https://24-ai.news/hr/news/2026-07-03/amd-agentkernelarena-benchmark/ 24 AI hr 2026-07-03T00:00:00Z AMD AgentKernelArena: otvoreni benchmark za AI agente na GPU kernel optimizaciji https://24-ai.news/hr/news/2026-07-03/amd-eagle3-speculative-decoding-quark/ 24 AI hr 2026-07-03T00:00:00Z AMD Eagle3 i Quark FP8: spekulativno dekodiranje donosi do 2,00x throughput na MI355X https://24-ai.news/hr/news/2026-07-03/amd-yolo26-gpu-resident-pipeline/ 24 AI hr 2026-07-03T00:00:00Z AMD ROCm: GPU-rezidentni YOLO26 pipeline drži video okvire u VRAM-u od dekodiranja do detekcije https://24-ai.news/hr/news/2026-07-03/anthropic-claude-code-2-1-200/ 24 AI hr 2026-07-03T00:00:00Z Claude Code v2.1.200: zadani permission mode promijenjen na Manual https://24-ai.news/hr/news/2026-07-03/arxiv-distributed-attacks-ai-control/ 24 AI hr 2026-07-03T00:00:00Z Distribuirani napadi na AI coding-agente: nijedan monitor ih ne zaustavlja istovremeno https://24-ai.news/hr/news/2026-07-03/arxiv-eo-agents-earth-observation/ 24 AI hr 2026-07-03T00:00:00Z EO-Agents: troagentni LLM pipeline generirao 160 znanstvenih hipoteza iz 1.475 NASA skupova podataka https://24-ai.news/hr/news/2026-07-03/arxiv-online-safety-monitoring-llm/ 24 AI hr 2026-07-03T00:00:00Z Jednostavan kalibrirani monitoring LLM-ova premašuje složene sekvencijalne pristupe https://24-ai.news/hr/news/2026-07-03/arxiv-pase-cloud-healing-neurosymbolic/ 24 AI hr 2026-07-03T00:00:00Z PASE: neuro-simbolički sustav skraćuje oporavak cloud kvarova za više od 40 posto https://24-ai.news/hr/news/2026-07-03/arxiv-recontext-long-context-replay/ 24 AI hr 2026-07-03T00:00:00Z ReContext poboljšava iskorištenost 128K kontekstnih prozora bez ponovnog treniranja https://24-ai.news/hr/news/2026-07-03/arxiv-vrrl-visual-self-reflection/ 24 AI hr 2026-07-03T00:00:00Z VRRL: pojačano učenje primorava vizualne modele da stvarno koriste sliku pri samokorekciji https://24-ai.news/hr/news/2026-07-03/pytorch-test-infrastructure/ 24 AI hr 2026-07-03T00:00:00Z Kako PyTorch generira tisuće testova iz nekoliko template klasa — vodič kroz test-infrastrukturu https://24-ai.news/en/news/2026-07-03/amd-agentkernelarena-benchmark/ 24 AI en 2026-07-03T00:00:00Z AMD AgentKernelArena: Open Benchmark for AI Agents on GPU Kernel Optimization https://24-ai.news/en/news/2026-07-03/amd-eagle3-speculative-decoding-quark/ 24 AI en 2026-07-03T00:00:00Z AMD Eagle3 and Quark FP8: Speculative Decoding Delivers Up to 2.00x Throughput on MI355X https://24-ai.news/en/news/2026-07-03/amd-yolo26-gpu-resident-pipeline/ 24 AI en 2026-07-03T00:00:00Z AMD ROCm: GPU-Resident YOLO26 Pipeline Keeps Video Frames in VRAM from Decode to Detection https://24-ai.news/en/news/2026-07-03/anthropic-claude-code-2-1-200/ 24 AI en 2026-07-03T00:00:00Z Claude Code v2.1.200: Default Permission Mode Changed to Manual https://24-ai.news/en/news/2026-07-03/arxiv-distributed-attacks-ai-control/ 24 AI en 2026-07-03T00:00:00Z Distributed Attacks on AI Coding Agents: No Monitor Stops Both Attack Types Simultaneously https://24-ai.news/en/news/2026-07-03/arxiv-eo-agents-earth-observation/ 24 AI en 2026-07-03T00:00:00Z EO-Agents: Three-Agent LLM Pipeline Generated 160 Scientific Hypotheses from 1,475 NASA Datasets https://24-ai.news/en/news/2026-07-03/arxiv-online-safety-monitoring-llm/ 24 AI en 2026-07-03T00:00:00Z Simple Calibrated LLM Monitoring Matches Sophisticated Sequential Approaches https://24-ai.news/en/news/2026-07-03/arxiv-pase-cloud-healing-neurosymbolic/ 24 AI en 2026-07-03T00:00:00Z PASE: Neurosymbolic System Cuts Cloud Failure Recovery Time by More Than 40 Percent https://24-ai.news/en/news/2026-07-03/arxiv-recontext-long-context-replay/ 24 AI en 2026-07-03T00:00:00Z ReContext Improves Utilization of 128K Context Windows Without Retraining https://24-ai.news/en/news/2026-07-03/arxiv-vrrl-visual-self-reflection/ 24 AI en 2026-07-03T00:00:00Z VRRL: Reinforcement Learning Forces Visual Models to Actually Use the Image During Self-Correction https://24-ai.news/en/news/2026-07-03/pytorch-test-infrastructure/ 24 AI en 2026-07-03T00:00:00Z How PyTorch Generates Thousands of Tests from a Few Template Classes — A Guide to the Test Infrastructure https://24-ai.news/de/news/2026-07-03/amd-agentkernelarena-benchmark/ 24 AI de 2026-07-03T00:00:00Z AMD AgentKernelArena: offener Benchmark für KI-Agenten zur GPU-Kernel-Optimierung https://24-ai.news/de/news/2026-07-03/amd-eagle3-speculative-decoding-quark/ 24 AI de 2026-07-03T00:00:00Z AMD Eagle3 und Quark FP8: Spekulatives Decoding bringt bis zu 2,00x Durchsatz auf MI355X https://24-ai.news/de/news/2026-07-03/amd-yolo26-gpu-resident-pipeline/ 24 AI de 2026-07-03T00:00:00Z AMD ROCm: GPU-residente YOLO26-Pipeline hält Video-Frames von der Dekodierung bis zur Erkennung im VRAM https://24-ai.news/de/news/2026-07-03/anthropic-claude-code-2-1-200/ 24 AI de 2026-07-03T00:00:00Z Claude Code v2.1.200: Standard-Permission-Modus auf Manual umgestellt https://24-ai.news/de/news/2026-07-03/arxiv-distributed-attacks-ai-control/ 24 AI de 2026-07-03T00:00:00Z Verteilte Angriffe auf KI-Coding-Agenten: Kein Monitor stoppt beide Angriffstypen gleichzeitig https://24-ai.news/de/news/2026-07-03/arxiv-eo-agents-earth-observation/ 24 AI de 2026-07-03T00:00:00Z EO-Agents: Drei-Agenten-LLM-Pipeline generierte 160 wissenschaftliche Hypothesen aus 1.475 NASA-Datensätzen https://24-ai.news/de/news/2026-07-03/arxiv-online-safety-monitoring-llm/ 24 AI de 2026-07-03T00:00:00Z Einfaches kalibriertes LLM-Monitoring übertrifft komplexe sequenzielle Ansätze https://24-ai.news/de/news/2026-07-03/arxiv-pase-cloud-healing-neurosymbolic/ 24 AI de 2026-07-03T00:00:00Z PASE: Neuro-symbolisches System verkürzt Cloud-Fehlerwiederherstellung um mehr als 40 Prozent https://24-ai.news/de/news/2026-07-03/arxiv-recontext-long-context-replay/ 24 AI de 2026-07-03T00:00:00Z ReContext verbessert die Nutzung von 128K-Kontextfenstern ohne Retraining https://24-ai.news/de/news/2026-07-03/arxiv-vrrl-visual-self-reflection/ 24 AI de 2026-07-03T00:00:00Z VRRL: Reinforcement Learning zwingt Visuelle Modelle dazu, beim Selbstkorrigieren wirklich das Bild zu nutzen https://24-ai.news/de/news/2026-07-03/pytorch-test-infrastructure/ 24 AI de 2026-07-03T00:00:00Z Wie PyTorch Tausende von Tests aus wenigen Template-Klassen generiert — ein Leitfaden zur Test-Infrastruktur https://24-ai.news/zh/news/2026-07-03/amd-agentkernelarena-benchmark/ 24 AI zh 2026-07-03T00:00:00Z AMD AgentKernelArena：针对GPU内核优化的AI代理开放基准测试框架 https://24-ai.news/zh/news/2026-07-03/amd-eagle3-speculative-decoding-quark/ 24 AI zh 2026-07-03T00:00:00Z AMD Eagle3与Quark FP8：推测性解码在MI355X上实现最高2.00倍吞吐量 https://24-ai.news/zh/news/2026-07-03/amd-yolo26-gpu-resident-pipeline/ 24 AI zh 2026-07-03T00:00:00Z AMD ROCm：GPU驻留YOLO26流水线将视频帧从解码到检测全程保留在VRAM中 https://24-ai.news/zh/news/2026-07-03/anthropic-claude-code-2-1-200/ 24 AI zh 2026-07-03T00:00:00Z Claude Code v2.1.200：默认权限模式更改为手动 https://24-ai.news/zh/news/2026-07-03/arxiv-distributed-attacks-ai-control/ 24 AI zh 2026-07-03T00:00:00Z 针对AI编程代理的分布式攻击：现有监控器无一能同时阻止两类攻击 https://24-ai.news/zh/news/2026-07-03/arxiv-eo-agents-earth-observation/ 24 AI zh 2026-07-03T00:00:00Z EO-Agents：三代理LLM流水线从1,475个NASA数据集生成160个科学假设 https://24-ai.news/zh/news/2026-07-03/arxiv-online-safety-monitoring-llm/ 24 AI zh 2026-07-03T00:00:00Z 简单的校准LLM监控超越复杂的序贯方法 https://24-ai.news/zh/news/2026-07-03/arxiv-pase-cloud-healing-neurosymbolic/ 24 AI zh 2026-07-03T00:00:00Z PASE：神经符号系统将云故障恢复时间缩短40%以上 https://24-ai.news/zh/news/2026-07-03/arxiv-recontext-long-context-replay/ 24 AI zh 2026-07-03T00:00:00Z ReContext在无需重新训练的情况下改善128K上下文窗口的利用率 https://24-ai.news/zh/news/2026-07-03/arxiv-vrrl-visual-self-reflection/ 24 AI zh 2026-07-03T00:00:00Z VRRL：强化学习迫使视觉模型在自我修正时真正利用图像 https://24-ai.news/zh/news/2026-07-03/pytorch-test-infrastructure/ 24 AI zh 2026-07-03T00:00:00Z PyTorch如何从几个模板类生成数千个测试——测试基础设施指南 https://24-ai.news/ja/news/2026-07-03/amd-agentkernelarena-benchmark/ 24 AI ja 2026-07-03T00:00:00Z AMD AgentKernelArena：AIエージェントのGPUカーネル最適化を測るオープンベンチマーク https://24-ai.news/ja/news/2026-07-03/amd-eagle3-speculative-decoding-quark/ 24 AI ja 2026-07-03T00:00:00Z AMD Eagle3とQuark FP8：投機的デコーディングがMI355Xで最大2.00倍のスループットを実現 https://24-ai.news/ja/news/2026-07-03/amd-yolo26-gpu-resident-pipeline/ 24 AI ja 2026-07-03T00:00:00Z AMD ROCm：GPU常駐型YOLO26パイプラインがデコードから検出まで動画フレームをVRAMに保持 https://24-ai.news/ja/news/2026-07-03/anthropic-claude-code-2-1-200/ 24 AI ja 2026-07-03T00:00:00Z Claude Code v2.1.200：デフォルトのパーミッションモードがManualに変更 https://24-ai.news/ja/news/2026-07-03/arxiv-distributed-attacks-ai-control/ 24 AI ja 2026-07-03T00:00:00Z AIコーディングエージェントへの分散攻撃：既存のいかなるモニターも同時に防御できない https://24-ai.news/ja/news/2026-07-03/arxiv-eo-agents-earth-observation/ 24 AI ja 2026-07-03T00:00:00Z EO-Agents：3エージェントLLMパイプラインがNASA 1,475データセットから160件の科学的仮説を生成 https://24-ai.news/ja/news/2026-07-03/arxiv-online-safety-monitoring-llm/ 24 AI ja 2026-07-03T00:00:00Z シンプルなキャリブレーション済みLLMモニタリングが複雑な逐次的手法を上回る https://24-ai.news/ja/news/2026-07-03/arxiv-pase-cloud-healing-neurosymbolic/ 24 AI ja 2026-07-03T00:00:00Z PASE：ニューロシンボリックシステムがクラウド障害回復時間を40%以上短縮 https://24-ai.news/ja/news/2026-07-03/arxiv-recontext-long-context-replay/ 24 AI ja 2026-07-03T00:00:00Z ReContext：再訓練なしで128Kコンテキストウィンドウの活用を改善 https://24-ai.news/ja/news/2026-07-03/arxiv-vrrl-visual-self-reflection/ 24 AI ja 2026-07-03T00:00:00Z VRRL：強化学習が視覚言語モデルに自己修正時に実際に画像を使わせる https://24-ai.news/ja/news/2026-07-03/pytorch-test-infrastructure/ 24 AI ja 2026-07-03T00:00:00Z PyTorchが少数のテンプレートクラスから数千のテストを生成する仕組み――テストインフラストラクチャガイド https://24-ai.news/ko/news/2026-07-03/amd-agentkernelarena-benchmark/ 24 AI ko 2026-07-03T00:00:00Z AMD AgentKernelArena: GPU 커널 최적화를 위한 AI 에이전트 오픈 벤치마크 https://24-ai.news/ko/news/2026-07-03/amd-eagle3-speculative-decoding-quark/ 24 AI ko 2026-07-03T00:00:00Z AMD Eagle3과 Quark FP8: 투기적 디코딩으로 MI355X에서 최대 2.00배 처리량 달성 https://24-ai.news/ko/news/2026-07-03/amd-yolo26-gpu-resident-pipeline/ 24 AI ko 2026-07-03T00:00:00Z AMD ROCm: GPU 레지던트 YOLO26 파이프라인, 비디오 프레임을 디코딩부터 감지까지 VRAM에 유지 https://24-ai.news/ko/news/2026-07-03/anthropic-claude-code-2-1-200/ 24 AI ko 2026-07-03T00:00:00Z Claude Code v2.1.200: 기본 권한 모드가 Manual로 변경 https://24-ai.news/ko/news/2026-07-03/arxiv-distributed-attacks-ai-control/ 24 AI ko 2026-07-03T00:00:00Z AI 코딩 에이전트에 대한 분산 공격: 어떤 모니터도 동시에 차단하지 못함 https://24-ai.news/ko/news/2026-07-03/arxiv-eo-agents-earth-observation/ 24 AI ko 2026-07-03T00:00:00Z EO-Agents: 3에이전트 LLM 파이프라인이 1,475개 NASA 데이터셋에서 160개의 과학적 가설을 생성 https://24-ai.news/ko/news/2026-07-03/arxiv-online-safety-monitoring-llm/ 24 AI ko 2026-07-03T00:00:00Z 단순 보정 LLM 모니터링이 복잡한 순차적 접근 방식을 능가 https://24-ai.news/ko/news/2026-07-03/arxiv-pase-cloud-healing-neurosymbolic/ 24 AI ko 2026-07-03T00:00:00Z PASE: 뉴로심볼릭 시스템이 클라우드 장애 복구 시간을 40% 이상 단축 https://24-ai.news/ko/news/2026-07-03/arxiv-recontext-long-context-replay/ 24 AI ko 2026-07-03T00:00:00Z ReContext, 재훈련 없이 128K 컨텍스트 창 활용도를 개선 https://24-ai.news/ko/news/2026-07-03/arxiv-vrrl-visual-self-reflection/ 24 AI ko 2026-07-03T00:00:00Z VRRL: 강화 학습이 시각 모델이 자기 수정 시 이미지를 실제로 활용하도록 강제 https://24-ai.news/ko/news/2026-07-03/pytorch-test-infrastructure/ 24 AI ko 2026-07-03T00:00:00Z PyTorch가 몇 가지 템플릿 클래스에서 수천 개의 테스트를 생성하는 방법 — 테스트 인프라 가이드 https://24-ai.news/hr/news/2026-07-02/aisi-test-time-compute-agent-evals/ 24 AI hr 2026-07-02T00:00:00Z Britanski AISI: Evaluacije AI agenata moraju uzeti u obzir budžet računanja https://24-ai.news/hr/news/2026-07-02/arxiv-automem-memory-cognitive-skill/ 24 AI hr 2026-07-02T00:00:00Z AutoMem: memorijsko upravljanje kao naučiva vještina, ne arhitekturni izbor https://24-ai.news/hr/news/2026-07-02/arxiv-harc-harmfulness-refusal-coupling/ 24 AI hr 2026-07-02T00:00:00Z HARC: Nova metoda fine-tuninga koja sprječava jailbreak uparivanjem štetnosti i odbijanja https://24-ai.news/hr/news/2026-07-02/arxiv-openworld-agent-fragility-paft/ 24 AI hr 2026-07-02T00:00:00Z Istraživanje ICML 2026: SFT i RL agenti dramatično gube na performansama izvan kontroliranog benchmarka https://24-ai.news/hr/news/2026-07-02/arxiv-sea-self-evolving-agents-certificates/ 24 AI hr 2026-07-02T00:00:00Z SEA: Agenti koji se samo-modificiraju uz formalna sigurnosna jamstva u realnom vremenu https://24-ai.news/hr/news/2026-07-02/arxiv-two-ai-metrics-diverged/ 24 AI hr 2026-07-02T00:00:00Z Dva AI mjerila razišla se — hoće li to biti odlučujuće? https://24-ai.news/hr/news/2026-07-02/aws-bedrock-ai-phishing-detection/ 24 AI hr 2026-07-02T00:00:00Z Amazon Bedrock otkriva AI-generirani phishing analizom ponašanja sadržaja https://24-ai.news/hr/news/2026-07-02/aws-sagemaker-multiturn-rl-best-practices/ 24 AI hr 2026-07-02T00:00:00Z AWS SageMaker vodič: u multi-turn RL nagrada i evaluacija važniji od algoritma https://24-ai.news/hr/news/2026-07-02/github-copilot-agent-session-streaming/ 24 AI hr 2026-07-02T00:00:00Z GitHub Copilot agenti dobivaju streaming sesija u javnoj pretpregledu https://24-ai.news/hr/news/2026-07-02/github-copilot-cli-cost-controls/ 24 AI hr 2026-07-02T00:00:00Z GitHub Copilot u CI/CD-u: bez osobnih tokena i s kontrolom troškova po timu https://24-ai.news/hr/news/2026-07-02/langchain-coding-agent-cost-langsmith/ 24 AI hr 2026-07-02T00:00:00Z Kako obuzdati eksploziju troškova coding-agenata: četverofazni pristup LangSmithom https://24-ai.news/en/news/2026-07-02/aisi-test-time-compute-agent-evals/ 24 AI en 2026-07-02T00:00:00Z UK AISI: Agent Evaluations Must Account for Compute Budget https://24-ai.news/en/news/2026-07-02/arxiv-automem-memory-cognitive-skill/ 24 AI en 2026-07-02T00:00:00Z AutoMem: Memory Management as a Learnable Skill, Not an Architectural Choice https://24-ai.news/en/news/2026-07-02/arxiv-harc-harmfulness-refusal-coupling/ 24 AI en 2026-07-02T00:00:00Z HARC: A New Fine-Tuning Method That Prevents Jailbreaks by Coupling Harmfulness and Refusal https://24-ai.news/en/news/2026-07-02/arxiv-openworld-agent-fragility-paft/ 24 AI en 2026-07-02T00:00:00Z ICML 2026 Research: SFT and RL Agents Suffer Dramatic Performance Drops Outside Controlled Benchmarks https://24-ai.news/en/news/2026-07-02/arxiv-sea-self-evolving-agents-certificates/ 24 AI en 2026-07-02T00:00:00Z SEA: Agents That Self-Modify With Formal Safety Guarantees in Real Time https://24-ai.news/en/news/2026-07-02/arxiv-two-ai-metrics-diverged/ 24 AI en 2026-07-02T00:00:00Z Two AI Metrics Diverged — Will It Make All the Difference? https://24-ai.news/en/news/2026-07-02/aws-bedrock-ai-phishing-detection/ 24 AI en 2026-07-02T00:00:00Z Amazon Bedrock Detects AI-Generated Phishing Through Content Behavior Analysis https://24-ai.news/en/news/2026-07-02/aws-sagemaker-multiturn-rl-best-practices/ 24 AI en 2026-07-02T00:00:00Z AWS SageMaker Guide: In Multi-Turn RL, Reward and Evaluation Matter More Than the Algorithm https://24-ai.news/en/news/2026-07-02/github-copilot-agent-session-streaming/ 24 AI en 2026-07-02T00:00:00Z GitHub Copilot Agents Get Session Streaming in Public Preview https://24-ai.news/en/news/2026-07-02/github-copilot-cli-cost-controls/ 24 AI en 2026-07-02T00:00:00Z GitHub Copilot in CI/CD: No Personal Tokens and Team-Level Cost Controls https://24-ai.news/en/news/2026-07-02/langchain-coding-agent-cost-langsmith/ 24 AI en 2026-07-02T00:00:00Z How to Tame Exploding Coding Agent Costs: A Four-Phase Approach with LangSmith https://24-ai.news/de/news/2026-07-02/aisi-test-time-compute-agent-evals/ 24 AI de 2026-07-02T00:00:00Z Britisches AISI: KI-Agenten-Evaluierungen müssen Rechenbudget berücksichtigen https://24-ai.news/de/news/2026-07-02/arxiv-automem-memory-cognitive-skill/ 24 AI de 2026-07-02T00:00:00Z AutoMem: Gedächtnismanagement als erlernbare Fähigkeit, nicht als Architekturentscheidung https://24-ai.news/de/news/2026-07-02/arxiv-harc-harmfulness-refusal-coupling/ 24 AI de 2026-07-02T00:00:00Z HARC: Neue Fine-Tuning-Methode verhindert Jailbreaks durch Kopplung von Schädlichkeit und Ablehnung https://24-ai.news/de/news/2026-07-02/arxiv-openworld-agent-fragility-paft/ 24 AI de 2026-07-02T00:00:00Z ICML 2026-Studie: SFT- und RL-Agenten verlieren dramatisch an Leistung außerhalb kontrollierter Benchmarks https://24-ai.news/de/news/2026-07-02/arxiv-sea-self-evolving-agents-certificates/ 24 AI de 2026-07-02T00:00:00Z SEA: Agenten, die sich selbst modifizieren – mit formalen Sicherheitsgarantien in Echtzeit https://24-ai.news/de/news/2026-07-02/arxiv-two-ai-metrics-diverged/ 24 AI de 2026-07-02T00:00:00Z Zwei KI-Metriken divergierten – wird das den entscheidenden Unterschied machen? https://24-ai.news/de/news/2026-07-02/aws-bedrock-ai-phishing-detection/ 24 AI de 2026-07-02T00:00:00Z Amazon Bedrock erkennt KI-generierten Phishing durch Verhaltensanalyse https://24-ai.news/de/news/2026-07-02/aws-sagemaker-multiturn-rl-best-practices/ 24 AI de 2026-07-02T00:00:00Z AWS SageMaker-Leitfaden: Bei Multi-Turn-RL sind Belohnungsfunktion und Evaluierung wichtiger als der Algorithmus https://24-ai.news/de/news/2026-07-02/github-copilot-agent-session-streaming/ 24 AI de 2026-07-02T00:00:00Z GitHub Copilot-Agenten erhalten Session-Streaming in der öffentlichen Vorschau https://24-ai.news/de/news/2026-07-02/github-copilot-cli-cost-controls/ 24 AI de 2026-07-02T00:00:00Z GitHub Copilot in CI/CD: Kein persönlicher Token mehr und Kostenkontrolle pro Team https://24-ai.news/de/news/2026-07-02/langchain-coding-agent-cost-langsmith/ 24 AI de 2026-07-02T00:00:00Z Wie man Kostenexplosionen bei Coding-Agenten eindämmt: Vierphasiger Ansatz mit LangSmith https://24-ai.news/zh/news/2026-07-02/aisi-test-time-compute-agent-evals/ 24 AI zh 2026-07-02T00:00:00Z 英国AISI：AI智能体评估必须考虑计算预算 https://24-ai.news/zh/news/2026-07-02/arxiv-automem-memory-cognitive-skill/ 24 AI zh 2026-07-02T00:00:00Z AutoMem：将记忆管理作为可学习技能而非架构选择 https://24-ai.news/zh/news/2026-07-02/arxiv-harc-harmfulness-refusal-coupling/ 24 AI zh 2026-07-02T00:00:00Z HARC：通过耦合有害性与拒绝方向防止越狱攻击的新型微调方法 https://24-ai.news/zh/news/2026-07-02/arxiv-openworld-agent-fragility-paft/ 24 AI zh 2026-07-02T00:00:00Z ICML 2026研究：SFT与RL智能体在受控基准之外性能大幅下降 https://24-ai.news/zh/news/2026-07-02/arxiv-sea-self-evolving-agents-certificates/ 24 AI zh 2026-07-02T00:00:00Z SEA：具备实时形式化安全保证的自我修改智能体 https://24-ai.news/zh/news/2026-07-02/arxiv-two-ai-metrics-diverged/ 24 AI zh 2026-07-02T00:00:00Z 两种AI衡量标准分歧——这将带来决定性影响吗？ https://24-ai.news/zh/news/2026-07-02/aws-bedrock-ai-phishing-detection/ 24 AI zh 2026-07-02T00:00:00Z Amazon Bedrock通过内容行为分析识别AI生成的网络钓鱼 https://24-ai.news/zh/news/2026-07-02/aws-sagemaker-multiturn-rl-best-practices/ 24 AI zh 2026-07-02T00:00:00Z AWS SageMaker指南：多轮强化学习中奖励与评估比算法更重要 https://24-ai.news/zh/news/2026-07-02/github-copilot-agent-session-streaming/ 24 AI zh 2026-07-02T00:00:00Z GitHub Copilot智能体获得公开预览版会话流传输功能 https://24-ai.news/zh/news/2026-07-02/github-copilot-cli-cost-controls/ 24 AI zh 2026-07-02T00:00:00Z GitHub Copilot在CI/CD中：无需个人令牌，支持按团队成本控制 https://24-ai.news/zh/news/2026-07-02/langchain-coding-agent-cost-langsmith/ 24 AI zh 2026-07-02T00:00:00Z 如何遏制编码智能体的成本爆炸：LangSmith四阶段方法 https://24-ai.news/ja/news/2026-07-02/aisi-test-time-compute-agent-evals/ 24 AI ja 2026-07-02T00:00:00Z 英国AISI：AIエージェント評価はテスト時計算リソースを考慮すべき https://24-ai.news/ja/news/2026-07-02/arxiv-automem-memory-cognitive-skill/ 24 AI ja 2026-07-02T00:00:00Z AutoMem：人手なしに記憶管理を学習する認知スキルとしてのメモリ https://24-ai.news/ja/news/2026-07-02/arxiv-harc-harmfulness-refusal-coupling/ 24 AI ja 2026-07-02T00:00:00Z HARC：有害性と拒否の方向を結合したジェイルブレーク耐性ファインチューニング手法 https://24-ai.news/ja/news/2026-07-02/arxiv-openworld-agent-fragility-paft/ 24 AI ja 2026-07-02T00:00:00Z ICML 2026研究：SFTとRLエージェントは制御されたベンチマーク外で劇的に性能が低下 https://24-ai.news/ja/news/2026-07-02/arxiv-sea-self-evolving-agents-certificates/ 24 AI ja 2026-07-02T00:00:00Z SEA：リアルタイムの形式的安全保証を持つ自己修正エージェント https://24-ai.news/ja/news/2026-07-02/arxiv-two-ai-metrics-diverged/ 24 AI ja 2026-07-02T00:00:00Z 2つのAI指標が分岐した——これは決定的な問題になるか？ https://24-ai.news/ja/news/2026-07-02/aws-bedrock-ai-phishing-detection/ 24 AI ja 2026-07-02T00:00:00Z Amazon Bedrockがコンテンツ行動分析でAI生成フィッシングを検出 https://24-ai.news/ja/news/2026-07-02/aws-sagemaker-multiturn-rl-best-practices/ 24 AI ja 2026-07-02T00:00:00Z AWS SageMakerガイド：マルチターンRLでは報酬と評価がアルゴリズムより重要 https://24-ai.news/ja/news/2026-07-02/github-copilot-agent-session-streaming/ 24 AI ja 2026-07-02T00:00:00Z GitHub Copilotエージェントがパブリックプレビューでセッションストリーミングを取得 https://24-ai.news/ja/news/2026-07-02/github-copilot-cli-cost-controls/ 24 AI ja 2026-07-02T00:00:00Z GitHub CopilotのCI/CD：個人トークン不要、チーム別コスト制御 https://24-ai.news/ja/news/2026-07-02/langchain-coding-agent-cost-langsmith/ 24 AI ja 2026-07-02T00:00:00Z コーディングエージェントのコスト爆発を抑える：LangSmithによる4段階アプローチ https://24-ai.news/ko/news/2026-07-02/aisi-test-time-compute-agent-evals/ 24 AI ko 2026-07-02T00:00:00Z 영국 AISI: AI 에이전트 평가는 컴퓨팅 예산을 반드시 고려해야 한다 https://24-ai.news/ko/news/2026-07-02/arxiv-automem-memory-cognitive-skill/ 24 AI ko 2026-07-02T00:00:00Z AutoMem: 메모리 관리를 아키텍처적 선택이 아닌 학습 가능한 기술로 https://24-ai.news/ko/news/2026-07-02/arxiv-harc-harmfulness-refusal-coupling/ 24 AI ko 2026-07-02T00:00:00Z HARC: 유해성과 거부 방향 결합으로 탈옥 방지하는 새로운 파인튜닝 방법 https://24-ai.news/ko/news/2026-07-02/arxiv-openworld-agent-fragility-paft/ 24 AI ko 2026-07-02T00:00:00Z ICML 2026 연구: SFT 및 RL 에이전트, 통제된 벤치마크 외부에서 성능 급락 https://24-ai.news/ko/news/2026-07-02/arxiv-sea-self-evolving-agents-certificates/ 24 AI ko 2026-07-02T00:00:00Z SEA: 실시간 공식 안전 보장을 갖춘 자기 수정 에이전트 https://24-ai.news/ko/news/2026-07-02/arxiv-two-ai-metrics-diverged/ 24 AI ko 2026-07-02T00:00:00Z 두 AI 지표가 엇갈렸다 — 그것이 결정적인 차이를 만들까? https://24-ai.news/ko/news/2026-07-02/aws-bedrock-ai-phishing-detection/ 24 AI ko 2026-07-02T00:00:00Z Amazon Bedrock, 콘텐츠 행동 분석으로 AI 생성 피싱 탐지 https://24-ai.news/ko/news/2026-07-02/aws-sagemaker-multiturn-rl-best-practices/ 24 AI ko 2026-07-02T00:00:00Z AWS SageMaker 가이드: 멀티턴 RL에서 알고리즘보다 보상과 평가가 중요하다 https://24-ai.news/ko/news/2026-07-02/github-copilot-agent-session-streaming/ 24 AI ko 2026-07-02T00:00:00Z GitHub Copilot 에이전트, 퍼블릭 프리뷰에서 세션 스트리밍 지원 시작 https://24-ai.news/ko/news/2026-07-02/github-copilot-cli-cost-controls/ 24 AI ko 2026-07-02T00:00:00Z GitHub Copilot CI/CD: 개인 토큰 불필요, 팀별 비용 통제 지원 https://24-ai.news/ko/news/2026-07-02/langchain-coding-agent-cost-langsmith/ 24 AI ko 2026-07-02T00:00:00Z 코딩 에이전트 비용 폭발을 억제하는 방법: LangSmith 4단계 접근법 https://24-ai.news/hr/news/2026-07-01/anthropic-claude-code-2-1-198/ 24 AI hr 2026-07-01T00:00:00Z Claude Code v2.1.198: background agenti sada sami otvaraju PR-ove, stigao /dataviz skill https://24-ai.news/hr/news/2026-07-01/anthropic-enterprise-model-entitlements/ 24 AI hr 2026-07-01T00:00:00Z Anthropic uvodi kontrolu pristupa modelima za enterprise administratore https://24-ai.news/hr/news/2026-07-01/anthropic-fable-5-mythos-5-restored/ 24 AI hr 2026-07-01T00:00:00Z Fable 5 i Mythos 5 ponovno dostupni: Anthropic vratio modele korisnicima https://24-ai.news/hr/news/2026-07-01/aws-a2a-gateway-serverless/ 24 AI hr 2026-07-01T00:00:00Z AWS objavljuje serverless A2A gateway koji zamjenjuje 190 point-to-point veza središnjim registrom https://24-ai.news/hr/news/2026-07-01/aws-agentcore-memory-metadata-filtering/ 24 AI hr 2026-07-01T00:00:00Z AWS AgentCore Memory dobiva filtriranje metapodacima — točnost skočila s 40% na 64% https://24-ai.news/hr/news/2026-07-01/aws-nemotron-gptoss-bedrock-govcloud/ 24 AI hr 2026-07-01T00:00:00Z NVIDIA Nemotron i OpenAI GPT OSS modeli dostupni u AWS GovCloudu s FedRAMP High certifikatom https://24-ai.news/hr/news/2026-07-01/github-copilot-kimi-k2-7/ 24 AI hr 2026-07-01T00:00:00Z Kimi K2.7 Code stigao u GitHub Copilot kao prvi open-weight model https://24-ai.news/hr/news/2026-07-01/github-copilot-vision-browser-tools-ga/ 24 AI hr 2026-07-01T00:00:00Z GitHub Copilot Vision i browser tools: dvije GA sposobnosti u jednom danu https://24-ai.news/hr/news/2026-07-01/github-models-retirement-july-2026/ 24 AI hr 2026-07-01T00:00:00Z GitHub Models se potpuno gasi 30. srpnja 2026. https://24-ai.news/hr/news/2026-07-01/github-secret-scanning-public-monitoring/ 24 AI hr 2026-07-01T00:00:00Z GitHub proširuje secret scanning na cijelu javnu površinu GitHuba za enterprise korisnike https://24-ai.news/hr/news/2026-07-01/langchain-openwiki-repo-docs/ 24 AI hr 2026-07-01T00:00:00Z LangChain objavljuje OpenWiki: open-source alat koji automatski dokumentira repozitorije za AI agente https://24-ai.news/hr/news/2026-07-01/langchain-rlm-deep-agents/ 24 AI hr 2026-07-01T00:00:00Z LangChain uvodi RLM agente: rekurzivni modeli postižu 79% bolje rezultate na dugim kontekstima https://24-ai.news/en/news/2026-07-01/anthropic-claude-code-2-1-198/ 24 AI en 2026-07-01T00:00:00Z Claude Code v2.1.198: Background Agents Now Open PRs Automatically, /dataviz Skill Arrives https://24-ai.news/en/news/2026-07-01/anthropic-enterprise-model-entitlements/ 24 AI en 2026-07-01T00:00:00Z Anthropic Introduces Model Access Control for Enterprise Administrators https://24-ai.news/en/news/2026-07-01/anthropic-fable-5-mythos-5-restored/ 24 AI en 2026-07-01T00:00:00Z Fable 5 and Mythos 5 Back Online: Anthropic Restores Model Access https://24-ai.news/en/news/2026-07-01/aws-a2a-gateway-serverless/ 24 AI en 2026-07-01T00:00:00Z AWS Releases Serverless A2A Gateway Replacing 190 Point-to-Point Connections with a Central Registry https://24-ai.news/en/news/2026-07-01/aws-agentcore-memory-metadata-filtering/ 24 AI en 2026-07-01T00:00:00Z AWS AgentCore Memory Gains Metadata Filtering — Accuracy Jumps from 40% to 64% https://24-ai.news/en/news/2026-07-01/aws-nemotron-gptoss-bedrock-govcloud/ 24 AI en 2026-07-01T00:00:00Z NVIDIA Nemotron and OpenAI GPT OSS Models Available in AWS GovCloud with FedRAMP High Certification https://24-ai.news/en/news/2026-07-01/github-copilot-kimi-k2-7/ 24 AI en 2026-07-01T00:00:00Z Kimi K2.7 Code Arrives in GitHub Copilot as the First Open-Weight Model https://24-ai.news/en/news/2026-07-01/github-copilot-vision-browser-tools-ga/ 24 AI en 2026-07-01T00:00:00Z GitHub Copilot Vision and Browser Tools: Two GA Capabilities in One Day https://24-ai.news/en/news/2026-07-01/github-models-retirement-july-2026/ 24 AI en 2026-07-01T00:00:00Z GitHub Models Fully Shuts Down July 30, 2026 https://24-ai.news/en/news/2026-07-01/github-secret-scanning-public-monitoring/ 24 AI en 2026-07-01T00:00:00Z GitHub Expands Secret Scanning to the Entire Public GitHub Surface for Enterprise Users https://24-ai.news/en/news/2026-07-01/langchain-openwiki-repo-docs/ 24 AI en 2026-07-01T00:00:00Z LangChain Releases OpenWiki: Open-Source Tool That Automatically Documents Repositories for AI Agents https://24-ai.news/en/news/2026-07-01/langchain-rlm-deep-agents/ 24 AI en 2026-07-01T00:00:00Z LangChain Introduces RLM Agents: Recursive Models Achieve 79% Better Results on Long Contexts https://24-ai.news/de/news/2026-07-01/anthropic-claude-code-2-1-198/ 24 AI de 2026-07-01T00:00:00Z Claude Code v2.1.198: Hintergrund-Agenten öffnen jetzt selbst PRs, neuer /dataviz-Skill https://24-ai.news/de/news/2026-07-01/anthropic-enterprise-model-entitlements/ 24 AI de 2026-07-01T00:00:00Z Anthropic führt Zugriffskontrolle für Modelle für Enterprise-Administratoren ein https://24-ai.news/de/news/2026-07-01/anthropic-fable-5-mythos-5-restored/ 24 AI de 2026-07-01T00:00:00Z Fable 5 und Mythos 5 wieder verfügbar: Anthropic stellt Modellzugang wieder her https://24-ai.news/de/news/2026-07-01/aws-a2a-gateway-serverless/ 24 AI de 2026-07-01T00:00:00Z AWS veröffentlicht serverlosen A2A-Gateway, der 190 Punkt-zu-Punkt-Verbindungen durch eine zentrale Registry ersetzt https://24-ai.news/de/news/2026-07-01/aws-agentcore-memory-metadata-filtering/ 24 AI de 2026-07-01T00:00:00Z AWS AgentCore Memory erhält Metadaten-Filterung — Genauigkeit steigt von 40 % auf 64 % https://24-ai.news/de/news/2026-07-01/aws-nemotron-gptoss-bedrock-govcloud/ 24 AI de 2026-07-01T00:00:00Z NVIDIA Nemotron und OpenAI GPT OSS-Modelle in AWS GovCloud mit FedRAMP High-Zertifikat verfügbar https://24-ai.news/de/news/2026-07-01/github-copilot-kimi-k2-7/ 24 AI de 2026-07-01T00:00:00Z Kimi K2.7 Code ist jetzt in GitHub Copilot — als erstes Open-Weight-Modell https://24-ai.news/de/news/2026-07-01/github-copilot-vision-browser-tools-ga/ 24 AI de 2026-07-01T00:00:00Z GitHub Copilot Vision und Browser-Tools: zwei GA-Fähigkeiten an einem Tag https://24-ai.news/de/news/2026-07-01/github-models-retirement-july-2026/ 24 AI de 2026-07-01T00:00:00Z GitHub Models wird am 30. Juli 2026 vollständig eingestellt https://24-ai.news/de/news/2026-07-01/github-secret-scanning-public-monitoring/ 24 AI de 2026-07-01T00:00:00Z GitHub erweitert Secret Scanning auf die gesamte öffentliche GitHub-Oberfläche für Enterprise-Nutzer https://24-ai.news/de/news/2026-07-01/langchain-openwiki-repo-docs/ 24 AI de 2026-07-01T00:00:00Z LangChain veröffentlicht OpenWiki: Open-Source-Tool, das Repositories automatisch für KI-Agenten dokumentiert https://24-ai.news/de/news/2026-07-01/langchain-rlm-deep-agents/ 24 AI de 2026-07-01T00:00:00Z LangChain führt RLM-Agenten ein: Rekursive Modelle erzielen 79 % bessere Ergebnisse bei langen Kontexten https://24-ai.news/zh/news/2026-07-01/anthropic-claude-code-2-1-198/ 24 AI zh 2026-07-01T00:00:00Z Claude Code v2.1.198：后台智能体现可自动创建拉取请求，新增 /dataviz 技能 https://24-ai.news/zh/news/2026-07-01/anthropic-enterprise-model-entitlements/ 24 AI zh 2026-07-01T00:00:00Z Anthropic 为企业管理员引入模型访问控制 https://24-ai.news/zh/news/2026-07-01/anthropic-fable-5-mythos-5-restored/ 24 AI zh 2026-07-01T00:00:00Z Fable 5与Mythos 5重返服务：Anthropic为用户恢复模型访问权限 https://24-ai.news/zh/news/2026-07-01/aws-a2a-gateway-serverless/ 24 AI zh 2026-07-01T00:00:00Z AWS 发布无服务器 A2A 网关，用中央注册表替换 190 条点对点连接 https://24-ai.news/zh/news/2026-07-01/aws-agentcore-memory-metadata-filtering/ 24 AI zh 2026-07-01T00:00:00Z AWS AgentCore Memory 引入元数据过滤——准确率从 40% 跃升至 64% https://24-ai.news/zh/news/2026-07-01/aws-nemotron-gptoss-bedrock-govcloud/ 24 AI zh 2026-07-01T00:00:00Z NVIDIA Nemotron 和 OpenAI GPT OSS 模型在 AWS GovCloud 上获得 FedRAMP High 认证 https://24-ai.news/zh/news/2026-07-01/github-copilot-kimi-k2-7/ 24 AI zh 2026-07-01T00:00:00Z Kimi K2.7 Code 作为首款开放权重模型进入 GitHub Copilot https://24-ai.news/zh/news/2026-07-01/github-copilot-vision-browser-tools-ga/ 24 AI zh 2026-07-01T00:00:00Z GitHub Copilot Vision 和浏览器工具：同日两项功能正式发布 https://24-ai.news/zh/news/2026-07-01/github-models-retirement-july-2026/ 24 AI zh 2026-07-01T00:00:00Z GitHub Models 将于 2026 年 7 月 30 日完全关闭 https://24-ai.news/zh/news/2026-07-01/github-secret-scanning-public-monitoring/ 24 AI zh 2026-07-01T00:00:00Z GitHub 为企业用户将密钥扫描扩展至整个公共 GitHub 平台 https://24-ai.news/zh/news/2026-07-01/langchain-openwiki-repo-docs/ 24 AI zh 2026-07-01T00:00:00Z LangChain 发布 OpenWiki：自动为 AI 智能体生成和维护代码库文档的开源工具 https://24-ai.news/zh/news/2026-07-01/langchain-rlm-deep-agents/ 24 AI zh 2026-07-01T00:00:00Z LangChain 推出 RLM 智能体：递归模型在长上下文中实现 79% 的性能提升 https://24-ai.news/ja/news/2026-07-01/anthropic-claude-code-2-1-198/ 24 AI ja 2026-07-01T00:00:00Z Claude Code v2.1.198: バックグラウンドエージェントが自動でPRを作成、/datavizスキル追加 https://24-ai.news/ja/news/2026-07-01/anthropic-enterprise-model-entitlements/ 24 AI ja 2026-07-01T00:00:00Z Anthropicがエンタープライズ管理者向けモデルアクセス制御を導入 https://24-ai.news/ja/news/2026-07-01/anthropic-fable-5-mythos-5-restored/ 24 AI ja 2026-07-01T00:00:00Z Fable 5とMythos 5が復活：AnthropicがClaude両モデルのアクセスを再開 https://24-ai.news/ja/news/2026-07-01/aws-a2a-gateway-serverless/ 24 AI ja 2026-07-01T00:00:00Z AWSが190のポイントツーポイント接続を中央レジストリに置き換えるサーバーレスA2Aゲートウェイを発表 https://24-ai.news/ja/news/2026-07-01/aws-agentcore-memory-metadata-filtering/ 24 AI ja 2026-07-01T00:00:00Z AWS AgentCore Memoryがメタデータフィルタリングを導入——精度が40%から64%に向上 https://24-ai.news/ja/news/2026-07-01/aws-nemotron-gptoss-bedrock-govcloud/ 24 AI ja 2026-07-01T00:00:00Z NVIDIA NemotronとOpenAI GPT OSSモデルがFedRAMP High認定でAWS GovCloudに登場 https://24-ai.news/ja/news/2026-07-01/github-copilot-kimi-k2-7/ 24 AI ja 2026-07-01T00:00:00Z Kimi K2.7 CodeがGitHub Copilotに登場——初のオープンウェイトモデル https://24-ai.news/ja/news/2026-07-01/github-copilot-vision-browser-tools-ga/ 24 AI ja 2026-07-01T00:00:00Z GitHub Copilot VisionとBrowser Tools：1日に2つのGA機能が追加 https://24-ai.news/ja/news/2026-07-01/github-models-retirement-july-2026/ 24 AI ja 2026-07-01T00:00:00Z GitHub Modelsが2026年7月30日に完全廃止 https://24-ai.news/ja/news/2026-07-01/github-secret-scanning-public-monitoring/ 24 AI ja 2026-07-01T00:00:00Z GitHubがエンタープライズユーザー向けにシークレットスキャニングをGitHub公開面全体に拡張 https://24-ai.news/ja/news/2026-07-01/langchain-openwiki-repo-docs/ 24 AI ja 2026-07-01T00:00:00Z LangChainがOpenWikiを公開：AIエージェント向けにリポジトリを自動ドキュメント化するオープンソースツール https://24-ai.news/ja/news/2026-07-01/langchain-rlm-deep-agents/ 24 AI ja 2026-07-01T00:00:00Z LangChainがRLMエージェントを導入：再帰モデルが長いコンテキストで79%高い精度を達成 https://24-ai.news/ko/news/2026-07-01/anthropic-claude-code-2-1-198/ 24 AI ko 2026-07-01T00:00:00Z Claude Code v2.1.198: 백그라운드 에이전트가 스스로 PR을 열고, /dataviz 스킬 추가 https://24-ai.news/ko/news/2026-07-01/anthropic-enterprise-model-entitlements/ 24 AI ko 2026-07-01T00:00:00Z Anthropic, 엔터프라이즈 관리자를 위한 모델 접근 제어 도입 https://24-ai.news/ko/news/2026-07-01/anthropic-fable-5-mythos-5-restored/ 24 AI ko 2026-07-01T00:00:00Z Fable 5와 Mythos 5 재출시: Anthropic, 두 모델 접근권 복원 https://24-ai.news/ko/news/2026-07-01/aws-a2a-gateway-serverless/ 24 AI ko 2026-07-01T00:00:00Z AWS, 190개 포인트-투-포인트 연결을 중앙 레지스트리 하나로 대체하는 서버리스 A2A 게이트웨이 발표 https://24-ai.news/ko/news/2026-07-01/aws-agentcore-memory-metadata-filtering/ 24 AI ko 2026-07-01T00:00:00Z AWS AgentCore Memory에 메타데이터 필터링 추가 — 정확도 40%에서 64%로 상승 https://24-ai.news/ko/news/2026-07-01/aws-nemotron-gptoss-bedrock-govcloud/ 24 AI ko 2026-07-01T00:00:00Z NVIDIA Nemotron과 OpenAI GPT OSS 모델, FedRAMP High 인증과 함께 AWS GovCloud에서 가용 https://24-ai.news/ko/news/2026-07-01/github-copilot-kimi-k2-7/ 24 AI ko 2026-07-01T00:00:00Z Kimi K2.7 Code, GitHub Copilot 최초의 오픈 웨이트 모델로 등장 https://24-ai.news/ko/news/2026-07-01/github-copilot-vision-browser-tools-ga/ 24 AI ko 2026-07-01T00:00:00Z GitHub Copilot Vision과 브라우저 도구: 하루 만에 두 가지 GA 기능 출시 https://24-ai.news/ko/news/2026-07-01/github-models-retirement-july-2026/ 24 AI ko 2026-07-01T00:00:00Z GitHub Models, 2026년 7월 30일 완전 종료 https://24-ai.news/ko/news/2026-07-01/github-secret-scanning-public-monitoring/ 24 AI ko 2026-07-01T00:00:00Z GitHub, 엔터프라이즈 사용자를 위해 전체 공개 GitHub 표면으로 시크릿 스캐닝 확대 https://24-ai.news/ko/news/2026-07-01/langchain-openwiki-repo-docs/ 24 AI ko 2026-07-01T00:00:00Z LangChain, 저장소를 AI 에이전트용으로 자동 문서화하는 오픈소스 도구 OpenWiki 출시 https://24-ai.news/ko/news/2026-07-01/langchain-rlm-deep-agents/ 24 AI ko 2026-07-01T00:00:00Z LangChain이 RLM 에이전트 도입: 재귀적 모델이 긴 컨텍스트에서 79% 향상된 성능 달성