Oracle Gece Arastirma — 2026-07-06

Curado por Mahsum Aktaş · Escaneo diario automatizado del sector de IA

Oracle Gece Arastirma — 2026-07-06

Otomatik derleme | v3 pipeline | 4.931 ham kaynak | 3.456 benzersiz item | 80 benzersiz kaynak | 5/5 kaynak ailesi

Gunun Ozeti

Bugunun ana ekseni agent ekosisteminin urunlesirken ayni anda risk yuzeyini buyutmesi: coding-agent tooling, MCP/skill sistemleri, AI destekli vulnerability kesfi ve agent enerji maliyeti ayni hatta toplandi. Kaynak: https://github.com/openai/codex-plugin-cc ve https://github.com/ChromeDevTools/chrome-devtools-mcp ve https://news.google.com/rss/articles/CBMirwFBVV95cUxPdExNb3J4dEotT1E1VUcxb3BlLU9id2laUHdtMTVkd0ZzWVZ3ek5DQ1BmcElOQnRDR0xVd1RJYjZjVHM3UG9JTlhVZ182aGQ2OEN1bkRYeFRWUlpxX25tblZ4R1BCdFhMeHZCNEQzM3BHUlNvLWgwNkdQSW9uUUZNR2RfbkJNVGxjUWJlMVZ3Wm9zZnNEQjl3OFZZOUZTc3VZWGN2U1p1WllwRG1aREMw?oc=5

Ikinci hat model ve altyapi dagilimi: ulusal/acik LLM girisimleri, optical inference, local browser LLM ve NVIDIA destekli egitim/enterprise model platformlari frontier API disinda alternatif dagitim modellerinin cogaldigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMilwFBVV95cUxQN0k3Zmx0OUpDTXhMQmwxUDhiaVhKU2JtXzQ5NTM5Ym44YUdfS1hmVEdQcjNtRG9QYWpvUlZ1N0ZYVnBnbzRwbTRkc2hqLURodTJSenhfTHJzbE5wajJ0ZkN3Slo1NmVnZ3gtYVVZV09ZVG8xb1kydzZOMHhUYVJzODlTQkFRRkY0SjROUGx5V01DdGlmTEFj?oc=5 ve https://news.google.com/rss/articles/CBMihgJBVV95cUxPSFduZXFkT0EtY1FnX1Z1dkZLMkp4OG03LVBPNFJqdUN3T2k0d3NRMFRJNU9HaWZldWVHZ3FvN0ZxS0ppVFNwWUN0dlljVzlhUF9rTkhaS3VSQzhqNUNqMXF1ZEhVTmo0LXhsWDRxQzlfbGdKSkpVeTNKUkVGWjJuckN2NzRLNGVQVC11bWZMSEdKdjN6TXRaeXVWOGFUV0VXMFZoNkhQNE8zdXZObWlpTTk2TVZ5RVBSa2ZZNW9PZ0lEazRWTGRWbXFWWXdaUE9vaGgwbFdSY0VQWGhQLThUVGYzTEVBbTBnazhmaGZ5S1JRQ0U1QktTZUVhcEh1R0VLYlQ4a2R3?oc=5

Ucuncu hat pratik operasyon: kaynaklar sosyal akista cok gurultulu, ama net sinyal su: skills, AGENTS.md, browser/MCP ve test-harness odakli agent ops artik ayri bir disiplin haline geliyor. Kaynak: https://github.com/agentskills/agentskills ve https://github.com/dotnet/skills ve https://role-confusion.github.io

Top 7

1. Amazon Mechanical Turk yeni musteri kabulunu durduruyor; human-labeling altyapisinda eski kalabalik is gucu modelinin kapanis sinyali. Kaynak: https://techcrunch.com/2026/07/05/amazon-will-stop-accepting-new-customers-for-mechanical-turk/
2. Baidu Unlimited OCR, uzun dokumanlari tek geciste isleyen memory-aware OCR yaklasimiyla dokuman AI hattinda yeni sinyal verdi. Kaynak: https://the-decoder.com/baidus-unlimited-ocr-processes-dozens-of-document-pages-in-one-pass-by-treating-memory-like-human-forgetting/
3. AWS S3 Annotations, object storage uzerine AI-generated metadata ve compliance context ekleyerek veri altyapisini agent-readable hale getiriyor. Kaynak: https://www.infoq.com/news/2026/07/aws-s3-annotations/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
4. OpenAI codex-plugin-cc, Claude Code icinden Codex'e review/delegasyon akisi acarak multi-agent coding interoperability hattini guclendiriyor. Kaynak: https://github.com/openai/codex-plugin-cc
5. Strix, open-source AI pentesting araci olarak agentic security otomasyonunun GitHub tarafindaki en net sinyallerinden biri. Kaynak: https://github.com/usestrix/strix
6. AI ajanlarinin ChatGPT tarzı sohbet botlarindan 136 kat fazla enerji tukettigi iddiasi, agent runtime maliyetini teknik degil operasyonel risk olarak gundeme tasiyor. Kaynak: https://www.donanimhaber.com/yapay-zeka-ajanlari-chatgpt-den-136-kat-fazla-enerji-tuketiyor--207709
7. AI ile vulnerability raporlarinin patlamasi, guvenlik ekipleri icin artik "bulgu sayisi" degil "onceliklendirme ve dogrulama" krizini one cikariyor. Kaynak: https://news.google.com/rss/articles/CBMirwFBVV95cUxPdExNb3J4dEotT1E1VUcxb3BlLU9id2laUHdtMTVkd0ZzWVZ3ek5DQ1BmcElOQnRDR0xVd1RJYjZjVHM3UG9JTlhVZ182aGQ2OEN1bkRYeFRWUlpxX25tblZ4R1BCdFhMeHZCNEQzM3BHUlNvLWgwNkdQSW9uUUZNR2RfbkJNVGxjUWJlMVZ3Wm9zZnNEQjl3OFZZOUZTc3VZWGN2U1p1WllwRG1aREMw?oc=5

TRENDS

• SPIKE: Claude 4, Flux, Inflection AI, Sora ve Suno trend penceresinde spike verdi; bugunku yeni aksiyon model duyurusundan cok agent maliyeti, agent tooling ve medya/model ekosistemi tarafinda. Kaynak: https://intelligence.exponentialview.co/
• YUKSELIS: GPT-5, Hugging Face, Databricks ve Midjourney basliklari yukseliyor; bugunku veri Hugging Face model ekosistemi ve Google News release akisi tarafindan domine edildi. Kaynak: https://huggingface.co/AliesTaha/fable-traces
• AZALIS: AI Regulation hacmi trend penceresinde dusuyor gorunse de bugunku Five Eyes, executive order ve vulnerability reporting akisi reguelasyonun guvenlik tarafina kaydigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiygFBVV95cUxNa2pMc0VsZmQ4V002ZWdNeEEyZkkxTWVFdGlpenBFdFRfbFNhR3REcFRLQmpxU0F5eERkcERfcUZwc1RKb0FYT2l4S0s3WWZlLU1nYmpraGpWV1JfMEZUbDNsVVVlS3FnLWp4X0o5NFVFNExmYV9nbC1vbmxNbXgwYlcwbGJyblU2Tk90NkN5MWVYZERybFdiYUlVY1JSdlNHUklkMDdYcDZLS01MRjd6aUJiejF3dXhkOGN0Q0dkSlhwNmRsR2tialJ3?oc=5

KAT-1 — Endustri / Sirketler

• Higharc'in 90M dolar Series C turu, AI'nin sadece yazilimda degil konut/insaat workflow'larinda da sermaye cekmeye devam ettigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMimgFBVV95cUxQN1psdWtnamlUVWlqSC05X1BBU1gya3NCdVZncGs3UV9Ma1l1RXE5M3BQTjlBV1RfZXR1VkR5RWFpeE1DRmYwNXlocWtrRW9lVzhTaG8zNVhhMFdKc1JOVkdmSVdJblk1VzAyWkl1YnNCZ1ZVNkpjRVo2bThXZnFKQkpGckdmMlppTjJMQkhfYmFwdmk1OUFqdW5n?oc=5
• Venice.ai'nin 65M dolar Series A ile unicorn olmasi, privacy-first AI narratifinin hala yatirimci ilgisi cekebildigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiwAFBVV95cUxNX3FjbE45Sjc1bXdFQS1XaW9ucnpTb2lVRFRha0sxSTNhU2dNQTJIU295eVlkVEJ5eWQ4UUctV1hBOEZsVWFXRXJVVEt0Rll2clVoaXhsZk9pUElVT2RXN3B5cmtpMXBPbnpXdWR2Q240eVpIWC1HeE9TU0I0QWJtNlN6R1J3bkU2MGxrWWcyQzEyRzVRbzIwYTVXQVNLRzZySGhNWFdnSnh2SU8tXzhhbjlCOGlLT3psdjVEUGJ1X2w?oc=5
• AI inference software startup Gimlet Labs'in 80M dolar Series A turu, compute cost optimizasyonunun ayri bir pazar haline geldigini dogruluyor. Kaynak: https://news.google.com/rss/articles/CBMiqAFBVV95cUxNVS14cVhYVXB6V2N3QzVpbE5pc0lGYzNfT2JiQ0dwOTNqYVhhNml2WVpuWUFtaGhBazdYOEJlMkNEWFFWblI1NmlzLTZQTlJzM0xSY3gtUExacjE5YUNobE5IQnRScUhDRnhiSXRoLWh3a0h5REkwQ1FEemN3NW1xMWlfQzJHNkUzSUJmSFFha1oySmRvZjBBYjE5LUdBcGdZdmtKN2RfRi0?oc=5
• Perceptic'in drug discovery otomasyonu icin seed turu, AI-science tarafindaki sermayenin eski biotech yazilimindan agentic discovery araclarina kaydigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMi-gFBVV95cUxNSXMzdW42QXZBY2xuTEo0RnVvLTZHZ0VKSy14eU5wdWNOM3ltZXc1QjN3MDY3bFBLYVBHUVdjRTVScGlxQk1KRzl0NDlReVB5TTh3U211cWVWbllhNEszRk81bDgtdDZWeVA2bllzN3YyTUYwRzJYdUtGckoyb3J6RXpiT0FNZlRhM25uNU5SNHYxRVBsNDh2Y3AxVmxoZHBMTmpVd2lrbjUtZlQ4VW1LZWNydUROQVh6THIwb25BSU9mTzJERHlJNmd1eXZCaHI3YXQzd2dTQU1QNEhJcFlSOWsyX1JfVkxPdXNocFg2aXNET2RCZWRQRGV3?oc=5

KAT-2 — Model / Release

• Switzerland open-source AI model akisi, ulusal LLM projelerinin ABD frontier lab'larina alternatif olarak konumlandigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMilwFBVV95cUxQN0k3Zmx0OUpDTXhMQmwxUDhiaVhKU2JtXzQ5NTM5Ym44YUdfS1hmVEdQcjNtRG9QYWpvUlZ1N0ZYVnBnbzRwbTRkc2hqLURodTJSenhfTHJzbE5wajJ0ZkN3Slo1NmVnZ3gtYVVZV09ZVG8xb1kydzZOMHhUYVJzODlTQkFRRkY0SjROUGx5V01DdGlmTEFj?oc=5
• Thinking Machines Tinker, distributed LLM fine-tuning icin dusuk seviye API sunarak "managed but not hidden" egitim altyapisi sinyali verdi. Kaynak: https://news.google.com/rss/articles/CBMi-wFBVV95cUxNZG5qWGN0WjFsalJROHFVT1QxMF9xZ3pMVk5kS05kTzlsall1NG1Kc2ZSbTNwX1lnNDJnZFNfdi1GY3BtTkloYUJWRXpFa3M3b0k5NG5idDFhSjFyOHRvWnBVNXg2WUNScFV0dW5VdUtyVEJOWlNvSzNVd2FFWUw3Qzc4Vi1YOWZuLXEtZC1oMEhjYVZNalhwUElkb3JzVXlsZDYwR2tJZHFvcjFlUHk4dllISjZiM3pSWjBpS0lteTFrczM1VFVmenpYQjFWMGRuNkdmb0FDMzVNTDRDekU1MmJ6bkZOdG91Y3lhTUEwOG1oQWpsbUZuUHFoNA?oc=5
• Zoho Zia LLM, prebuilt agents, custom agent builder, MCP ve marketplace ile enterprise AI platformlarini tek pakete toplamaya calisiyor. Kaynak: https://news.google.com/rss/articles/CBMi-AFBVV95cUxNZlBFYmc4bzhLTTNxZ2JGenF2ODhfU3pqcDZZUUtrMlZHRjVGV2VwSTdIQU4wR1FkMnA3R0FqRkhqYVhzMDdNNjBHcU1lVVVqUi1lbV9KUzN5cUpadWVWaE9aaEhpcDVsamt1TElqbVpEM3EzOER4WGdkdlVXWm83SXZvWmhCSzJWVVNfOXozVmRfLVVibFNGcm5Ic0Z5WTRqQlFBRDRvemZyVG0ycnIydm1vX28wMG5EZ0tEMV90YkF1UFRhMUJnU2tGOUxTQ3R0eGdMaEdVSloyYlBaTWR3M3YwWWJCVEVjbzFVT0xoSm1KbklkeDhLMw?oc=5
• Sigma'nin local LLM'li privacy-focused AI browser lansmani, browser runtime'in agent ve local inference icin yeni cephe oldugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMimwFBVV95cUxOd0RnaS13NmpFNU40d1VGSU02ZjBROEk4cXhHVWpTREpZTUVZVHlmRU45MjZiNkNGcm96ek9UYlMtS1NpNkE3cmFoZzJOcUIxeENyMXVmSFZRSlBRWTVXd1c4X0ViS1JvQnozdVRRQVJ3TE1YRnE3ZHpTa3BZNG9zaVlRdXNBdXBBNFd3eFZXWEVnM0JobGw3YzZZSQ?oc=5

KAT-3 — Guvenlik

• Prompt Injection as Role Confusion, prompt injection'i sadece string sanitization degil rol/otorite karisikligi problemi olarak cerceveliyor. Kaynak: https://role-confusion.github.io
• The Register'in MFA-optional banks haberi, AI caginda bile temel kimlik guvenligi eksiklerinin ana saldiri yuzeyi oldugunu hatirlatiyor. Kaynak: https://www.theregister.com/security/2026/07/05/mfa-optional-banks-leave-safe-doors-and-accounts-wide-open-for-thieves-to-pillage/5266161
• MCP'nin kritik supply chain yuzeyi olarak isaretlenmesi, tool metadata ve server discovery katmaninin agent guvenliginde birincil risk oldugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiuwFBVV95cUxQLTFFczJfemlJZG1TR2tiM3RCNElHRnc2RThEWEl1M3lITzkzLUNkVzI3aGhYZklBWnBER1NSRzZWbnBLNkd4QlBDWmdHaWxSdUN3RzFnMWxVQ3lON2dGMEx4Z1JLaTFpYS1jNkJtQTktQzFIY1JHeWtLWGZtWVB6NVJMbUpnZ0p2WS10UGwtRF83bUJNT1pfSnZzT2ctZGxqaG5wanljUTlyNUJ1WkxSdFlNaXJiemtYMnZR?oc=5
• Copilot vulnerability basligi, e-posta, 2FA kodu ve hassas veriye uzanan AI-tooling riskinin enterprise ortamda dogrudan veri sizintisina donebilecegini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMikgFBVV95cUxNZkR6WGE0R0FBbV9IM2hfZGJxRTJFQ3YzZ19SNVRfRloyTWZ2MzY3QllrRE1Ic1U2TzNGMngxdVRGYUtCd1c2d25QdzFMd21rZWNOMFdkXzAyT0gyR19hQnlST2Nueml5M0xhYlhuNEJaWlVMSmpOazBabFdjb2luTjBtcl9UME9RdkI5OWRFZDMtdw?oc=5

KAT-4 — Arastirma / Paper

• Matrix Orthogonalization'in recurrent model memory uzerindeki etkisi, long-context/discrete memory disinda eski mimari ailelerinde de kapasite kazanimi arandigini gosteriyor. Kaynak: https://ayushtambde.com/blog/matrix-orthogonalization-improves-memory-in-recurrent-models/
• Transformer ve hybrid modelleri token seviyesinde karsilastiran calisma, mimari seciminde benchmark skorundan cok davranis profili okumayi one cikariyor. Kaynak: https://arxiv.org/pdf/2606.20936
• Dartmouth AI tutor sonucunda 0.71-1.30 SD etki buyuklugu iddiasi, egitim AI'inda "yardimci chatbot"tan olculebilir pedagogik etkiye gecis sinyali. Kaynak: https://intextbooks.science.uu.nl/workshop2026/files/itb26_s1s2.pdf
• Google AMIE medical AI arastirmasi, klinik karar destek sistemlerinde model benchmark'inin hasta yonetimi simulasyonuna kaydigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMisgFBVV95cUxPM1l6b1JxaHFNLUI0R2wyVlVwM041aW1QUXJkMTZjWmcxYUlRME9DNDZfLVVBcGtWdDVXVEtPbDFXcE90U0lSOE85UE9VcVkzU3ppRTUzaEdUUFlDTDVoZ1dqWjlGbzE5Rlh5QlNuNmstWl9mdkh6QS05TGZMcUJMTEhiVU42YTBvQWJTTkpsYkVmZlEwM0RzQklURFJVdjR6N2Z1YXVvSGxtYjNIUTFKWUNR?oc=5

KAT-5 — Reguelasyon / Politika

• "Promoting Advanced Artificial Intelligence Innovation and Security" executive order basligi, ABD hattinda AI innovation ve security'nin ayni politika metninde paketlendigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiygFBVV95cUxNa2pMc0VsZmQ4V002ZWdNeEEyZkkxTWVFdGlpenBFdFRfbFNhR3REcFRLQmpxU0F5eERkcERfcUZwc1RKb0FYT2l4S0s3WWZlLU1nYmpraGpWV1JfMEZUbDNsVVVlS3FnLWp4X0o5NFVFNExmYV9nbC1vbmxNbXgwYlcwbGJyblU2Tk90NkN5MWVYZERybFdiYUlVY1JSdlNHUklkMDdYcDZLS01MRjd6aUJiejF3dXhkOGN0Q0dkSlhwNmRsR2tialJ3?oc=5
• Five Eyes uyarisi, model kabiliyetlerinin devlet ve is dunyasina yikici saldiri kapasitesine yaklastigi iddiasini politika seviyesine tasiyor. Kaynak: https://news.google.com/rss/articles/CBMiwgFBVV95cUxNTklxTWJwcms1UlBCeUYyTGtNTGFtdEtGcmR3bEJJWHY2VHRpbUJqR3lUdW13RkF6MWJ4SklkWEtEYUlOLURENnRDNWNjN1dCclZVVGNOQVV0OEg5U0gyY2pGWDVDY3JUR3lKelJGSHdxUUM1UXY2UXFpa09idjN1Z0ZLNFZWX2dmR1FyaDRrM2NVY2trNFJwbjJHdWZuYUUtck1rWU9VRWZEbV9kT2hXOUNxbm1iU3FwTlp3dzkzdUhQdw?oc=5
• AI security ve vulnerability reporting icin tasari haberi, AI ile bulunan guvenlik aciklarinin raporlama rejimini yasama gundemine sokuyor. Kaynak: https://news.google.com/rss/articles/CBMizgFBVV95cUxOTklYeEJISGUwN1dnam5NREtTVlRNRllIZU5LcEs3RllyRnNUY0RuSnhqekV1TDZkMHg2bUdHVi1UQmhkb09Rb2tLaC1LcHNHTWFWaWVsSUppWEZUWGRmeWlqU1o2YXBoMUM0MGdBSXdfRy1FdnIwZGVaQzdMa29CbWNvbGYxNE9qcllubXpxUDdBQjNqOUxWbzE5OTI3S2RoMnY1YTJuSmxVUkpGYWthZFBqT2hNXzZKczBjaTlweFhaWnpmRExfM0Myb3hiQQ?oc=5

KAT-6 — Core AI

• AI search agents haberinde sorun "arama yapamamak" degil, belirsiz sorguda dogru soruyu soramamak olarak tarif ediliyor; bu agent UX icin kritik. Kaynak: https://the-decoder.com/ai-search-agents-dont-fail-at-searching-they-fail-at-asking-the-right-questions-when-queries-get-ambiguous/
• LlamaIndex legal-kb, agentic retrieval'i retrieve/find/read/grep gibi dosya sistemi benzeri araclarla sunarak RAG'i daha denetlenebilir hale getiriyor. Kaynak: https://www.marktechpost.com/2026/07/05/llamaindex-legal-kb-agentic-retrieval-over-index-v2-with-retrieve-find-read-and-grep-tools/
• Competence Gate, kucuk modelin sozlu ozguveninden degil internal confidence sinyalinden tool-use karari vermeyi deniyor. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1unw5un/competence_gate_gating_tooluse_on_a_small_models/
• Mistral CEO Arthur Mensch'in proprietary model uyarisi, model saglayicisinin business process gorunurlugu kazanmasini stratejik risk olarak isaretliyor. Kaynak: https://the-decoder.com/mistral-ceo-mensch-says-proprietary-ai-models-give-labs-a-front-row-seat-to-your-business-processes/

KAT-7 — Engineering / Tooling

• Chrome DevTools MCP, browser introspection'u coding agent'lar icin standart tool katmanina ceviriyor. Kaynak: https://github.com/ChromeDevTools/chrome-devtools-mcp
• openai/codex-plugin-cc, agent'lar arasi review/delegasyon paternini plugin seviyesine indiriyor. Kaynak: https://github.com/openai/codex-plugin-cc
• agentskills/agentskills, skill formatini agent capability dagitimi icin standartlastirma denemesi olarak one cikiyor. Kaynak: https://github.com/agentskills/agentskills
• TIRx compiler stack, frontier ML kernel evrimi icin compiler seviyesinde acik kaynak altyapi sinyali veriyor. Kaynak: https://tvm.apache.org/2026/06/22/tirx

KAT-8 — Community

• AI Agents Enable Adaptive Computer Worms tartismasi, agent otonomisi ile kendini uyarlayan saldiri yuzeyi arasindaki bagin toplulukta ciddiye alindigini gosteriyor. Kaynak: https://cleverhans.io/worm.html
• "What does Safe AI look like?" tartismasi, open-weight modellerde post-release fine-tuning ile safety'nin nasil korunacagi sorusunu one cikariyor. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1um9bs7/what_does_safe_ai_look_like_d/
• arXiv'in Cornell'den ayrilip bagimsiz nonprofit olmasi, AI arastirma altyapisinda kurumsal yonetisim degisikligi olarak takip edilmeli. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1ukjtlm/on_july_1_2026_arxiv_will_spin_out_from_cornell/
• Yerel/uncensored model tartismasi, kullanici tarafinda privacy, kontrol ve policy friction nedeniyle local modele gecisin community narratifi haline geldigini gosteriyor. Kaynak: https://www.reddit.com/r/artificial/comments/1uocn4j/why_are_more_and_more_people_switching_to/

KAT-9 — Meta / Piyasa

• Exponential View raporu, generative AI'nin son yilda 110B dolar satis yarattigi iddiasiyla pazar buyuklugu narratifini besliyor. Kaynak: https://intelligence.exponentialview.co/
• Simon Willison'in inference margin tartismasi, AI sirketlerinin zarar edip etmedigi sorusundan cok birim inference marjininin bilinmesi gerektigini soyluyor. Kaynak: https://bsky.app/profile/simonwillison.net/post/3mpw5mmxya22q
• Global startup investment ve AI boom haberleri, AI sermaye akisini hala genis piyasa anlatilarinin merkezinde tutuyor. Kaynak: https://news.google.com/rss/articles/CBMiqwFBVV95cUxQZ2V6SXlCUVdJVWV2NjNlMkFHZVFlSzlLa1EzZ3Q2cGtkeHRFTWNkVGJXQ0J5UDJaOHZRSDhKVzdKdEN6UVI2Z19GenhHR1ZhNmVLZFItQjNWNnFybGloaGVkRmlvQU1PZm1tSEFDQ3V1Q3ZWZHU5SzBYVlUyY05peElqRW5XRUxLVUJnNllNTFpJSTBNajBwV1hlTVRTZzN2cFF3NWJEcHAwQm8?oc=5

SECURITY

• AI-driven vulnerability discovery 2026 CVE hacmini 66.000'e itebilir iddiasi, savunma ekipleri icin otomatik triage ve exploitability scoring ihtiyacini buyutuyor. Kaynak: https://news.google.com/rss/articles/CBMidkFVX3lxTE42eHpidGVVcjR3XzBOYTNkaGNyNDR6M3hUaG5RWUNEREpURHRrUmlIaHRlZTRjanRqNzVDelJ0OG9MS0VOWTBzREpNMVhia2JKcmtkZkY0ZjZFZ240UmtmZjdDbW5scHZ6Wl8zazYzcnV6RHJUUWc?oc=5
• OpenSSL'in AI ile bulunan high-severity vulnerability patch'i, AI-assisted security'nin gercek patch akisini etkilemeye basladigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMikAFBVV95cUxOQTMwX3pWQ2t3Y1Fpd3hpak5uX0JJLWNrVEtjY1JiRnJqWG0yYVpWbzM1VVlac3hJOGFHS2ZvZUNMYmVWblU5RnpEOHRMNmEwc1FjeEZ1TmxfUU8xUUN0aE1fRzhsRlMwTDhDNjE0YjN2ckNSQVhpMjA4UHR6NlplczdHSlU4dTQxeWFCVHVnMi3SAZYBQVVfeXFMT2xkcUY2ZGtMczc2cjhxVzFnUHlrcGp6cE1CT0NtUTZlUWlrbE13MW8wczRUMlg0dVBBR0JTRHE1T1NQZXNuOUtGd2FJcmJCakthc3pDNTB1and2OTNZYjVHeFBpeUNDWHB2S09JSzFScjFkQ2V0QnNJWGtQT1kyUkVjWVZ4amtXNFdzUE5lS1VJMUdPVDF3?oc=5
• OpenAI Daybreak haberleri otomatik vulnerability detection ve patch validation hattini urunlestirme yarisi olarak izlenmeli. Kaynak: https://news.google.com/rss/articles/CBMihAFBVV95cUxQalhKa28zNGFmVW0tVVRPTjlGVFJfbmJoMFZ1VVpIWHljTWR1S1lHbzVnSEstNklfRUE2dTBkXy1kRTU5cmN5MkFLN0NTakZvSWVaR28yV00zcjQ0dG4yYWM0dENVOG9hSlVzckVwaW5JeXFvU2U4ZkZKdjI5aTg1VUhhN0I?oc=5

REGULATION

• Global AI governance bugunku veride daha cok security reporting, vulnerability coordination ve privacy-by-design basliklarina indi. Kaynak: https://news.google.com/rss/articles/CBMi-AFBVV95cUxPYTZXMTNsTy1rZWdCQWlWNmpQRzJ3TlpzS3VzSnFUYU5YaDJySkFCNzNQSVZhNGFWNWM4bi1xZVZSWGtpUDQzS3R4SUhSWHdmdjI1bXdrUE9JSTJsYkUzNHFOT2hoNDlqeXlHeWhpYkhrTGpCaG5yNW9LQUw3Q0FRLUZCRENaY0NGeUxIVFZ1YmNNWkNGUGgza1ZzaUJCbVllbkt2Z1hJdy01RVJmZl9mQW4ySVdxUmRFVXlFcDRpT0RxNWd2ZmNDWEViMVlJNkZnQmpoTG02aGg1U0dxUDQ4eU1qZnlzbjhTUTktYXdtUUlneUo5bjFsdQ?oc=5
• AI governance mistakes haberi, kurumlarin AI policy yazmasindan cok operasyonel kontrol ve audit trail kurmakta zorlandigini isaret ediyor. Kaynak: https://news.google.com/rss/articles/CBMiswFBVV95cUxObmowQXVsSmtYZUZuMkhVdWNTVXcyWjR5ZTExeUpiWHdkbHZrcWpESUZYSDdqWjRWR0hHV21iUy1mcjdVdkpwSEZoVDVSUU0tdW5qclN5ZzBGVTFDQ3NFU3A4M2xpWHVweUN4cTZ0Vy11U0NEaHhCQXZKX3htQjBpU2ZSTDRwV3hGdDJ6ZkxkUVNUTEZ1Qk80MWoyMUdKR0xCcENaWG1vU1lHbm5rTHJKUnE1NA?oc=5

AI-SCIENCE

• Takeda'nin Insilico Medicine ile AI drug discovery hamlesi, pharma tarafinda modelden cok discovery pipeline ortakliginin degerli oldugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMilAFBVV95cUxQaFI3TEZkWkliQzVnVE5uY19GMUNHTkpneUs2RjQybkFJdzloWjU4dFp6MG44Tjk2TUFaak1DNXVKNkRMejE5OXhWajFGa0RQakd4UWRfRkhDczRQVVh5dTBDVXpTb0FOVjZfVDF2dXJ0eWtNekQ0RmNMTkUwTTd1Q3REN1VTTjJnZ3V5Q19sa21mMDR2?oc=5
• Machine learning ile iki yeni superconductor kesfi haberi, materials discovery hattinda AI'nin arama uzayini daraltma rolunu one cikariyor. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4acd242d5d4296814af6209849f2d7&url=https%3a%2f%2fwww.electronicsforu.com%2fnews%2fmachine-learning-discovers-two-new-superconductors&c=10999773609423138788&mkt=tr-tr
• Medical LLM ve KMed.ai akisi, ulusal dil/model stratejisinin saglik domainine ozellesmis modellerle genisledigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMilAFBVV95cUxNTHI3V05tM05vUzlXYXdVMjZnUHQ2ZWJCNnlEWi1yOXpKZnFxMHp6eFFnY1hPOHlKdGpRWG42UFR6ZzVpenBoTmFyMml2NWg2bUNrMFFLd3g0SGpJOXExS2ZOYnNxTDhFUXdFaVVXa0prTmRIZEY0S3h2a1Jvc2dLaHk5SmNJMkQyeFc3Rm5mVmVNOGt1?oc=5

INFRA

• Lumai'nin optical computing inference sistemi, GPU disi inference donanimi anlatilarinin somut urun duyurularina donustugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMihgJBVV95cUxPSFduZXFkT0EtY1FnX1Z1dkZLMkp4OG03LVBPNFJqdUN3T2k0d3NRMFRJNU9HaWZldWVHZ3FvN0ZxS0ppVFNwWUN0dlljVzlhUF9rTkhaS3VSQzhqNUNqMXF1ZEhVTmo0LXhsWDRxQzlfbGdKSkpVeTNKUkVGWjJuckN2NzRLNGVQVC11bWZMSEdKdjN6TXRaeXVWOGFUV0VXMFZoNkhQNE8zdXZObWlpTTk2TVZ5RVBSa2ZZNW9PZ0lEazRWTGRWbXFWWXdaUE9vaGgwbFdSY0VQWGhQLThUVGYzTEVBbTBnazhmaGZ5S1JRQ0U1QktTZUVhcEh1R0VLYlQ4a2R3?oc=5
• MAX modellerinin Apple silicon GPU'larda calismasi, local/on-device inference icin Apple GPU yolunun olgunlastigini gosteriyor. Kaynak: https://forum.modular.com/t/max-models-can-now-run-on-apple-silicon-gpus/3283
• Hon Hai'nin AI talebiyle satis artisi, AI infra tedarik zincirinin sadece NVIDIA degil tedarikci agi tarafinda da genisledigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMimwFBVV95cUxPd29lOURmS3NfYmprZGFsQU9udlJ0eVFYWlJFX3p5QzR1c0FTSDh2VWt1alpVcDZUUXFOYUZpX2FCd2JUQnRWeUFkU2g4d2daTGJ0RkhheWJuUFhkQURyMGJvclNVVDYtTG9nRzlDWllHQ0x3WkRuVy1BUzk1SnhaWnRndGVQVGdtZXY1LTFYWDZqMGlBdlJaVVFUWQ?oc=5

SAFETY

• Robust AI Security and Alignment makalesi, alignment ve security'nin ayri kutular olmadigini; robustluk probleminin hem model hem sistem katmaninda cozulmesi gerektigini hatirlatiyor. Kaynak: https://ieeexplore.ieee.org/document/11475847/
• Safe AI tartismasi, open-weight modellerde post-release finetune ile refusal/safety zayiflatma riskinin pratik savunma sorusuna donustugunu gosteriyor. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1um9bs7/what_does_safe_ai_look_like_d/
• Anthropic Responsible Scaling Policy guncellemesi onceki gunlerde raporlandi; bugun yeni safety sinyali daha cok topluluk ve policy tartismasinda devam etti. Kaynak: https://www.anthropic.com/news/announcing-our-updated-responsible-scaling-policy

WATCHLIST

• Codex/Claude Code interoperability: codex-plugin-cc ve sosyal akistaki Codex iyilestirme basliklari multi-agent coding stack'inin hizlandigini gosteriyor. Kaynak: https://github.com/openai/codex-plugin-cc ve https://news.google.com/rss/articles/CBMiWEFVX3lxTFBhT1pfeU81cGt2Um1ZNEZrZGx6a3lGR3gwZ1ZnbzdpQzB3NkhDbkJ5UTl4UFRNd0Z5akl0em1TN3pPNlptdmtmV0RuelBMdWxtamw1a0ljNlI?oc=5
• MCP riskleri: Chrome DevTools MCP urunlesirken MCP supply-chain uyarilari ayni anda geldi; entegrasyonlar security review olmadan genisletilmemeli. Kaynak: https://github.com/ChromeDevTools/chrome-devtools-mcp ve https://news.google.com/rss/articles/CBMiuwFBVV95cUxQLTFFczJfemlJZG1TR2tiM3RCNElHRnc2RThEWEl1M3lITzkzLUNkVzI3aGhYZklBWnBER1NSRzZWbnBLNkd4QlBDWmdHaWxSdUN3RzFnMWxVQ3lON2dGMEx4Z1JLaTFpYS1jNkJtQTktQzFIY1JHeWtLWGZtWVB6NVJMbUpnZ0p2WS10UGwtRF83bUJNT1pfSnZzT2ctZGxqaG5wanljUTlyNUJ1WkxSdFlNaXJiemtYMnZR?oc=5
• Hugging Face / open model ekosistemi: trending_models ve sosyal akista model/skill indirme hacmi acik model dagitim kanalinin gucunu korudugunu gosteriyor. Kaynak: https://huggingface.co/AliesTaha/fable-traces ve https://news.google.com/rss/articles/CBMiZ0FVX3lxTE5yd2w5elBXNkNocW1HX2FSc2xoVHN6dWVqU0tfaHpQVTFIVzRPTkRnZW5vOWJzSHFyLUotSHRJbzBQeWRTU1dYNVg1U2VhRWQxY1VOenRuZk9zN3pBaEFYdmpETjdXbnM?oc=5

CikCik Paketi

• Simon Willison, AI ekonomisini anlamak icin toplam zarar yerine inference'in marjla satilip satilmadiginin bilinmesi gerektigini vurguladi. Kaynak: https://bsky.app/profile/simonwillison.net/post/3mpw5mmxya22q
• Simon Willison, Claude Fable ile sqlite-utils release review'unda bes release blocker buldurdugunu yazdi; bu, final-review agent pattern'i icin guclu pratik sinyal. Kaynak: https://simonwillison.net/2026/Jul/5/sqlite-utils-fable/
• Omar Saravia akisi, AGENTS.md dosyalarina ne koyulduguna dikkat edilmesi gerektigini soyluyor; agent instruction injection artik repo hijyeninin parcasi. Kaynak: https://news.google.com/rss/articles/CBMiXkFVX3lxTE5OZm9CdGFTV0dCeUt1TFB2aHFnZEJ2ZEl5M2ZmOVRHekdGay1NbEVqSkZGa1Q4NXVSV0ZoQlg3LXduLXNxNkVITUd2Z1dNQkRjeV93QkNwcUMzdDMwQXc?oc=5
• Philipp Schmid akisi, MCP'nin guclu ama context-window bloat riski yarattigini isaret ediyor. Kaynak: https://news.google.com/rss/articles/CBMiY0FVX3lxTE5ZTlVPM2g4N2dpbHhhaHp4ejI1NlpxaXIxYXAtQVlibTBqZFV5RnVNN2Jkai1IMlZ6X0Q3bDdYcWNOZEhCUHRraGFaTVQ1SHotT3dZcVFMNHpleG00ZGFLTnBISQ?oc=5
• Thomas Wolf akisi LeRobot/VLA-JEPA tarafinda robotics tooling'in acik kaynak ve kolay kurulum ekseninde ilerledigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiW0FVX3lxTE04TTNJcS1KdTdWS0J1THQ0SkVBYnFYSk9OeEtPZWNXU0ZsLW5LV3lxQ0tvX2lXOWZXOWtMNVJKX2ZibzlSWXJPZDgtTWpod25LbzFHZkludTI4c0E?oc=5

Oracle Self-Improvement Sinyalleri

• Final-review agent pattern'i production release oncesi zorunlu kalite kapisi olarak degerlendirilmeli; Simon Willison'in sqlite-utils ornegi dogrudan uygulanabilir. Kaynak: https://simonwillison.net/2026/Jul/5/sqlite-utils-fable/
• Agent skill standartlari, OpenClaw skill sisteminde metadata, scope ve rollback bilgisinin daha siki tutulmasi gerektigini dogruluyor. Kaynak: https://github.com/agentskills/agentskills
• Chrome DevTools MCP ve browser automation akisi, frontend/browser-heavy dogrulamada CDP/MCP screenshot ve a11y snapshot kapilarini daha sik kullanma sinyali veriyor. Kaynak: https://github.com/ChromeDevTools/chrome-devtools-mcp
• Role-confusion prompt injection arastirmasi, sistem talimatlarini sadece text policy degil authority-routing problemi olarak test etmeyi gerektiriyor. Kaynak: https://role-confusion.github.io
• AWS S3 Annotations, ham dosya depolamadan context-rich object store'a gecisin memory/wiki pipeline icin de model alinabilecegini gosteriyor. Kaynak: https://www.infoq.com/news/2026/07/aws-s3-annotations/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering

Coverage / Blind Spots

• Coverage tamam: rss/news, search, community, social ve academic/api ailelerinin tamami veri verdi; toplam 4.931 ham itemden 3.456 benzersiz item cikti.
• Baskin kaynak ailesi social oldu: 4.091 ham item ve 2.994 unique item; bu hacim faydali ama Google News tabanli Twitter fallback gürültüsü cok yuksek.
• Academic/API sadece 70 ham item ve 1 unique item verdi; arXiv/Semantic Scholar tarafinda 429 ve bos akışlar nedeniyle akademik derinlik zayif kaldi.
• Search ailesi 518 ham item ile security/release/company sinyallerini tasidi; ancak canonical URL cozumleme eksigi nedeniyle Google News RSS linkleri raporda fazla yer kapladi.

Kaynak Ozeti

• RSS/news: 184 ham, 71 unique; DonanimHaber, Technopat, 0xSMW AI Research ve The Decoder ana tasiyicilar.
• Search: 518 ham, 306 unique; Google News AI/releases/security ana kaynak oldu.
• Community: 68 ham, 84 unique; GitHub Trending, Lobsters ve Reddit MachineLearning pratik sinyal verdi.
• Social: 4.091 ham, 2.994 unique; X fallback cok hacimli ama tekrar/kalite filtresi gerektiriyor.
• Academic/API: 70 ham, 1 unique; Hugging Face tarafinda veri geldi, Semantic Scholar ve PapersWithCode zayif kaldi.

Dedupe & Kalite Notu

Bu rapordaki maddeler onceki 3 gun raporlarina gore tekrar etmemek icin filtrelendi; Alibaba Claude Code yasagi, Claude Science Beta ve Leanstral gibi son 3 gunde islenen ana haberler bilincli olarak tekrar edilmedi.

• Merge sonucu: 5 collector dosyasi, 4.931 giris, 4.913 cikis, 18 URL tekrari atildi.
• Dedupe sonucu: 4.913 giris, 3.456 yeni item, 1.455 fingerprint tekrari, 2 semantik tekrar.
• Trend update: 51 topic, pencere 2026-06-30 -> 2026-07-06.