresearch.com\" target=\"_blank\" rel=\"noopener\">Sanchit Vir Gogia\u003C\u002Fa> 說得很直。他認為 Microsoft 同時扮演平台商、安全廠商、\u003Ca href=\"\u002Ftag\u002Fai-\">AI 基礎設施\u003C\u002Fa>供應商、\u003Ca href=\"\u002Fnews\u002Fopencode-cli-acp-server-support-zh\">Open\u003C\u002Fa>AI 合作方、Mythos 整合者和 agentic security 供應商。\u003C\u002Fp>\u003Cp>這種位置很強。也很集中。安全工具市場如果越來越靠少數大廠，企業會更省事，但也更難分散風險。\u003C\u002Fp>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.beaglesecurity.com\" target=\"_blank\" rel=\"noopener\">Beagle Security\u003C\u002Fa> 顧問 Sunil Varkey 的看法也很直接。他把這件事看成 AI 對 AI 的競賽。攻擊者在用 AI 加速，防守方也得跟上。\u003C\u002Fp>\u003Cul>\u003Cli>Microsoft 在多個安全角色上同時發力\u003C\u002Fli>\u003Cli>高分 benchmark 不等於真實環境成功\u003C\u002Fli>\u003Cli>企業會更依賴大廠整合方案\u003C\u002Fli>\u003Cli>AI 防守已經進入對抗階段\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>這對資安團隊代表什麼\u003C\u002Fh2>\u003Cp>對企業來說，MDASH 代表漏洞管理可能會變形。以前是定期掃描，再人工分類，最後等 patch window。\u003C\u002Fp>\u003Cp>現在 Microsoft 想做的是連續發現。AI \u003Ca href=\"\u002Ftag\u002Fagents\">agents\u003C\u002Fa> 先跑第一輪。人類再做判斷。最後把修補流程接上去。\u003C\u002Fp>\u003Cp>講白了，這會讓資安團隊少做一點重複工作。但前提是流程要乾淨。否則你只會得到更多 findings。\u003C\u002Fp>\u003Cp>我覺得最實際的問題不是「它能不能找洞」。而是「它能不能縮短修補時間」。如果不能，那就是更快產生報告而已。\u003C\u002Fp>\u003Cp>還有一個現實問題。自動化越強，治理就越重要。沒有 validation、沒有 patch 優先級、沒有 change management，AI 找再多洞也只是堆資料。\u003C\u002Fp>\u003Cp>所以如果 Microsoft 6 月真的開 private preview，企業第一件事應該是看它能不能接進既有流程。像是 ticketing、patch triage、exploit validation 和風險排序。\u003C\u002Fp>\u003Cp>第二件事，是看它會不會把團隊帶進「報告很多、修補很慢」的老問題。這種情況下，AI 只是把噪音加速而已。\u003C\u002Fp>\u003Ch2>更大的背景是什麼\u003C\u002Fh2>\u003Cp>這波其實不是單一產品新聞。它是在看 AI 安全工具怎麼長大。從掃描器，到 fuzzing，再到 agentic workflow，路線已經很清楚。\u003C\u002Fp>\u003Cp>以前很多工具只會丟 CVE。現在廠商開始比誰能驗證、誰能重現、誰能直接接進修補流程。這才是企業會買單的地方。\u003C\u002Fp>\u003Cp>對台灣開發者來說，這也不是離很遠的事。只要你的軟體有 Windows 元件、VPN、DNS、網路堆疊，這些漏洞就會碰到你。不是每個人都會寫 kernel code，但很多人都得面對 patch 管理。\u003C\u002Fp>\u003Cp>我自己的判斷很簡單。接下來 12 個月，資安產品會越來越像「多 agent 工作流」。單一模型的 demo 會越來越少。能接流程、能驗證、能落地的系統，才會留下來。\u003C\u002Fp>\u003Ch2>結語：先看落地，不要先看話術\u003C\u002Fh2>\u003Cp>如果你是資安團隊，先問三件事。它能不能接你現有流程。它會不會產生太多假警報。它能不能真的縮短修補時間。\u003C\u002Fp>\u003Cp>如果答案都不錯，那 MDASH 這類工具就值得試。反過來說，如果只是多一個漂亮儀表板，那就不用太嗨。\u003C\u002Fp>\u003Cp>我會繼續看 2026 年 6 月的 private preview。真正的考驗，不在發表會，而在企業環境裡。\u003C\u002Fp>","Microsoft 的 MDASH AI 找出 16 個 Windows 漏洞，含 4 個重大 RCE，並將在 6 月開放企業私測。","www.csoonline.com","https:\u002F\u002Fwww.csoonline.com\u002Farticle\u002F4170785\u002Fmicrosofts-new-ai-system-finds-16-windows-flaws-including-four-critical-rces.html",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779041037625-66oq.png",[13,14,15,16,17,18,19],"Microsoft","MDASH","Windows 漏洞","AI 資安","遠端程式碼執行","Patch Tuesday","CyberGym","zh",0,false,"2026-05-17T18:03:35.214691+00:00","2026-05-17T18:03:35.01+00:00","done","3f06daf5-5415-4d40-a546-d7c64a2f1d13","microsoft-mdash-finds-16-windows-flaws-zh","research","94f7efad-6f63-4873-9d18-62267154c2c7","published",[32,33,34],"MDASH 已找出 16 個 Windows 漏洞，含 4 個 Critical。","它的重點不是單一模型，而是 100+ AI agents 的整套流程。","benchmark 成績亮眼，但企業落地還要看修補與治理流程。","0c35a120-52fc-41fc-afa3-d404eb934158","[-0.0375328,0.00034178552,0.0071258326,-0.064522326,-0.008017336,0.018630179,-0.0067302333,0.00977215,0.003686404,0.02258062,-0.00018157264,-0.02012978,0.045787163,0.008178233,0.10901795,0.0143460985,0.01715898,0.025953664,-0.008307651,-0.0011028466,-0.0002870592,-0.005205598,0.001360621,0.001918899,-0.020303555,-0.004877385,0.029725399,0.0054865284,0.03008519,-0.02684104,0.016214043,0.014621728,-0.0010724842,0.002410477,-0.011317288,0.017838089,0.027278924,-0.017253926,0.012303492,0.007976022,-0.010617864,-0.0054512513,-0.0024658646,0.0128916595,0.012964785,0.0055287112,-0.007194062,-0.0019879981,-0.014592384,-0.0083091725,0.014976263,0.009332149,-0.009778648,-0.150032,0.016654883,0.01753607,0.0027693848,0.006667073,0.0065916697,-0.01589281,-0.0075342534,0.0044948473,-0.03469983,-0.03536311,-0.004992752,-0.043424163,0.019173602,0.025247408,-0.018416516,-0.0018438895,-0.028082356,-0.006765357,0.009662754,-0.032151215,-0.00223788,-0.022735512,-0.025433289,-0.018962698,0.012021514,0.03516832,-0.0023825823,-0.034705173,0.00993939,0.011930882,-0.007371305,-0.002062364,-0.010754433,0.0027087221,0.0041322852,0.02569853,-0.0068148565,0.031798743,-0.0062273606,-0.016034236,-0.0062414645,-0.019778172,0.027575452,0.0030718108,0.021401962,0.004288914,-0.0011954925,-0.032040942,-0.003178648,0.011631027,0.015089561,-0.011321399,-0.014617675,0.01506711,0.017340736,-0.0050640525,-0.0025940961,-0.042747967,-0.019075116,0.029297091,-0.021415884,-0.13531888,0.0011856183,-0.0013839464,0.004633903,0.019151531,-0.029749773,-0.0066172616,-0.01611631,0.023522094,-0.009308567,-0.012686656,-0.012253459,-0.005095435,0.008996783,0.0073968577,-0.0048044114,-0.025355073,0.018947687,-0.016669389,-0.016651966,0.016878432,-0.011576487,-0.02410576,-0.02591325,-0.03501603,0.0155255115,0.016894994,-0.0036978892,-0.02283266,-0.03880496,0.021262133,-0.017725969,0.018481085,0.020868782,0.0025920898,0.020109352,-0.038866106,-0.0153667545,-0.011819186,-0.011757416,-0.042137273,-0.0129396245,-0.0017484829,-0.007508078,0.020517947,0.0066156406,-0.022780402,-0.005671279,0.032796636,0.024743544,0.0076043787,0.0055467375,0.005765273,-0.0030768448,0.004245693,0.024622858,-0.023674347,-0.014237053,0.011827828,-0.0051377,0.0077132885,0.011291513,-0.0015048917,-0.0032765854,0.0054835356,-0.015060565,-0.013056053,0.0058126673,0.01344877,0.0106834965,-0.0043435046,0.017727436,0.0029758748,0.017667858,-0.006506966,0.0035251966,-0.0014347531,0.005298207,0.0022683775,-0.007643705,-0.048395846,-0.01426895,-0.003974768,-0.0009334181,0.01217064,-0.018807445,-0.02736238,-0.019407835,-0.018258965,-0.0024678016,-0.0019341476,-0.009107444,-0.0074171335,-0.0031245206,-0.008190757,-0.010147685,0.015174393,0.018877799,-0.005706412,0.00055894844,-0.0028817558,-0.03416605,0.012287993,0.008111483,-0.0005154163,0.00049351016,0.0002586907,0.010642189,0.012383364,0.016890435,-0.02927832,0.005135807,-0.014576641,0.008015853,0.027759168,0.014809533,0.0010758821,0.0063255047,-0.0024198403,-0.0023806908,-0.018104214,-0.014202254,0.006870348,0.017027002,0.012856926,-0.0084752245,0.016061407,-0.014351326,0.022833448,0.026316334,-0.026506659,0.020066814,-0.02054144,-0.025160858,0.0032229533,-0.00092344295,0.0008672734,0.008415143,0.0074941376,0.02104798,-0.0027042155,0.011651449,0.021434953,-0.016119806,-0.004048379,0.011185626,0.011993583,-0.035522338,-0.02343013,0.03278215,0.019169236,-0.012541254,0.012892276,-0.025361046,0.009517048,-0.021348258,0.018592155,-0.016313432,0.010459337,0.019340577,-0.02602291,-0.052878752,-0.010881589,-0.0074591017,-0.036650296,0.003927915,0.017323285,-0.04177654,-0.003970423,-0.029796269,-0.020202048,-0.019382345,0.005648372,-0.024906367,0.0012349084,0.020877881,0.015236517,0.007994673,-0.001847482,0.026065907,-0.026871065,0.010536188,-0.0032354228,0.0058131414,-0.0069277184,-0.005924909,0.014845547,0.039367285,0.04890968,-0.005444516,-0.012412504,0.01447167,0.015953572,0.01729003,-0.0069719683,-0.032934584,-0.0062826416,-0.004095613,0.0005191394,-0.016638236,-0.038097292,-0.0010127546,-0.017002972,-0.01354021,-0.012916392,0.00054682256,0.0032542145,0.008134279,0.017387027,-0.024062453,-0.03649582,0.01058793,0.008843067,0.034593794,-0.008292061,0.0019231922,0.022777304,-0.009729437,0.021660438,-0.0024212014,0.012538968,0.02064657,0.0029913718,-0.0041903704,-0.013642252,0.017503541,0.009947071,-0.0005845168,-0.0037710103,-0.013341231,0.02437972,0.010620125,-0.0035990998,0.017040487,-0.020210834,0.022017937,-0.009791286,0.038975112,-0.016209062,-0.03592416,0.025031326,-0.0015592518,-0.015070882,0.006983996,0.015581044,-0.023273457,-0.010141447,0.004377071,-0.015103099,0.012049108,-0.036001313,0.0054911724,0.010136501,0.005365404,-0.0054434645,0.015708944,0.014867731,0.018174196,-0.015715262,-0.00028132112,0.0253656,0.007479426,0.008551428,0.012387864,-0.016640447,0.003708512,0.030718327,0.006679356,-0.041492715,-0.03169164,0.02567007,0.009164684,0.009937546,0.016988913,-0.012166138,-0.01208519,0.008921143,-0.009510768,-0.0015286865,0.013620575,-0.010082726,-0.01958004,-0.00027470026,-0.007448409,-0.017492006,0.005924272,0.014876075,0.011806908,0.0036609985,-0.019242376,0.009486955,-0.00019192214,0.019418534,0.0027235365,-0.034391303,0.0022637967,-0.008893006,-0.009645321,0.004581553,0.019710084,-0.0049693906,0.019155527,-0.01416952,-0.005048688,0.021643987,0.0059142364,-0.005296495,0.009354989,0.001002339,6.3086816e-05,0.022309091,-0.00596427,0.006992196,-0.007593033,-0.0131443385,0.023138724,-0.011095933,-0.0030762723,-0.044676248,-0.015803007,-0.04274412,-0.0071945577,-0.0010082987,-0.029564071,-0.02611364,0.0042889565,-0.023133058,-0.0015303591,-0.002243898,0.01695034,-0.034449335,0.0046280525,-0.03199045,-0.041344494,-0.005187041,0.02195165,0.028117118,0.03498472,-0.00076603354,-0.009986755,0.019531863,0.0023886906,0.014506963,0.0058292705,-0.012051995,-0.021877676,0.0019309483,0.023619978,-0.0042074523,0.009832975,-0.00828769,0.00669716,-0.006466695,0.0012634023,0.016076835,-0.021807054,-0.0113715585,-0.008701994,-0.01633844,-0.012813522,0.0018949762,-0.011433634,0.03712277,8.102628e-06,-0.003993662,-0.005563816,-0.009572739,-0.004549254,0.0025240388,0.01139786,0.023090126,-0.028654562,-0.020195419,0.005009297,0.032953046,0.024578357,0.02008635,0.011733479,0.012755432,-0.018303474,-0.020635309,-0.04872471,-0.008757819,-0.0063389973,-0.010948755,-0.0033084555,-0.003369742,0.0037181661,-0.025474066,-0.022007389,-0.03329075,0.014851072,0.0060903206,-0.011477162,0.00058881345,-0.025066432,-0.006315205,0.028470684,-0.0058908486,-0.0036313436,-0.019454358,-8.2596096e-05,0.018481657,0.0011110102,0.001518199,0.008555484,-0.019179337,-0.0135538615,0.009273306,0.0077802315,-0.034659144,0.028879317,-0.010243947,0.0082188845,-0.0004458522,-0.028931268,0.035882417,0.00264792,0.0038490822,0.021717243,0.010776343,0.003250798,0.004717707,0.056965694,0.004441017,-0.014721817,-0.0034555658,-0.042051654,0.005020696,0.013956476,-0.07641234,-0.024856085,0.009896825,0.002498668,-0.033775907,-0.04510414,0.031742718,-0.011824145,0.013571786,0.012167836,0.014126588,-0.01723075,0.032758605,0.009007713,-0.0056157405,-0.006202129,0.0031412246,0.017171048,0.009952209,0.00861538,0.031329155,-0.017843911,0.021803837,0.004080753,0.0094138505,-0.021883976,-0.005701195,0.0039967354,0.0053669875,0.015002807,0.016469467,-0.0028134105,-0.008683938,0.029295918,-0.0021536725,0.011369692,-0.004824833,0.013686019,0.012099551,0.0077447435,-0.007934614,-0.004989645,-0.023627589,0.007594752,-0.01074241,-0.00094261474,-0.02336633,0.0039618826,-0.0047519035,0.00042450085,-0.023280224,-0.024429908,-0.010778676,-0.035658546,-0.010462084,-0.004538966,-0.021127602,0.012918396,-0.017123817,-0.010999644,-0.022936672,-0.002936355,-0.029629935,0.01190667,-0.004951439,-0.007912293,-0.0015797194,0.017805247,0.0125940805,-0.012949408,-0.004827764,0.0054289247,0.019756373,0.008254882,-0.011982296,-0.0127808815,0.0051207957,0.02612657,-0.014317856,-0.019856565,-0.020007422,-0.0006887813,-0.10139864,-0.015867839,-0.019988148,-0.008544092,-0.01835535,-0.005352723,0.03299691,-0.023191923,-0.01032397,-0.03882249,-0.0009843203,0.0077164695,-0.014480297,-0.008849863,0.0018848644,-0.005119998,0.0015401068,0.021287683,-0.009452233,-0.018567603,-0.034077574,0.0118851,0.003961881,-0.034400165,-0.0096276505,-0.016677532,-0.0071835183,-0.008716151,-0.0036676934,-0.0064755753,-0.028521463,-0.14670607,0.010411229,-0.019012406,-0.0055250134,0.024283972,0.021947416,-0.0116917305,-0.02538041,0.020898376,-0.0055102813,-0.008105221,-0.012906663,-0.02265803,-0.01149066,-0.017372651,0.11027719,-0.022922145,0.0015013968,-0.00023635525,-0.010601367,0.008706469,-0.006315989,0.004263204,0.0019117153,0.0071134213,0.021842841,0.024062656,-0.025393011,-0.016092416,-0.0029423865,0.009025629,0.01375934,0.0039371355,-0.022799915,-0.01420946,-0.03419035,0.018241044,-0.003606369,0.01824806,0.0069278213,0.041221473,0.022130316,0.018842595,-0.008060528,-0.026192356,-0.00504552,0.019958444,-0.004000252,-0.021871705,0.012871796,-0.04319456,-0.07289655,0.029928321,-0.008884633,0.013799323,0.006354902,-0.025436355,0.0069265836,0.029830715,-0.03072596,-0.009529179,-0.010548375,-0.022256048,0.022675147,0.0069245864,-0.010218971,0.026055016,0.03287207,0.013357826,0.028252384,-0.013300978,-0.015665904,0.000609164,0.009916018,-0.013396753,-0.016123844,0.01590234,0.04198556,0.0074509983,-0.008276693,0.005491773,-0.028182432,0.017790306,-0.0113025885,0.015033761,0.008898748,0.01113835,-0.039943777,-0.01354507,-0.019269545,-0.008019219,0.03078535,-0.018061934,-0.005205589,0.020535959,-0.0008444665,-0.009873841,0.011595347,0.011609394,-0.0035040455,-0.011901675,-0.019408196,-0.007114082,0.02409785,0.039574116,0.0023830088,0.005858495,-0.0020000343,0.032776915,0.013802868]",{"tags":38,"relatedLang":48,"relatedPosts":52},[39,41,43,45,47],{"name":16,"slug":40},"ai-資安",{"name":13,"slug":42},"microsoft",{"name":14,"slug":44},"mdash",{"name":15,"slug":46},"windows-漏洞",{"name":17,"slug":17},{"id":29,"slug":49,"title":50,"language":51},"microsoft-mdash-finds-16-windows-flaws-en","Microsoft’s MDASH finds 16 Windows flaws","en",[53,59,65,71,77,83],{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":28},"902b314d-316c-48aa-9a2a-e4d16f32d2ac","browser-exploit-benchmarks-prove-ai-security-here-zh","為什麼瀏覽器 exploit 基準已證明 AI 安全威脅就在眼前","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779019382261-mfmw.png","2026-05-17T08:03:21.360298+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":28},"6ca303f0-7bd4-4bb2-be58-70d80da5ec40","why-ai-safety-teams-are-wrong-blame-only-alignment-zh","為什麼 AI 安全團隊錯把問題全怪在對齊","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778947417022-ak55.png","2026-05-16T16:03:16.319335+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":28},"50b2e74e-7248-43a3-8775-451bf2569f33","why-fine-tuning-llms-domain-tasks-right-default-zh","為什麼針對領域任務微調 LLM 才是預設選項","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778916229431-9olk.png","2026-05-16T07:23:32.255569+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":28},"001e062e-f246-4bf0-aa04-27506febcf7b","refdecoder-reference-conditioned-video-decoder-zh","RefDecoder 讓影片解碼器吃參考圖","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912646805-czy9.png","2026-05-16T06:23:33.170076+00:00",{"id":78,"slug":79,"title":80,"cover_image":81,"image_url":81,"created_at":82,"category":28},"b9516feb-41d5-42a3-887e-7b47c5c9ffb7","atlas-one-token-visual-reasoning-zh","ATLAS 用一個 token 做視覺推理","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912032775-hp0w.png","2026-05-16T06:13:34.693651+00:00",{"id":84,"slug":85,"title":86,"cover_image":87,"image_url":87,"created_at":88,"category":28},"bfd03801-a200-4222-9370-8b441be41483","entitybench-long-range-video-consistency-zh","EntityBench 盯住長片一致性","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778911845686-4mc8.png","2026-05-16T06:10:27.85068+00:00",[90,95,100,105,110,115,120,125,130,135],{"id":91,"slug":92,"title":93,"created_at":94},"f18dbadb-8c59-4723-84a4-6ad22746c77a","deepmind-bets-on-continuous-learning-ai-2026-zh","DeepMind 押注 2026 連續學習 AI","2026-03-26T08:16:02.367355+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"f4a106cb-02a6-4508-8f39-9720a0a93cee","ml-papers-of-the-week-github-research-desk-zh","每週 ML 論文清單，為何紅到 GitHub","2026-03-27T01:11:39.284175+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"c4f807ca-4e5f-47f1-a48c-961cf3fc44dc","ai-ml-conferences-to-watch-in-2026-zh","2026 AI 研討會投稿時程整理","2026-03-27T01:51:53.874432+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"9f50561b-aebd-46ba-94a8-363198aa7091","openclaw-agents-manipulated-self-sabotage-zh","OpenClaw Agent 會自己搞砸自己","2026-03-28T03:03:18.786425+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"11f22e92-7066-4978-a544-31f5f2156ec6","vega-learning-to-drive-with-natural-language-instructions-zh","Vega：使用自然語言指示進行自駕車控制","2026-03-28T14:54:04.847912+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"a4c7cfec-8d0e-4fec-93cf-1b9699a530b8","drive-my-way-en-zh","Drive My Way：個性化自駕車風格的實現","2026-03-28T14:54:26.207495+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"dec02f89-fd39-41ba-8e4d-11ede93a536d","training-knowledge-bases-with-writeback-rag-zh","用 WriteBack-RAG 強化知識庫提升檢索效能","2026-03-28T14:54:45.775606+00:00",{"id":126,"slug":127,"title":128,"created_at":129},"3886be5c-a137-40cc-b9e2-0bf18430c002","packforcing-efficient-long-video-generation-method-zh","PackForcing：短影片訓練也能生成長影片","2026-03-28T14:55:02.688141+00:00",{"id":131,"slug":132,"title":133,"created_at":134},"72b90667-d930-4cc9-8ced-aaa0f8968d44","pixelsmile-toward-fine-grained-facial-expression-editing-zh","PixelSmile：提升精細臉部表情編輯的新方法","2026-03-28T14:55:20.678181+00:00",{"id":136,"slug":137,"title":138,"created_at":139},"cf046742-efb2-4753-aef9-caed5da5e32e","adaptive-block-scaled-data-types-zh","IF4：神經網路量化的聰明選擇","2026-03-31T06:00:36.990273+00:00"]