Oracle Gece Arastirma - 2026-07-08

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

Oracle Gece Arastirma - 2026-07-08

Otomatik derleme | v3 pipeline | 9015 ham kaynak | 3046 benzersiz item | 106 benzersiz kaynak | 5/5 kaynak ailesi

Gunun Ozeti

Bugunun ana ekseni agent'larin urun ozelliginden cikup yonetilen runtime, guvenlik ve kurumsal operasyon katmanina yerlesmesi. Google Gemini API Managed Agents arka plan gorevleri ve remote MCP ile bu hattin platform tarafini guclendirirken, GitHub AI agent repo sizintisi ayni hattin guvenlik borcunu gosteriyor. Kaynaklar: https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api/ | https://www.theregister.com/security/2026/07/07/github-ai-agent-leaks-private-repos-when-asked-nicely/5267924

Ikinci eksen maliyet ve egemenlik: Microsoft Copilot tarafinda OpenAI/Anthropic bagimliligini azaltip kendi modellerine yaslanma sinyali verdi, DeepSeek ise ABD ihracat kontrolleri altinda kendi AI chip hattini gelistiriyor. Kaynaklar: https://techcrunch.com/2026/07/07/microsoft-joins-ai-cost-cutting-trend-by-relying-more-on-its-own-models/ | https://arstechnica.com/ai/2026/07/facing-us-export-controls-chinas-deepseek-plans-to-make-its-own-chips/

Ucuncu eksen safety ve moderation: Discord'un AI moderation hatasi, Meta Muse Image'in Instagram kimliklerini prompt olarak kullanmasi ve yeni agent-memory saldiri paper'lari, guvenilirlik probleminin hem urun hem research tarafinda buyudugunu gosteriyor. Kaynaklar: https://techcrunch.com/2026/07/07/discord-admits-ai-moderation-bug-wrongfully-banned-users-over-harmless-images/ | https://www.theverge.com/tech/962485/meta-muse-image-ai-model-instagram | https://arxiv.org/abs/2607.05029

TRENDS

• Trend penceresinde "AI Safety" ve "Autonomous" spike verdi; bugunku somut kanitlar agent safety testleri, otonom kara araclari ve AI moderation kazalari. Kaynaklar: https://huggingface.co/papers/2607.01793 | https://techcrunch.com/2026/07/07/the-first-american-autonomous-ground-vehicles-are-fighting-in-ukraine/ | https://techcrunch.com/2026/07/07/discord-admits-ai-moderation-bug-wrongfully-banned-users-over-harmless-images/
• "Models" ve "tooling" hacmi yuksek, fakat karar degeri daha cok runtime maliyeti, model egemenligi ve managed-agent API'lerinde. Kaynaklar: https://techcrunch.com/2026/07/07/microsoft-joins-ai-cost-cutting-trend-by-relying-more-on-its-own-models/ | https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api/
• "Regulation" toplam hacim dusus gosteriyor ama sinyal sert: Cin model ihrac kontrolu, Illinois AI safety yasasi ve UK tech sovereignty ayni gun geldi. Kaynaklar: https://the-decoder.com/china-eyes-export-curbs-on-its-top-ai-models-and-europe-is-caught-in-the-middle/ | http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4d701c7e694b818510a9f8d56302cd&url=https%3a%2f%2fwww.yahoo.com%2fnews%2fvideos%2fillinois-adopts-artificial-intelligence-safety-160602744.html&c=2816284873912463041&mkt=tr-tr | https://www.theregister.com/ai-and-ml/2026/07/07/mps-tell-brit-government-sort-out-your-tech-sovereignty-or-get-left-out-in-the-cold/5267076

Top 7

1. Google, Gemini API Managed Agents'i background tasks ve remote MCP ile genisletti; bu, agent runtime'in platform-native hale geldigini gosteriyor. Kaynak: https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api/
2. GitHub AI agent'in ozel repolari sizdirabildigi "GitLost" akisi, agent permission modelinin klasik CI/CD guvenliginden daha kirilgan oldugunu gosterdi. Kaynak: https://www.theregister.com/security/2026/07/07/github-ai-agent-leaks-private-repos-when-asked-nicely/5267924
3. Microsoft, AI maliyet baskisi nedeniyle Copilot'ta kendi modellerine daha fazla yaslanma hattina giriyor; frontier API arbitraji kurumsal maliyet stratejisine donusuyor. Kaynak: https://techcrunch.com/2026/07/07/microsoft-joins-ai-cost-cutting-trend-by-relying-more-on-its-own-models/
4. Meta Muse Image, Instagram hesaplarini gorsel prompt olarak kullanabilen model sinyali verdi; creator identity ve consent riski yukseliyor. Kaynak: https://www.theverge.com/tech/962485/meta-muse-image-ai-model-instagram
5. DeepSeek'in kendi AI chip'ini gelistirmesi, ihrac kontrolunun model katmanindan silicon katmanina cevap urettigini gosteriyor. Kaynak: https://arstechnica.com/ai/2026/07/facing-us-export-controls-chinas-deepseek-plans-to-make-its-own-chips/
6. Discord, AI moderation bug'i nedeniyle zararsiz gorseller uzerinden hatali ban verdigini kabul etti; safety sadece model kalitesi degil appeal ve operasyon meselesi. Kaynak: https://techcrunch.com/2026/07/07/discord-admits-ai-moderation-bug-wrongfully-banned-users-over-harmless-images/
7. AWS Bedrock AgentCore icin image editing ve support companion ornekleri, agent runtime'in demo seviyesinden serverless is akisi kaliplarina indigini gosteriyor. Kaynaklar: https://aws.amazon.com/blogs/machine-learning/build-a-serverless-image-editing-agent-with-amazon-bedrock-agentcore-harness/ | https://aws.amazon.com/blogs/machine-learning/build-an-ai-powered-aws-support-companion-with-amazon-bedrock-agentcore/

KAT-1 - LLM & Model Guncellemeleri

• Cohere Transcribe Arabic, Arapca ASR icin acik kaynak model olarak domain-specific speech modellerinin stratejik hale geldigini gosteriyor. Kaynak: https://the-decoder.com/cohere-transcribe-arabic-is-an-open-source-model-built-for-arabics-toughest-transcription-problems/
• Hugging Face Models on Foundry Managed Compute, model dagitimini Microsoft managed compute uzerinden enterprise akisa yaklastiriyor. Kaynak: https://huggingface.co/blog/microsoft/foundry-managed-compute
• Hugging Face'ten SageMaker Studio'ya tek tik gecis, model kesfinden enterprise experimentation'a surtmeyi azaltiyor. Kaynak: https://huggingface.co/blog/amazon/one-click-to-sagemaker-studio
• Omar Sanseviero akisinda Gemma 4 technical report sinyali geldi; sosyal tarafta model mimarisi ve token hacmi tartismasi canli. Kaynak: https://news.google.com/rss/articles/CBMiYkFVX3lxTE5fZ1B3TW1CYngwWlRtdlhnYklfaWNWR04tREJDR3FOWjdxenJzbUtLOWtvQ0lkM1hZMnc1enRtV2FBNTdfaG84dTNTSm50Nlg0S3FVZ3RWbVhNV2NIYVY2Uld3?oc=5

KAT-2 - Arastirma & Paper'lar

• Anthropic Jacobian Lens haberi, Claude'un gizli ic temsilini okunabilir hale getirme iddiasiyla interpretability hattinda pratik sinyal verdi. Kaynak: https://the-decoder.com/claudes-hidden-inner-monologue-is-now-readable-thanks-to-anthropics-new-jacobian-lens/
• SelfMem, AI agent memory sistemleri icin self-optimizing memory yaklasimini one cikariyor; persistent memory artik kalite kadar saldiri yuzeyi. Kaynak: https://arxiv.org/abs/2607.03726
• ToolFailBench, LLM agent'larda tool-use failure siniflandirmasini ayrica benchmark haline getiriyor. Kaynak: https://arxiv.org/abs/2607.04686
• MIRAGE, uzun-form RAG akisini misinformation pollution'a karsi savunmayi hedefliyor. Kaynak: https://arxiv.org/abs/2607.05069
• CausalGame, LLM agent'larin oyun ortaminda causal thinking kapasitesini olcmeye calisiyor. Kaynak: https://arxiv.org/abs/2607.04293

KAT-3 - Araclar & Framework'ler

• Google Managed Agents, background execution ve remote MCP ile agent orchestration'i API urunu haline getiriyor. Kaynak: https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api/
• AWS DevOps Agent release management genislemesi, agent'larin production oncesi kod validasyonu ve test hattina yerlestigini gosteriyor. Kaynak: https://www.infoq.com/news/2026/07/aws-devops-ai-agent/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• GitHub Copilot app'in tum planlara acilmasi, coding-agent operasyonunu masaustu uygulama yuzeyine tasiyor. Kaynak: https://github.blog/changelog/2026-07-07-github-copilot-app-available-to-all
• Secret scanning extended metadata, leaked secret ownership ve impact triage tarafinda operasyon kalitesini artiriyor. Kaynak: https://github.blog/changelog/2026-07-07-secret-scanning-extended-metadata-and-multipart-validation

KAT-4 - Acik Kaynak

• addyosmani/agent-skills, coding agent'lar icin production-grade workflow ve quality gate'leri skill formatinda paketliyor. Kaynak: https://github.com/addyosmani/agent-skills
• mvanhorn/last30days-skill, Reddit, X, YouTube, HN ve web uzerinden agent-led son 30 gun arastirmasini tek skill'e indiriyor. Kaynak: https://github.com/mvanhorn/last30days-skill
• alibaba/zvec, hafif ve in-process vector database olarak local/embedded RAG senaryolari icin izlenmeli. Kaynak: https://github.com/alibaba/zvec
• bradautomates/claude-video, Claude'a video izletme akisini frame extraction ve transkripsiyonla agent aracina donusturuyor. Kaynak: https://github.com/bradautomates/claude-video
• karakeep, self-hosted bookmark ve AI tagging kombinasyonuyla personal knowledge stack tarafinda faydali sinyal verdi. Kaynak: https://github.com/karakeep-app/karakeep

KAT-5 - Endustri & Sirketler

• Anthropic, open source AI yukselisine ragmen henuz talep kaybi yasamiyor; fark daha cok enterprise guven ve agent urunlesmesi tarafinda. Kaynak: https://techcrunch.com/2026/07/07/why-the-rise-of-open-source-ai-isnt-hurting-anthropic-yet/
• Australian Payments Plus, ChatGPT ve Codex ile is akisi hizlandirma case'i olarak OpenAI enterprise adoption hattina eklendi. Kaynak: https://openai.com/index/australian-payments-plus
• Norm, AI law startup olarak 120 milyon dolar fonlama ve unicorn degerleme sinyali verdi; legal AI pazari kurumsal compliance'a kayiyor. Kaynak: https://news.google.com/rss/articles/CBMikwFBVV95cUxNZkxSeG9MLUpVWEctU1ZhWlJJMFNYVTBuX2lqLThUNFM4b215TGEtN1hnOGZDTWZvekR6QnhUSUNmNTBzQnlQNW5HU0RSVnZkZW5abmpRVGRCU3BtY25zQjlEdS1IZFJMQnVPaUVsa280TjllZkJZeEF2cnNhRkpCZUNXaG5YdzJEN2JfWHlSbFI1SDg?oc=5
• Apollo ekonomisti, AI kar kazanclarinin teknoloji disi sektorlerde Wall Street beklentisinden cok daha gec gelebilecegi uyarisi yapti. Kaynak: https://the-decoder.com/apollo-economist-warns-ai-profit-gains-outside-tech-could-take-well-beyond-what-wall-street-expects/

KAT-6 - AI Agent'lar

• Claude Cowork mobil ve web'e aciliyor; coding-agent savasi office/collaboration yuzeyine tasiyor. Kaynak: https://techcrunch.com/2026/07/07/the-coding-agent-wars-are-spilling-into-the-rest-of-the-office-claude-cowork/
• AgentLTL, tool-using LLM agent'larda procedural compliance'i trace verification ile olcmeyi hedefliyor. Kaynak: https://arxiv.org/abs/2607.02599
• SkillFab, agent-native skill production platform fikrini akademik tarafta paketliyor. Kaynak: https://arxiv.org/abs/2607.03780
• CAGE-1, enterprise agentic AI icin control, assurance ve governance benchmark ihtiyacini somutlastiriyor. Kaynak: https://arxiv.org/abs/2607.03510
• AgentGym2, idealize edilmemis gercek dunya ortamlari icin LLM agent benchmark'i sunuyor. Kaynak: https://arxiv.org/abs/2607.05174

KAT-7 - Multimodal

• Meta Muse Image, Instagram kullanicilarini prompt olarak kullanabilen gorsel model yetenegiyle identity/consent riskini one cikardi. Kaynak: https://www.theverge.com/tech/962485/meta-muse-image-ai-model-instagram
• Homer, uzun videolari hierarchical memory ve agentic reasoning ile anlamayi hedefliyor. Kaynak: https://arxiv.org/abs/2607.02588
• Latent Visual Cache, video reasoning icin gorsel cache fikrini one cikariyor. Kaynak: https://arxiv.org/abs/2607.02607
• PixWorld, 3D scene generation ve reconstruction'i pixel space'te birlestirmeye calisiyor. Kaynak: https://huggingface.co/papers/2607.05373
• Solos AirGo A6, camerasiz daha hafif smart glasses formuyla AI assistant cihazlarinda privacy pozisyonunu kullaniyor. Kaynak: https://www.theverge.com/tech/961711/solos-airgo-a6-smart-glasses-ai-assistant-privacy

KAT-8 - Robotik & Embodied AI

• Amerikan otonom kara araclari Ukrayna'da savas alanina girdi; embodied autonomy artik lab disi operasyonel alanda test ediliyor. Kaynak: https://techcrunch.com/2026/07/07/the-first-american-autonomous-ground-vehicles-are-fighting-in-ukraine/
• Ars Technica, AI destekli robot isci hattinin isyeri ve ev senaryolarina nasil yaklastigini isliyor. Kaynak: https://arstechnica.com/features/2026/07/robot-workers-rising-how-ai-may-drive-general-purpose-autonomy-in-robotics/
• InternVLA-A1.5, understanding, latent foresight ve action'i compositional generalization icin birlestiriyor. Kaynak: https://huggingface.co/papers/2607.04988
• GigaWorld-1, robot policy evaluation icin world model roadmap'i sunuyor. Kaynak: https://huggingface.co/papers/2607.02642
• SoftVTBench, deformable object manipulation icin safety-aware visuo-tactile benchmark sinyali verdi. Kaynak: https://arxiv.org/abs/2607.04234

KAT-9 - Veri & Altyapi

• Hugging Face ve SkyPilot zero-egress storage, multi-cloud AI workload maliyetini storage/egress seviyesinde azaltmaya calisiyor. Kaynak: https://huggingface.co/blog/skypilot-hf-storage
• NVIDIA Vera CPU, agentic workload'lar icin single-thread CPU performansinin AI factory throughput'ta kritik oldugunu vurguluyor. Kaynak: https://developer.nvidia.com/blog/nvidia-vera-cpu-boosts-ai-factory-throughput-to-accelerate-agentic-workloads/
• Momentic'in PostgreSQL'den ClickHouse'a gecisi, AI testing/cache workload'larinda analitik veri tabani ihtiyacini gosteriyor. Kaynak: https://www.infoq.com/news/2026/07/momentic-postgres-clickhouse/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• FuriosaAI'nin Avrupa datacenter hamlesi, inference chip alternatiflerinin Nvidia disi pazarda yer aradigini gosteriyor. Kaynak: https://www.theregister.com/ai-and-ml/2026/07/07/south-korean-chip-startup-furiosaai-invades-european-datacenters/5267884

SECURITY

• GitHub AI agent leak, agent permission ve repo isolation icin kritik uyari verdi. Kaynak: https://www.theregister.com/security/2026/07/07/github-ai-agent-leaks-private-repos-when-asked-nicely/5267924
• Januscape Linux flaw, Intel ve AMD cihazlarda VM escape riskini gundeme tasidi. Kaynak: https://www.bleepingcomputer.com/news/linux/new-januscape-linux-kernel-flaw-allows-vm-escape-on-intel-amd-devices/
• Tenda router firmware icindeki hidden backdoor, edge cihaz guvenliginin hala en zayif halka oldugunu gosteriyor. Kaynak: https://www.bleepingcomputer.com/news/security/hidden-backdoor-in-tenda-router-firmware-grants-admin-access/
• GitHub Actions attack pattern, CI security scanner'larin kacirdigi workflow seviyesindeki risklere isaret ediyor. Kaynak: https://www.bleepingcomputer.com/news/security/the-github-actions-attack-pattern-your-ci-security-scanners-miss/
• Agent data injection ve forged memory paper'lari, persistent agent memory icin yeni threat model'i aciyor. Kaynaklar: https://arxiv.org/abs/2607.05120 | https://arxiv.org/abs/2607.05029

REGULATION

• Cin'in en iyi AI modellerine yabanci erisimi sinirlama planlari, open model jeopolitigini ihrac kontrolu hattina tasiyor. Kaynak: https://the-decoder.com/china-eyes-export-curbs-on-its-top-ai-models-and-europe-is-caught-in-the-middle/
• Illinois, AI safety duzenlemesiyle sirket sorumlulugunu eyalet seviyesinde daha net kurmaya calisiyor. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4d701c7e694b818510a9f8d56302cd&url=https%3a%2f%2fwww.yahoo.com%2fnews%2fvideos%2fillinois-adopts-artificial-intelligence-safety-160602744.html&c=2816284873912463041&mkt=tr-tr
• UK tech sovereignty uyarisi, AI altyapisi ve kamu kapasitesi icin ulusal kontrol tartismasini buyutuyor. Kaynak: https://www.theregister.com/ai-and-ml/2026/07/07/mps-tell-brit-government-sort-out-your-tech-sovereignty-or-get-left-out-in-the-cold/5267076
• Cloudflare'in UK Cyber Resilience Pledge'e katilmasi, siber dayaniklilik icin gonullu governance hattina kurumsal destek sinyali. Kaynak: https://blog.cloudflare.com/cloudflare-joins-uk-cyber-resilience-pledge/

AI-SCIENCE

• Rethinking Scientific Discovery in an Agentic Era, bilimsel kesif akisini agent'lar uzerinden yeniden tasarlama tartismasini guclendiriyor. Kaynak: https://arxiv.org/abs/2607.03863
• Compressing the Validation Bottleneck, self-driving lab'larda dogrulama darbogazini agentic sistemle azaltmayi hedefliyor. Kaynak: https://arxiv.org/abs/2607.04508
• AutoResearch, execution-grounded multi-agent research workflow automation icin pratik framework sunuyor. Kaynak: https://arxiv.org/abs/2607.02520
• ResearchStudio-Reel, paper'dan poster, video ve blog'a son kilometre research dissemination otomasyonunu hedefliyor. Kaynak: https://huggingface.co/papers/2607.04438

INFRA

• Data center enerji talebi, ABD uretim maliyetleri ve "Made in America" planlariyla carpismaya basladi. Kaynak: https://arstechnica.com/tech-policy/2026/07/us-manufacturers-energy-costs-soar-because-of-ai-data-center-demand/
• UK datacenter planlama sureclerini hizlandirarak yerel itiraz bariyerini azaltmayi hedefliyor. Kaynak: https://www.theregister.com/on-prem/2026/07/07/uk-guts-planning-red-tape-so-datacenters-can-bypass-the-neighbors-faster/5266865
• Samsung karinin 19 kat artmasina ragmen hisselerin dusmesi, AI chip beklentilerinin ne kadar sert fiyatlandigini gosteriyor. Kaynak: https://www.cnbc.com/2026/07/07/chip-stocks-ai-selloff-samsung.html
• Broadcom ve Apple custom silicon ortakligini 2031'e uzatti; cihaz uzeri AI icin silicon tedariki stratejik kalmaya devam ediyor. Kaynak: https://www.theregister.com/personal-tech/2026/07/07/broadcom-and-apple-extend-custom-silicon-pact-to-2031/5267491

SAFETY

• Safety Testing LLM Agents at Scale, risk discovery'den evidence-grounded verification'a agent safety test hattini sistemlestiriyor. Kaynak: https://huggingface.co/papers/2607.01793
• When Agents Lie, repeated games icinde agent'larin premeditation, persistence ve exploitation davranislarini inceliyor. Kaynak: https://arxiv.org/abs/2607.05132
• Uncertainty-Aware Abstention, LLM'lerde cevap vermeme davranisini provable alignment guarantee ile baglamaya calisiyor. Kaynak: https://arxiv.org/abs/2607.04430
• RetroCoT, model nesilleri boyunca forensic reconstruction prompt'larini safety diagnostic olarak kullaniyor. Kaynak: https://arxiv.org/abs/2607.04645
• Global AI industry safety ranking haberleri, buyuk lab'larin hala A alamadigi bir safety governance boslugu oldugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMikwFBVV95cUxOYThPNHF6SDhnSFdfOFFJNVl2dXB6RzhxMnNDblQ1VG9jeWgzazZXTUhYcm5ONU5JVWF2OHhxbnRTQnR2LTQ5TGdfRFU3MzJjbTJ3dzI0eVBteWFlV0lDZU5Za0U5dW9PcW4tMk54a2J6QjczU3c1UVZRSk04aGM2LTNUZE5yVDVhWXBmOHBaTC1HN00?oc=5

WATCHLIST

• Managed-agent platform savasi: Google Managed Agents, AWS AgentCore ve Claude Cowork ayni hafta agent runtime'i kurumsal yuzeye tasidi. Kaynaklar: https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api/ | https://aws.amazon.com/blogs/machine-learning/build-an-ai-powered-aws-support-companion-with-amazon-bedrock-agentcore/ | https://techcrunch.com/2026/07/07/the-coding-agent-wars-are-spilling-into-the-rest-of-the-office-claude-cowork/
• Agent memory threat model: SelfMem, MOSS, forged memory ve stealth memory injection paper'lari ayni gun kume olusturdu. Kaynaklar: https://arxiv.org/abs/2607.03726 | https://arxiv.org/abs/2607.04391 | https://arxiv.org/abs/2607.05029 | https://arxiv.org/abs/2607.05189
• China AI sovereignty: DeepSeek chip, model export curbs ve OpenRouter uzerinde Chinese model payi beraber izlenmeli. Kaynaklar: https://arstechnica.com/ai/2026/07/facing-us-export-controls-chinas-deepseek-plans-to-make-its-own-chips/ | https://the-decoder.com/china-eyes-export-curbs-on-its-top-ai-models-and-europe-is-caught-in-the-middle/ | https://the-decoder.com/chinese-ai-models-regularly-pass-30-percent-on-openrouter-as-cost-gap-widens/
• AI moderation risk: Discord bug ve Meta Muse identity prompt akisi consumer AI governance icin tekrar eden risk hatti. Kaynaklar: https://techcrunch.com/2026/07/07/discord-admits-ai-moderation-bug-wrongfully-banned-users-over-harmless-images/ | https://www.theverge.com/tech/962485/meta-muse-image-ai-model-instagram

CikCik Paketi

• Simon Willison, sqlite-utils 4.0 release review'unda Claude Fable 5'in dort release blocker yakaladigini paylasti; coding-agent review kalitesi icin guclu saha sinyali. Kaynak: https://simonwillison.net/2026/Jul/7/sqlite-utils-4/#claude-fable-5-helped-a-lot
• Hacker News/Mastodon akisi "30papers.com" ile Ilya'nin 30 temel ML paper'ini beginner-friendly formatta one cikardi. Kaynak: https://mastodon.social/@h4ckernews/116879831597945715
• "Weighing smoke" tartismasi, AI visibility dashboard'larinin olcum kalitesi konusunda supheyi topluluk gundemine tasidi. Kaynak: https://mastodon.social/@h4ckernews/116879764798552448
• NPR, AI sirketlerinin model gelistirme surecinde filozoflari ise almasini isledi; safety ve deger tasarimi hiring tarafina indi. Kaynak: https://www.npr.org/2026/07/07/nx-s1-5875184/why-ai-companies-are-hiring-philosophers-to-help-develop-their-models
• Agent Name Service, AI agent identity layer icin sosyal akista yeni altyapi fikri olarak belirdi. Kaynak: https://mastodon.social/@sjvn/116880783049280152
• OfficeCLI, Word/Excel/PowerPoint'i AI agent'lara kontrol ettirme araciyla desktop automation risk/firsat hattina girdi. Kaynak: https://web.brid.gy/r/https://gigazine.net/news/20260708-officecli/

Oracle Self-Improvement Sinyalleri

• Agent-memory paper kumesi, OpenClaw persistent memory icin "read/write authority", forged reasoning ve stealth injection kontrollerini ayri test kategorisi yapmayi gerektiriyor. Kaynaklar: https://arxiv.org/abs/2607.05029 | https://arxiv.org/abs/2607.05189
• ToolFailBench ve AgentLTL, subagent verification gate'lerini sadece task success ile degil tool-use failure ve procedural compliance ile olcmek gerektigini gosteriyor. Kaynaklar: https://arxiv.org/abs/2607.04686 | https://arxiv.org/abs/2607.02599
• GitHub AI agent leak, agent'larin repo izinlerinin "prompt ile erisilebilirlik" perspektifinden test edilmesi gerektigini hatirlatiyor. Kaynak: https://www.theregister.com/security/2026/07/07/github-ai-agent-leaks-private-repos-when-asked-nicely/5267924
• Dedupe tarafinda 4156 fingerprint tekrar ve 27 semantik tekrar elendi; sosyal ve akademik collector hacmi buyuk, kalite icin kaynak guvenilirligi agirligi korunmali. Kaynak: raw-sources/2026-07-08/unique.json

Kaynak Ozeti

• Toplam 9015 ham item toplandi; 3046 benzersiz item ve 106 benzersiz kaynak rapora girdi. Kaynak: raw-sources/2026-07-08/unique.json
• Kaynak aileleri 5/5 tamam: RSS/news, search, community, social, academic/api. Kaynak: raw-sources/2026-07-08/unique.json
• RSS/news tarafinda DonanimHaber, Nikkei Asia ve The Register; social tarafta twitter_fallback/@sama, twitter_fallback/@lilianweng ve Mastodon #MachineLearning baskin. Kaynak: raw-sources/2026-07-08/unique.json
• Academic/API tarafinda arxiv/cs.LG, arxiv/cs.CV ve arxiv/cs.AI agirligi cok yuksek; agent/security paper'lari bu yuzden bol. Kaynak: raw-sources/2026-07-08/unique.json

Coverage / Blind Spots

• Eksik kaynak ailesi yok; thin family yok; bos topic yok; thin topic yok. Kaynak: raw-sources/2026-07-08/unique.json
• Search ailesi aggregator linklerine dayaniyor; canonical URL cozumleme hala kalite artisi saglar. Kaynak: raw-sources/2026-07-08/search.json
• Reddit tarafinda 429 hatalari surdu; community family GitHub Trending ve Lobsters ile dolduruldu. Kaynak: raw-sources/2026-07-08/community.json
• Social collector hacmi yuksek ama Google News fallback gurultulu; CikCik seciminde dogrudan sosyal URL ve guvenilir blog linkleri tercih edildi. Kaynak: raw-sources/2026-07-08/social.json

Bu Gece Sistem Ne Ogrendi

• Kalici ders: Launches, Regulation ve Models yine baskin, ancak bugunku ayrim agent runtime + safety + infra katmaninda; skorlamada "agent governance" boost'u korunmali. Kaynaklar: https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api/ | https://www.theregister.com/security/2026/07/07/github-ai-agent-leaks-private-repos-when-asked-nicely/5267924
• Rising entities artifact'indeki Google, AI Agents, Anthropic ve OpenAI sinyalleri bugun de devam etti; fakat Microsoft ve DeepSeek maliyet/egemenlik tarafinda daha yuksek karar degeri uretmeye basladi. Kaynaklar: https://techcrunch.com/2026/07/07/microsoft-joins-ai-cost-cutting-trend-by-relying-more-on-its-own-models/ | https://arstechnica.com/ai/2026/07/facing-us-export-controls-chinas-deepseek-plans-to-make-its-own-chips/
• Tekrar korumasi calisti: onceki 3 gun ana maddeleri olan LeRobot v0.6, x402/MCP auth, SkillCloak, MTurk kapanisi, Alibaba-Claude yasagi ve Qwen/Leanstral ana anlatilari Top 7'ye alinmadi. Kaynak: reports/2026-07-05.md
• Bir sonraki run icin ders: "agent memory attack", "managed agents", "AI moderation bug" ve "model export controls" watchlist terimleri trend tracker'a daha yuksek agirlikla eklenmeli. Kaynaklar: https://arxiv.org/abs/2607.05029 | https://techcrunch.com/2026/07/07/discord-admits-ai-moderation-bug-wrongfully-banned-users-over-harmless-images/ | https://the-decoder.com/china-eyes-export-curbs-on-its-top-ai-models-and-europe-is-caught-in-the-middle/

Dedupe & Kalite Notu

Bu rapordaki tum maddeler onceki 3 gunun raporlarindan elendi/dedupe edilmistir. Toplam 9015 ham item islendi; merge asamasinda 7229 kayit kaldi; dedupe sonrasi 3046 benzersiz item rapor havuzuna girdi. Dedupe sonucu: 4156 fingerprint tekrar, 27 semantik tekrar, 0 guncelleme. Kaynak: raw-sources/2026-07-08/unique.json