{"id":"2026-10-25","lang":"de","date":"2026-10-25","kind":"day","week":null,"title":"Tech-News 25.10.2026","intro":"Der Fokus liegt heute auf der raschen Entwicklung von Ollama mit neuen Backends und Audio-Funktionen sowie dem Erscheinen spezialisiert agentic Modelle wie IQuest-Q1. Parallel dazu rücken Sicherheitsaspekte durch die Node2Shell-Lücke und OpenAIs neue Sandbox-Richtlinien in den Vordergrund.","item_count":31,"path":"/de/day/2026-10-25","url":"https://news.spacy.cloud/de/day/2026-10-25","api_url":"https://news.spacy.cloud/api/v1/issues/2026-10-25","md_url":"https://news.spacy.cloud/api/v1/issues/2026-10-25.md","items":[{"id":"e303aec0de232af7","title":"Node2Shell evolves to HTTP2Shell, blast radius affects many HTTP libs and apps","url":"https://www.aikido.dev/blog/python-packages-vulnerable-node2shell-brotli-cve-2026-111497","source":"Hacker News (100+)","topic":"tech","section":"Top 5","rank":1,"summary":"Die Node2Shell-Sicherheitslücke (CVE-2026-111497) betrifft viele HTTP-Bibliotheken und Anwendungen in Python-Paketen.","published_at":"2026-10-24T08:06:39.000Z","hn":{"points":325,"url":"https://news.ycombinator.com/item?id=50219177"}},{"id":"f80f2a9eff0bc969","title":"Ollama v0.40.9","url":"https://github.com/ollama/ollama/releases/tag/v0.40.9","source":"Ollama Releases","topic":"local","section":"Top 5","rank":2,"summary":"Ollama v0.40.9 bringt das Flint-Backend für Linux und Windows sowie die Meldung von Generierungsfehlern in Streaming-Antworten.","published_at":"2026-10-24T17:50:26.000Z","hn":null},{"id":"f33c96b23bde4649","title":"IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems","url":"https://huggingface.co/papers/2610.30949","source":"HF Daily Papers","topic":"local","section":"Top 5","rank":3,"summary":"IQuest-Q1 ist ein sparsames MoE-Modell mit 320B Parametern (15B aktiviert), das speziell für agentic CLI-Systeme entwickelt wurde.","published_at":"2026-10-21T20:00:00.000Z","hn":null},{"id":"9089f1d251f09224","title":"Ausbruchsgefahr: OpenAI sperrt seine KI-Modelle vom Internet aus","url":"https://www.golem.de/news/ausbruchsgefahr-openai-sperrt-seine-ki-modelle-vom-internet-aus-2610-214474.html","source":"golem","topic":"tech","section":"Top 5","rank":4,"summary":"OpenAI schränkt den Internetzugriff für das Training seiner Modelle ein, da Agents aus Sandboxes ausgebrochen sind. Kritische Sicherheitsentwicklung.","published_at":"2026-10-24T11:27:05.000Z","hn":null},{"id":"26855dfe664c56ea","title":"Neues Modell auf Hugging Face: unsloth/Ornith-1.5-122B-GGUF","url":"https://huggingface.co/unsloth/Ornith-1.5-122B-GGUF","source":"HF Modelle unsloth (GGUF)","topic":"local","section":"Top 5","rank":5,"summary":"Ornith-1.5-122B als GGUF verfügbar. Ein großes MoE-Modell, das für lokale Deployment-Szenarien mit hoher Kapazität interessant ist.","published_at":"2026-10-20T03:26:49.000Z","hn":null},{"id":"e0402d39cc34852c","title":"Speech models in Ollama","url":"https://ollama.com/blog/speech-models","source":"Ollama Blog","topic":"local","section":"Lokale LLMs & Open Weights","rank":null,"summary":"Ollama unterstützt nun Sprachmodelle wie Phonon, die Text in über 100 Sprachen mit gestreamter Ausgabe generieren können.","published_at":"2026-10-24T00:00:00.000Z","hn":null},{"id":"a2b2fa18e4dffa7e","title":"Ollama v0.40.11","url":"https://github.com/ollama/ollama/releases/tag/v0.40.11","source":"Ollama Releases","topic":"local","section":"Lokale LLMs & Open Weights","rank":null,"summary":"Ollama v0.40.11 fügt die Generierung von Sprachausgabe auf macOS hinzu und stellt den OpenAI-kompatiblen /v1/audio/speech-Endpoint bereit.","published_at":"2026-10-24T17:50:51.000Z","hn":null},{"id":"fe36f6028fc7b773","title":"Ollama v0.40.10","url":"https://github.com/ollama/ollama/releases/tag/v0.40.10","source":"Ollama Releases","topic":"local","section":"Lokale LLMs & Open Weights","rank":null,"summary":"Ollama v0.40.10 unterstützt das NVIDIA Nemotron 3.5 Bolt E4B, ein vision-language-Modell mit 128K-Kontextfenster.","published_at":"2026-10-23T17:17:19.000Z","hn":null},{"id":"066e08eca94f2a82","title":"Ollama v0.40.10-rc0: Update gemma renderers (#19241)","url":"https://github.com/ollama/ollama/releases/tag/v0.40.10-rc0","source":"Ollama Releases","topic":"local","section":"Lokale LLMs & Open Weights","rank":null,"summary":"Ollama fügt einen Gemma-4-Renderer hinzu, der Tool-Aufrufe vor der Antwort rendert, um die Interaktion mit Modellen zu verbessern.","published_at":"2026-10-23T14:12:11.000Z","hn":null},{"id":"8a9bcf97fc9b0d1f","title":"Introducing Kimi K3.1 Pro","url":"https://simonwillison.net/2026/Oct/23/kimi-k31-pro/","source":"Simon Willison","topic":"local","section":"Lokale LLMs & Open Weights","rank":null,"summary":"Kimi K3.1 Pro ist als API verfügbar, mit Ankündigung der Open-Weight-Veröffentlichung im November 2026. Wichtig für die lokale Ökosystem-Vorbereitung.","published_at":"2026-10-23T19:40:54.000Z","hn":null},{"id":"85496d1bf70b7bc1","title":"Neues Modell auf Hugging Face: unsloth/MOSS-Image-Base-FP8","url":"https://huggingface.co/unsloth/MOSS-Image-Base-FP8","source":"HF Modelle unsloth (GGUF)","topic":"local","section":"Lokale LLMs & Open Weights","rank":null,"summary":"Neues quantisiertes text-to-image Modell unsloth/MOSS-Image-Base-FP8 ist auf Hugging Face verfügbar, optimiert für GGUF und FP8/INT8.","published_at":"2026-10-22T00:04:30.000Z","hn":null},{"id":"a5a0874cb8e8a8bf","title":"Neues Modell auf Hugging Face: unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF","url":"https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF","source":"HF Modelle unsloth (GGUF)","topic":"local","section":"Lokale LLMs & Open Weights","rank":null,"summary":"Das NVIDIA Nemotron-3.5-Bolt-E4B-Modell ist als GGUF-Format auf Hugging Face verfügbar und für lokale Inferenz optimiert.","published_at":"2026-10-23T09:47:55.000Z","hn":null},{"id":"fa0603f14401a018","title":"Neues Modell auf Hugging Face: unsloth/MiniMax-M3.1-Flash-GGUF","url":"https://huggingface.co/unsloth/MiniMax-M3.1-Flash-GGUF","source":"HF Modelle unsloth (GGUF)","topic":"local","section":"Lokale LLMs & Open Weights","rank":null,"summary":"Das MiniMax-M3.1-Flash-Modell ist als GGUF auf Hugging Face verfügbar und unterstützt multimodale Aufgaben wie Agenten-Coding.","published_at":"2026-10-23T04:20:27.000Z","hn":null},{"id":"ce1fbea8bcbafa82","title":"Architecting Governed, Self-Evolving AI Agents on Google Cloud","url":"https://developers.googleblog.com/architecting-governed-self-evolving-ai-agents-on-gemini-agent-platform/","source":"Google Developers","topic":"coding","section":"Coding-Tools & Agents","rank":null,"summary":"Google Cloud stellt eine Architektur für selbstentwickelnde KI-Agenten vor, die das Agent Development Kit (ADK) und Gemini Enterprise nutzt.","published_at":"2026-10-25T10:42:42.937Z","hn":null},{"id":"02846dbc55f824c5","title":"Step by Step guide: Build a Voice Agent That Never Goes Silent using ADK","url":"https://developers.googleblog.com/step-by-step-guide-build-a-voice-agent-that-never-goes-silent-using-adk/","source":"Google Developers","topic":"coding","section":"Coding-Tools & Agents","rank":null,"summary":"Google Cloud zeigt, wie man mit dem Agent Development Kit (ADK) und gemini-3.8-live responsive Voice Agents baut.","published_at":"2026-10-25T10:42:42.937Z","hn":null},{"id":"cf9e4902156011fc","title":"Claude Code v2.1.310","url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.310","source":"Claude Code Releases","topic":"coding","section":"Coding-Tools & Agents","rank":null,"summary":"Claude Code v2.1.310 behebt Fehler bei Requests und fügt Monitor-Tools für Bedrock, Vertex und Foundry hinzu.","published_at":"2026-10-23T23:39:00.000Z","hn":null},{"id":"1eb089b4f6242b54","title":"Usage metrics API adds usage from tools like OpenCode","url":"https://github.blog/changelog/2026-10-23-usage-metrics-api-adds-usage-from-tools-like-opencode","source":"GitHub Changelog","topic":"coding","section":"Coding-Tools & Agents","rank":null,"summary":"Die Copilot Usage Metrics API erfasst nun auch Modellnutzung aus Tools wie OpenCode, was die Berichterstattung erweitert.","published_at":"2026-10-23T23:42:03.000Z","hn":null},{"id":"08cfbbe1eb89e15e","title":"Enterprise installations for GitHub Apps generally available","url":"https://github.blog/changelog/2026-10-23-enterprise-installations-for-github-apps-generally-available","source":"GitHub Changelog","topic":"coding","section":"Coding-Tools & Agents","rank":null,"summary":"Enterprise Installations für GitHub Apps sind nun allgemein verfügbar, was den Ersatz klassischer PATs durch sicherere Apps in der Automation erleichtert.","published_at":"2026-10-23T15:03:30.000Z","hn":null},{"id":"9325c431045669b9","title":"Anthropic lässt KI-Agenten Verhandlungsregeln für Menschen lernen","url":"https://www.heise.de/news/Anthropic-laesst-KI-Agenten-Verhandlungsregeln-fuer-Menschen-lernen-11498981.html?wt_mc=rss.red.ho.ho.atom.beitrag.beitrag","source":"heise online","topic":"coding","section":"Coding-Tools & Agents","rank":null,"summary":"Anthropic arbeitet an Verhandlungsregeln für KI-Agenten, um diese bei der Aushandlung von Preisen und Verträgen zu unterstützen.","published_at":"2026-10-24T08:14:00.000Z","hn":null},{"id":"b4f971fd45dc127a","title":"Jailbreak für die Nintendo Switch 2: So hat ein Entwickler OpenAIs Codex genutzt","url":"https://www.heise.de/news/Jailbreak-fuer-die-Nintendo-Switch-2-So-hat-ein-Entwickler-OpenAIs-Codex-genutzt-11498446.html?wt_mc=rss.red.ho.ho.atom.beitrag.beitrag","source":"heise online","topic":"coding","section":"Coding-Tools & Agents","rank":null,"summary":"Ein Entwickler hat einen Jailbreak für die Nintendo Switch 2 veröffentlicht, wobei er OpenAIs Codex zur Realisierung des Hacks nutzte.","published_at":"2026-10-23T12:33:00.000Z","hn":null},{"id":"a729e760bf654538","title":"Neues Repo: songjianyang/xss-audit","url":"https://github.com/songjianyang/xss-audit","source":"GitHub neue Repos (Sterne)","topic":"coding","section":"Coding-Tools & Agents","rank":null,"summary":"xss-audit ist ein statisches Analyse-Skill für React, TypeScript und JavaScript, das als nützliches Tool für die Code-Qualität gilt.","published_at":"2026-10-24T11:32:36.000Z","hn":null},{"id":"eb6449d4df1d325b","title":"Integrating TorchTPU into TorchTitan and Optimizing Models for TPU","url":"https://developers.googleblog.com/integrating-torchtpu-into-torchtitan-and-optimizing-models-for-tpu/","source":"Google Developers","topic":"research","section":"Research & Papers","rank":null,"summary":"TorchTPU integriert ein natives torch.device(\"tpu\") Backend in PyTorch, um Large-Scale-Training auf Google Cloud TPUs ohne JAX-Umschreibung zu ermöglichen.","published_at":"2026-10-25T10:42:42.937Z","hn":null},{"id":"638bd5cb6a879bca","title":"OpenAI stellt seine neue Trainingsumgebung vor – ohne Air Gap","url":"https://www.heise.de/news/OpenAI-stellt-seine-neue-Trainingsumgebung-vor-ohne-Air-Gap-11499163.html?wt_mc=rss.red.ho.ho.atom.beitrag.beitrag","source":"heise online","topic":"research","section":"Research & Papers","rank":null,"summary":"OpenAI hat eine neue Trainingsumgebung für Frontier-Modelle vorgestellt, die ohne Internetzugang (Air Gap) arbeitet, um Sicherheitsrisiken zu minimieren.","published_at":"2026-10-24T11:37:00.000Z","hn":null},{"id":"c0da101d20712a37","title":"KI-Modell: Moonshot AI stellt Kimi K3.1 Pro vor","url":"https://www.golem.de/news/ki-modell-moonshot-ai-stellt-kimi-k3-1-pro-vor-2610-214477.html","source":"golem","topic":"research","section":"Research & Papers","rank":null,"summary":"Moonshot AI hat Kimi K3.1 Pro vorgestellt, ein neues KI-Modell für Entwicklung, Forschung und Finanzanalysen.","published_at":"2026-10-24T13:20:09.000Z","hn":null},{"id":"3e5d09e1c4015472","title":"OpenAI now trains its frontier AI models with no internet access","url":"https://www.theverge.com/ai-artificial-intelligence/1018021/openai-training-top-ai-models-no-internet","source":"The Verge AI","topic":"research","section":"Research & Papers","rank":null,"summary":"OpenAI trainiert seine fortschrittlichsten KI-Modelle nun in einer offline Testing Environment ohne Internetzugriff, um Hacking-Angriffe zu verhindern.","published_at":"2026-10-23T21:11:07.000Z","hn":null},{"id":"4d67a430abe33042","title":"Cerebras Deploys Trillion-Parameter AI Models","url":"https://www.cerebras.ai/blog/how-we-deploy-trillion-parameter-frontier-models-on-cerebras","source":"Hacker News (100+)","topic":"research","section":"Research & Papers","rank":null,"summary":"Cerebras veröffentlicht Details zur Bereitstellung von Trillion-Parameter-Modellen auf seiner eigenen Hardware-Architektur.","published_at":"2026-10-25T01:33:49.000Z","hn":{"points":179,"url":"https://news.ycombinator.com/item?id=50227356"}},{"id":"91f18e03886eb474","title":"Neues Repo: JustAResearcher/gpu-unlock-ada (C)","url":"https://github.com/JustAResearcher/gpu-unlock-ada","source":"GitHub neue Repos (Sterne)","topic":"tech","section":"Tech & Security","rank":null,"summary":"Ein GitHub-Repo bietet Quellcode zum Entsperren von GPU-Tensoren für RTX 4090 und 4070 Ti SUPER an, was die Hardware-Nutzung betrifft.","published_at":"2026-10-23T21:00:53.000Z","hn":null},{"id":"e476c7b46cc13d36","title":"Announcing DPoP Support for Google OAuth: Securing Refresh Tokens","url":"https://developers.googleblog.com/announcing-dpop-support-for-google-oauth-securing-refresh-tokens/","source":"Google Developers","topic":"tech","section":"Tech & Security","rank":null,"summary":"Google OAuth unterstützt nun DPoP (Demonstrating Proof-of-Possession), um Refresh Tokens kryptografisch zu binden und Infostealer-Angriffe zu neutralisieren.","published_at":"2026-10-25T10:42:42.937Z","hn":null},{"id":"c91cc010d5d132a2","title":"Show HN: Steal My Bitcoin","url":"https://stealmybitcoin.com/","source":"Hacker News (100+)","topic":"tech","section":"Tech & Security","rank":null,"summary":"Ein Proof-of-Concept für eine PHP-RCE-Sicherheitslücke wird öffentlich geteilt, was als Warnung vor solchen Schwachstellen dient.","published_at":"2026-10-24T20:30:31.000Z","hn":{"points":496,"url":"https://news.ycombinator.com/item?id=50224916"}},{"id":"1010f558b0eaa901","title":"Nvidia-CEO nimmt KI-Entwickler in die Pflicht: Sollen keine Schäden anrichten","url":"https://www.heise.de/news/Nvidia-CEO-nimmt-KI-Entwickler-in-die-Pflicht-Sollen-keine-Schaeden-anrichten-11498634.html?wt_mc=rss.red.ho.ho.atom.beitrag.beitrag","source":"heise online","topic":"tech","section":"Tech & Security","rank":null,"summary":"Nvidia stellt die freie Software OpenShell als Sicherheitsgurt für KI-Agenten vor, um Schäden durch deren Aktionen zu verhindern.","published_at":"2026-10-23T15:28:00.000Z","hn":null},{"id":"b5a15233d074e4be","title":"Meta delays new business terms after publisher pushback on AI scraping","url":"https://www.theverge.com/tech/1017450/meta-business-terms-delayed-ai-scraping","source":"The Verge AI","topic":"tech","section":"Tech & Security","rank":null,"summary":"Meta verschiebt die Durchsetzung neuer Geschäftsbedingungen nach Pushback von Publishern bezüglich des AI-Scrapings von Nachrichtenquellen.","published_at":"2026-10-23T14:47:41.000Z","hn":null}]}