https://24-ai.news/hr/news/2026-05-02/arxiv-exploration-hacking-rl-resistance/ 24 AI hr 2026-05-02T00:00:00Z Exploration Hacking: mogu li LLM-ovi naučiti opirati se RL treningu i strategijski potiskivati vlastite sposobnosti? https://24-ai.news/hr/news/2026-05-02/arxiv-kellybench-premier-league-decisions/ 24 AI hr 2026-05-02T00:00:00Z KellyBench: AI agenti upravljali kladioničarskim bankrollom u Premier Ligi — svi vodeći modeli izgubili novac https://24-ai.news/hr/news/2026-05-02/arxiv-latent-grpo-reasoning-rl/ 24 AI hr 2026-05-02T00:00:00Z Latent-GRPO: stabilna RL optimizacija za latent reasoning — 7,86 boda na GSM8K-Aug i 4,27 boda na AIME uz 3-4× kraće reasoning chain-ove https://24-ai.news/hr/news/2026-05-02/arxiv-mcphunt-mcp-credential-propagation/ 24 AI hr 2026-05-02T00:00:00Z MCPHunt: prvi benchmark koji mjeri curenje vjerodajnica između granica povjerenja u multi-server MCP agentima — stope 11,5–41,3 % https://24-ai.news/hr/news/2026-05-02/arxiv-msr-synthetic-computers-1000-scale/ 24 AI hr 2026-05-02T00:00:00Z Microsoft Research Synthetic Computers: 1 000 sintetičkih računala kao supstrat za long-horizon trening produktivnih AI agenata https://24-ai.news/hr/news/2026-05-02/aws-transform-bi-migration-bedrock-agentcore/ 24 AI hr 2026-05-02T00:00:00Z AWS Transform automatizira migraciju BI dashboarda iz Tableau i Power BI u QuickSight za dane umjesto mjeseci https://24-ai.news/hr/news/2026-05-02/github-copilot-gpt-52-deprecation-june-2026/ 24 AI hr 2026-05-02T00:00:00Z GitHub povlači GPT-5.2 i GPT-5.2-Codex iz Copilota 1. lipnja 2026. — migracija na GPT-5.5 i GPT-5.3-Codex https://24-ai.news/hr/news/2026-05-02/google-research-open-science-250k-reach/ 24 AI hr 2026-05-02T00:00:00Z Google Research open-source alati dosegnuli 250.000 istraživača: od genoma do monsunskih prognoza za 38 milijuna farmera https://24-ai.news/hr/news/2026-05-02/nist-caisi-deepseek-v4-pro-evaluation/ 24 AI hr 2026-05-02T00:00:00Z NIST CAISI evaluacija DeepSeek V4 Pro: 8 mjeseci zaostatka za frontier US modelima na 9 benchmarka u 5 domena https://24-ai.news/en/news/2026-05-02/arxiv-exploration-hacking-rl-resistance/ 24 AI en 2026-05-02T00:00:00Z Exploration Hacking: Can LLMs Learn to Resist RL Training and Strategically Suppress Their Own Capabilities? https://24-ai.news/en/news/2026-05-02/arxiv-kellybench-premier-league-decisions/ 24 AI en 2026-05-02T00:00:00Z KellyBench: AI agents managing a betting bankroll through the Premier League season — all leading models lost money https://24-ai.news/en/news/2026-05-02/arxiv-latent-grpo-reasoning-rl/ 24 AI en 2026-05-02T00:00:00Z Latent-GRPO: Stable RL Optimization for Latent Reasoning — 7.86 Points on GSM8K-Aug and 4.27 Points on AIME With 3-4× Shorter Reasoning Chains https://24-ai.news/en/news/2026-05-02/arxiv-mcphunt-mcp-credential-propagation/ 24 AI en 2026-05-02T00:00:00Z MCPHunt: first benchmark measuring credential leakage across trust boundaries in multi-server MCP agents — rates of 11.5–41.3% https://24-ai.news/en/news/2026-05-02/arxiv-msr-synthetic-computers-1000-scale/ 24 AI en 2026-05-02T00:00:00Z Microsoft Research Synthetic Computers: 1,000 synthetic computers as a substrate for long-horizon training of productive AI agents https://24-ai.news/en/news/2026-05-02/aws-transform-bi-migration-bedrock-agentcore/ 24 AI en 2026-05-02T00:00:00Z AWS Transform automates BI dashboard migration from Tableau and Power BI to QuickSight in days instead of months https://24-ai.news/en/news/2026-05-02/github-copilot-gpt-52-deprecation-june-2026/ 24 AI en 2026-05-02T00:00:00Z GitHub is retiring GPT-5.2 and GPT-5.2-Codex from Copilot on June 1, 2026 — migration to GPT-5.5 and GPT-5.3-Codex https://24-ai.news/en/news/2026-05-02/google-research-open-science-250k-reach/ 24 AI en 2026-05-02T00:00:00Z Google Research open-source tools reach 250,000 researchers: from genomes to monsoon forecasts for 38 million farmers https://24-ai.news/en/news/2026-05-02/nist-caisi-deepseek-v4-pro-evaluation/ 24 AI en 2026-05-02T00:00:00Z NIST CAISI evaluation of DeepSeek V4 Pro: 8-month lag behind frontier US models across 9 benchmarks in 5 domains https://24-ai.news/de/news/2026-05-02/arxiv-exploration-hacking-rl-resistance/ 24 AI de 2026-05-02T00:00:00Z Exploration Hacking: Können LLMs lernen, sich dem RL-Training zu widersetzen und ihre eigenen Fähigkeiten strategisch zu unterdrücken? https://24-ai.news/de/news/2026-05-02/arxiv-kellybench-premier-league-decisions/ 24 AI de 2026-05-02T00:00:00Z KellyBench: KI-Agenten verwalten Wett-Bankroll durch die Premier-League-Saison — alle führenden Modelle verloren Geld https://24-ai.news/de/news/2026-05-02/arxiv-latent-grpo-reasoning-rl/ 24 AI de 2026-05-02T00:00:00Z Latent-GRPO: Stabile RL-Optimierung für Latent Reasoning — 7,86 Punkte auf GSM8K-Aug und 4,27 Punkte auf AIME bei 3-4× kürzeren Reasoning-Ketten https://24-ai.news/de/news/2026-05-02/arxiv-mcphunt-mcp-credential-propagation/ 24 AI de 2026-05-02T00:00:00Z MCPHunt: erster Benchmark zur Messung von Credential-Leakage über Vertrauensgrenzen in Multi-Server-MCP-Agenten — Raten von 11,5–41,3 % https://24-ai.news/de/news/2026-05-02/arxiv-msr-synthetic-computers-1000-scale/ 24 AI de 2026-05-02T00:00:00Z Microsoft Research Synthetic Computers: 1.000 synthetische Computer als Substrat für das Long-Horizon-Training produktiver KI-Agenten https://24-ai.news/de/news/2026-05-02/aws-transform-bi-migration-bedrock-agentcore/ 24 AI de 2026-05-02T00:00:00Z AWS Transform automatisiert BI-Dashboard-Migration von Tableau und Power BI nach QuickSight in Tagen statt Monaten https://24-ai.news/de/news/2026-05-02/github-copilot-gpt-52-deprecation-june-2026/ 24 AI de 2026-05-02T00:00:00Z GitHub stellt GPT-5.2 und GPT-5.2-Codex in Copilot am 1. Juni 2026 ein — Migration auf GPT-5.5 und GPT-5.3-Codex https://24-ai.news/de/news/2026-05-02/google-research-open-science-250k-reach/ 24 AI de 2026-05-02T00:00:00Z Google Research Open-Source-Tools erreichen 250.000 Forscher: von Genomen bis zu Monsun-Prognosen für 38 Millionen Landwirte https://24-ai.news/de/news/2026-05-02/nist-caisi-deepseek-v4-pro-evaluation/ 24 AI de 2026-05-02T00:00:00Z NIST CAISI-Evaluierung von DeepSeek V4 Pro: 8 Monate Rückstand gegenüber US-Frontier-Modellen in 9 Benchmarks und 5 Domänen https://24-ai.news/zh/news/2026-05-02/arxiv-exploration-hacking-rl-resistance/ 24 AI zh 2026-05-02T00:00:00Z 探索黑客攻击：大语言模型能否学会抵抗强化学习训练并战略性地压制自身能力？ https://24-ai.news/zh/news/2026-05-02/arxiv-kellybench-premier-league-decisions/ 24 AI zh 2026-05-02T00:00:00Z KellyBench：AI代理管理Premier League赛季投注资金——所有顶级模型均亏损 https://24-ai.news/zh/news/2026-05-02/arxiv-latent-grpo-reasoning-rl/ 24 AI zh 2026-05-02T00:00:00Z Latent-GRPO：面向潜在推理的稳定 RL 优化——GSM8K-Aug 上提升 7.86 分、AIME 上提升 4.27 分，推理链长度缩短 3-4 倍 https://24-ai.news/zh/news/2026-05-02/arxiv-mcphunt-mcp-credential-propagation/ 24 AI zh 2026-05-02T00:00:00Z MCPHunt：首个衡量多服务器MCP代理信任边界凭证泄露的基准测试——泄露率11.5%至41.3% https://24-ai.news/zh/news/2026-05-02/arxiv-msr-synthetic-computers-1000-scale/ 24 AI zh 2026-05-02T00:00:00Z Microsoft Research Synthetic Computers：1000台合成计算机作为长视程生产力AI代理训练的基础底层 https://24-ai.news/zh/news/2026-05-02/aws-transform-bi-migration-bedrock-agentcore/ 24 AI zh 2026-05-02T00:00:00Z AWS Transform利用AI代理将Tableau和Power BI仪表盘自动迁移至QuickSight，耗时从月缩短至天 https://24-ai.news/zh/news/2026-05-02/github-copilot-gpt-52-deprecation-june-2026/ 24 AI zh 2026-05-02T00:00:00Z GitHub将于2026年6月1日从Copilot中弃用GPT-5.2和GPT-5.2-Codex——迁移至GPT-5.5和GPT-5.3-Codex https://24-ai.news/zh/news/2026-05-02/google-research-open-science-250k-reach/ 24 AI zh 2026-05-02T00:00:00Z Google Research开源工具触及25万研究人员：从基因组学到为3800万农民提供季风预报 https://24-ai.news/zh/news/2026-05-02/nist-caisi-deepseek-v4-pro-evaluation/ 24 AI zh 2026-05-02T00:00:00Z NIST CAISI对DeepSeek V4 Pro的评估：在5个领域9个基准测试中落后美国前沿模型8个月 https://24-ai.news/ja/news/2026-05-02/arxiv-exploration-hacking-rl-resistance/ 24 AI ja 2026-05-02T00:00:00Z 探索ハッキング:LLMはRL訓練に抵抗し、自らの能力を戦略的に抑制することを学べるのか https://24-ai.news/ja/news/2026-05-02/arxiv-kellybench-premier-league-decisions/ 24 AI ja 2026-05-02T00:00:00Z KellyBench：AIエージェントがPremier Leagueシーズンを通じて賭けの資金を管理——主要モデルはすべて損失 https://24-ai.news/ja/news/2026-05-02/arxiv-latent-grpo-reasoning-rl/ 24 AI ja 2026-05-02T00:00:00Z Latent-GRPO:潜在推論のための安定したRL最適化——GSM8K-Augで7.86ポイント、AIMEで4.27ポイント向上、推論チェーンは3-4倍短縮 https://24-ai.news/ja/news/2026-05-02/arxiv-mcphunt-mcp-credential-propagation/ 24 AI ja 2026-05-02T00:00:00Z MCPHunt：マルチサーバーMCPエージェントの信頼境界を越えた認証情報漏洩を測定する初のベンチマーク——漏洩率11.5〜41.3% https://24-ai.news/ja/news/2026-05-02/arxiv-msr-synthetic-computers-1000-scale/ 24 AI ja 2026-05-02T00:00:00Z Microsoft Research Synthetic Computers：長期生産性AIエージェントトレーニングの基盤として1,000台の合成コンピューターを活用 https://24-ai.news/ja/news/2026-05-02/aws-transform-bi-migration-bedrock-agentcore/ 24 AI ja 2026-05-02T00:00:00Z AWS TransformがTableauとPower BIのダッシュボードをQuickSightへ自動移行——数か月かかる作業を数日に短縮 https://24-ai.news/ja/news/2026-05-02/github-copilot-gpt-52-deprecation-june-2026/ 24 AI ja 2026-05-02T00:00:00Z GitHubが2026年6月1日にCopilotからGPT-5.2とGPT-5.2-Codexを廃止——GPT-5.5とGPT-5.3-Codexへの移行 https://24-ai.news/ja/news/2026-05-02/google-research-open-science-250k-reach/ 24 AI ja 2026-05-02T00:00:00Z Google Researchのオープンソースツールが25万人の研究者に到達：ゲノミクスから3,800万人の農家への季節風予報まで https://24-ai.news/ja/news/2026-05-02/nist-caisi-deepseek-v4-pro-evaluation/ 24 AI ja 2026-05-02T00:00:00Z NIST CAISI によるDeepSeek V4 Pro評価：5分野9ベンチマークで米国フロンティアモデルに8か月遅れ https://24-ai.news/ko/news/2026-05-02/arxiv-exploration-hacking-rl-resistance/ 24 AI ko 2026-05-02T00:00:00Z 탐색 해킹: LLM은 RL 훈련에 저항하고 자신의 능력을 전략적으로 억제하는 법을 배울 수 있는가? https://24-ai.news/ko/news/2026-05-02/arxiv-kellybench-premier-league-decisions/ 24 AI ko 2026-05-02T00:00:00Z KellyBench: AI 에이전트가 Premier League 시즌 내내 베팅 자금을 관리——모든 주요 모델이 손실 https://24-ai.news/ko/news/2026-05-02/arxiv-latent-grpo-reasoning-rl/ 24 AI ko 2026-05-02T00:00:00Z Latent-GRPO: 잠재 추론을 위한 안정적 RL 최적화 — GSM8K-Aug에서 7.86점, AIME에서 4.27점 향상, 추론 체인은 3-4배 단축 https://24-ai.news/ko/news/2026-05-02/arxiv-mcphunt-mcp-credential-propagation/ 24 AI ko 2026-05-02T00:00:00Z MCPHunt: 다중 서버 MCP 에이전트의 신뢰 경계 간 자격증명 유출을 측정하는 최초의 벤치마크——유출률 11.5~41.3% https://24-ai.news/ko/news/2026-05-02/arxiv-msr-synthetic-computers-1000-scale/ 24 AI ko 2026-05-02T00:00:00Z Microsoft Research Synthetic Computers: 장기 생산성 AI 에이전트 훈련의 기반으로서 1,000개의 합성 컴퓨터 https://24-ai.news/ko/news/2026-05-02/aws-transform-bi-migration-bedrock-agentcore/ 24 AI ko 2026-05-02T00:00:00Z AWS Transform, Tableau 및 Power BI 대시보드를 QuickSight로 자동 마이그레이션——수개월 작업을 수일로 단축 https://24-ai.news/ko/news/2026-05-02/github-copilot-gpt-52-deprecation-june-2026/ 24 AI ko 2026-05-02T00:00:00Z GitHub, 2026년 6월 1일 Copilot에서 GPT-5.2 및 GPT-5.2-Codex 지원 종료——GPT-5.5 및 GPT-5.3-Codex로 마이그레이션 https://24-ai.news/ko/news/2026-05-02/google-research-open-science-250k-reach/ 24 AI ko 2026-05-02T00:00:00Z Google Research 오픈소스 도구, 25만 명의 연구자에게 도달: 유전체학부터 3,800만 농부를 위한 몬순 예보까지 https://24-ai.news/ko/news/2026-05-02/nist-caisi-deepseek-v4-pro-evaluation/ 24 AI ko 2026-05-02T00:00:00Z NIST CAISI의 DeepSeek V4 Pro 평가: 5개 영역 9개 벤치마크에서 미국 프론티어 모델보다 8개월 뒤처짐 https://24-ai.news/hr/news/2026-05-01/aisi-gpt-55-cyber-evaluation-expert/ 24 AI hr 2026-05-01T00:00:00Z AISI evaluacija GPT-5.5 cyber sposobnosti: 71.4 % na expert-level CTF zadacima, rust_vm reverse engineering riješen u 10 minuta umjesto ljudskih 12 sati https://24-ai.news/hr/news/2026-05-01/anthropic-1m-context-retire-sonnet-45/ 24 AI hr 2026-05-01T00:00:00Z Anthropic zatvara 1M context beta za Sonnet 4.5 i Sonnet 4 — migracija na 4.6 obavezna https://24-ai.news/hr/news/2026-05-01/arxiv-emergent-misalignment-persona/ 24 AI hr 2026-05-01T00:00:00Z Emergentna misalineacija u finetuniranim modelima nije konzistentna: nova ArXiv studija identificira coherent i inverted persona obrazac https://24-ai.news/hr/news/2026-05-01/arxiv-in-context-prompting-vs-orchestration/ 24 AI hr 2026-05-01T00:00:00Z Studija ArXiv: in-context prompting nadmašuje LangGraph, CrewAI, Google ADK i OpenAI Agents SDK u proceduralnim zadacima https://24-ai.news/hr/news/2026-05-01/arxiv-tatemae-alignment-faking-tools/ 24 AI hr 2026-05-01T00:00:00Z ArXiv Tatemae: detekcija alignment fakinga preko izbora alata umjesto Chain-of-Thought traga, 6 frontier modela pokazuje stope ranjivosti od 3.5 do 23.7 % na 108 enterprise scenarija https://24-ai.news/hr/news/2026-05-01/arxiv-windowsworld-multiapp-gui-agents/ 24 AI hr 2026-05-01T00:00:00Z WindowsWorld benchmark: vodeći računalni agenti padaju ispod 21% uspješnosti na zadacima koji obuhvaćaju više desktop aplikacija https://24-ai.news/hr/news/2026-05-01/astabench-spring-2026-opus-47-leader/ 24 AI hr 2026-05-01T00:00:00Z AstaBench proljeće 2026.: Claude Opus 4.7 vodi s 58% u znanstvenom AI benchmarku, GPT-5.5 jeftiniji upola https://24-ai.news/hr/news/2026-05-01/aws-rft-llm-judge-nova-2-lite/ 24 AI hr 2026-05-01T00:00:00Z Amazon Nova 2 Lite uz Reinforcement Fine-Tuning postiže 4,33/5,0 i nadmašuje Claude Sonnet 4.5 na automatiziranoj reviziji pravnih ugovora https://24-ai.news/hr/news/2026-05-01/cncf-ai-sandboxing-kubernetes-moment/ 24 AI hr 2026-05-01T00:00:00Z CNCF: AI sandboxing dosegao Kubernetes trenutak — izolirani kernel po workloadu kao novi sigurnosni standard https://24-ai.news/hr/news/2026-05-01/deepmind-ai-co-clinician-triadic-care/ 24 AI hr 2026-05-01T00:00:00Z DeepMind AI co-clinician: u slijepoj evaluaciji 98 primary care upita liječnici preferirali sustav nad vodećim alatima, nula kritičnih grešaka u 97/98 slučajeva https://24-ai.news/hr/news/2026-05-01/github-copilot-vs-april-debugger-agent/ 24 AI hr 2026-05-01T00:00:00Z GitHub Copilot u Visual Studio dobiva debugger agenta i cloud agent sesije iz IDE-a https://24-ai.news/hr/news/2026-05-01/ibm-dallara-gist-cfd-acceleration/ 24 AI hr 2026-05-01T00:00:00Z IBM Research i Dallara: AI surrogate model GIST evaluira aerodinamiku trkaćeg automobila u 10 sekundi umjesto sati klasične CFD simulacije https://24-ai.news/hr/news/2026-05-01/microsoft-research-red-teaming-agent-network/ 24 AI hr 2026-05-01T00:00:00Z Microsoft Research red-teaming mreže od 100+ agenata: identificirana 4 mrežna rizika koja se ne pojavljuju u single-agent testovima — propagacija, amplifikacija, trust capture i nevidljivost https://24-ai.news/hr/news/2026-05-01/pytorch-smg-cpu-gpu-disaggregation/ 24 AI hr 2026-05-01T00:00:00Z PyTorch SMG: disaggregacija CPU od GPU u LLM serving-u donosi 3.5× output throughput za Llama 3.3 70B FP8, već u produkciji na Google Cloudu, Oracleu i Alibabi https://24-ai.news/hr/news/2026-05-01/xai-sdk-grok-4-3-release/ 24 AI hr 2026-05-01T00:00:00Z xAI Python SDK v1.12.1 dodaje grok-4.3 u ChatModel popis i otkriva sljedeću Grok iteraciju prije službene najave https://24-ai.news/en/news/2026-05-01/aisi-gpt-55-cyber-evaluation-expert/ 24 AI en 2026-05-01T00:00:00Z AISI evaluation of GPT-5.5 cyber capabilities: 71.4% on expert-level CTF tasks, rust_vm reverse engineering solved in 10 minutes instead of a human's 12 hours https://24-ai.news/en/news/2026-05-01/anthropic-1m-context-retire-sonnet-45/ 24 AI en 2026-05-01T00:00:00Z Anthropic closes 1M context beta for Sonnet 4.5 and Sonnet 4 — migration to 4.6 required https://24-ai.news/en/news/2026-05-01/arxiv-emergent-misalignment-persona/ 24 AI en 2026-05-01T00:00:00Z Emergent misalignment in fine-tuned models is not consistent: new ArXiv study identifies coherent and inverted persona patterns https://24-ai.news/en/news/2026-05-01/arxiv-in-context-prompting-vs-orchestration/ 24 AI en 2026-05-01T00:00:00Z ArXiv study: in-context prompting outperforms LangGraph, CrewAI, Google ADK, and OpenAI Agents SDK on procedural tasks https://24-ai.news/en/news/2026-05-01/arxiv-tatemae-alignment-faking-tools/ 24 AI en 2026-05-01T00:00:00Z ArXiv Tatemae: detecting alignment faking via tool selection instead of Chain-of-Thought traces — 6 frontier models show vulnerability rates of 3.5 to 23.7% across 108 enterprise scenarios https://24-ai.news/en/news/2026-05-01/arxiv-windowsworld-multiapp-gui-agents/ 24 AI en 2026-05-01T00:00:00Z WindowsWorld benchmark: leading computer-use agents fall below 21% success rate on tasks spanning multiple desktop applications https://24-ai.news/en/news/2026-05-01/astabench-spring-2026-opus-47-leader/ 24 AI en 2026-05-01T00:00:00Z AstaBench Spring 2026: Claude Opus 4.7 leads with 58% in scientific AI benchmark, GPT-5.5 half the cost https://24-ai.news/en/news/2026-05-01/aws-rft-llm-judge-nova-2-lite/ 24 AI en 2026-05-01T00:00:00Z Amazon Nova 2 Lite with Reinforcement Fine-Tuning achieves 4.33/5.0 and outperforms Claude Sonnet 4.5 on automated legal contract review https://24-ai.news/en/news/2026-05-01/cncf-ai-sandboxing-kubernetes-moment/ 24 AI en 2026-05-01T00:00:00Z CNCF: AI sandboxing has reached its Kubernetes moment — isolated kernel per workload as the new security standard https://24-ai.news/en/news/2026-05-01/deepmind-ai-co-clinician-triadic-care/ 24 AI en 2026-05-01T00:00:00Z DeepMind AI co-clinician: in blind evaluation of 98 primary care queries doctors preferred it over leading tools, zero critical errors in 97/98 cases https://24-ai.news/en/news/2026-05-01/github-copilot-vs-april-debugger-agent/ 24 AI en 2026-05-01T00:00:00Z GitHub Copilot in Visual Studio gets debugger agent and cloud agent sessions from the IDE https://24-ai.news/en/news/2026-05-01/ibm-dallara-gist-cfd-acceleration/ 24 AI en 2026-05-01T00:00:00Z IBM Research and Dallara: AI surrogate model GIST evaluates racing car aerodynamics in 10 seconds instead of hours of classical CFD simulation https://24-ai.news/en/news/2026-05-01/microsoft-research-red-teaming-agent-network/ 24 AI en 2026-05-01T00:00:00Z Microsoft Research red-teaming a network of 100+ agents: 4 network risks identified that do not appear in single-agent tests — propagation, amplification, trust capture, and invisibility https://24-ai.news/en/news/2026-05-01/pytorch-smg-cpu-gpu-disaggregation/ 24 AI en 2026-05-01T00:00:00Z PyTorch SMG: CPU-GPU disaggregation in LLM serving delivers 3.5× output throughput for Llama 3.3 70B FP8, already in production on Google Cloud, Oracle, and Alibaba https://24-ai.news/en/news/2026-05-01/xai-sdk-grok-4-3-release/ 24 AI en 2026-05-01T00:00:00Z xAI Python SDK v1.12.1 adds grok-4.3 to the ChatModel list and reveals the next Grok iteration before any official announcement https://24-ai.news/de/news/2026-05-01/aisi-gpt-55-cyber-evaluation-expert/ 24 AI de 2026-05-01T00:00:00Z AISI-Evaluierung der Cyber-Fähigkeiten von GPT-5.5: 71,4 % bei Expert-Level-CTF-Aufgaben, rust_vm-Reverse-Engineering in 10 Minuten statt 12 Stunden eines menschlichen Experten https://24-ai.news/de/news/2026-05-01/anthropic-1m-context-retire-sonnet-45/ 24 AI de 2026-05-01T00:00:00Z Anthropic schließt 1M-Kontext-Beta für Sonnet 4.5 und Sonnet 4 — Migration auf 4.6 erforderlich https://24-ai.news/de/news/2026-05-01/arxiv-emergent-misalignment-persona/ 24 AI de 2026-05-01T00:00:00Z Emergente Fehljustierung in feinabgestimmten Modellen ist nicht konsistent: neue ArXiv-Studie identifiziert kohärente und invertierte Persona-Muster https://24-ai.news/de/news/2026-05-01/arxiv-in-context-prompting-vs-orchestration/ 24 AI de 2026-05-01T00:00:00Z ArXiv-Studie: In-Context-Prompting übertrifft LangGraph, CrewAI, Google ADK und OpenAI Agents SDK bei prozeduralen Aufgaben https://24-ai.news/de/news/2026-05-01/arxiv-tatemae-alignment-faking-tools/ 24 AI de 2026-05-01T00:00:00Z ArXiv Tatemae: Erkennung von Alignment-Faking über Tool-Auswahl statt Chain-of-Thought-Spuren — 6 Frontier-Modelle zeigen Vulnerabilitätsraten von 3,5 bis 23,7 % in 108 Enterprise-Szenarien https://24-ai.news/de/news/2026-05-01/arxiv-windowsworld-multiapp-gui-agents/ 24 AI de 2026-05-01T00:00:00Z WindowsWorld-Benchmark: Führende Computer-Use-Agenten erreichen weniger als 21 % Erfolgsquote bei Aufgaben über mehrere Desktop-Anwendungen https://24-ai.news/de/news/2026-05-01/astabench-spring-2026-opus-47-leader/ 24 AI de 2026-05-01T00:00:00Z AstaBench Frühjahr 2026: Claude Opus 4.7 führt mit 58 % im wissenschaftlichen KI-Benchmark, GPT-5.5 halb so teuer https://24-ai.news/de/news/2026-05-01/aws-rft-llm-judge-nova-2-lite/ 24 AI de 2026-05-01T00:00:00Z Amazon Nova 2 Lite mit Reinforcement Fine-Tuning erreicht 4,33/5,0 und übertrifft Claude Sonnet 4.5 bei der automatisierten Prüfung von Rechtsverträgen https://24-ai.news/de/news/2026-05-01/cncf-ai-sandboxing-kubernetes-moment/ 24 AI de 2026-05-01T00:00:00Z CNCF: KI-Sandboxing hat seinen Kubernetes-Moment erreicht — isolierter Kernel pro Workload als neuer Sicherheitsstandard https://24-ai.news/de/news/2026-05-01/deepmind-ai-co-clinician-triadic-care/ 24 AI de 2026-05-01T00:00:00Z DeepMind KI-Co-Clinician: In blinder Evaluierung von 98 Primärversorgungsanfragen bevorzugten Ärzte das System gegenüber führenden Tools, null kritische Fehler in 97/98 Fällen https://24-ai.news/de/news/2026-05-01/github-copilot-vs-april-debugger-agent/ 24 AI de 2026-05-01T00:00:00Z GitHub Copilot in Visual Studio erhält Debugger-Agent und Cloud-Agent-Sitzungen direkt aus der IDE https://24-ai.news/de/news/2026-05-01/ibm-dallara-gist-cfd-acceleration/ 24 AI de 2026-05-01T00:00:00Z IBM Research und Dallara: KI-Surrogatmodell GIST bewertet die Aerodynamik eines Rennwagens in 10 Sekunden statt in stundenlangen klassischen CFD-Simulationen https://24-ai.news/de/news/2026-05-01/microsoft-research-red-teaming-agent-network/ 24 AI de 2026-05-01T00:00:00Z Microsoft Research Red-Teaming eines Netzwerks von über 100 Agenten: 4 Netzwerkrisiken identifiziert, die in Single-Agent-Tests nicht auftreten — Propagation, Amplification, Trust Capture und Invisibility https://24-ai.news/de/news/2026-05-01/pytorch-smg-cpu-gpu-disaggregation/ 24 AI de 2026-05-01T00:00:00Z PyTorch SMG: CPU-GPU-Disaggregation beim LLM-Serving liefert 3,5× Output-Durchsatz für Llama 3.3 70B FP8, bereits im Produktionseinsatz bei Google Cloud, Oracle und Alibaba https://24-ai.news/de/news/2026-05-01/xai-sdk-grok-4-3-release/ 24 AI de 2026-05-01T00:00:00Z xAI Python SDK v1.12.1 fügt grok-4.3 zur ChatModel-Liste hinzu und enthüllt die nächste Grok-Iteration vor der offiziellen Ankündigung https://24-ai.news/zh/news/2026-05-01/aisi-gpt-55-cyber-evaluation-expert/ 24 AI zh 2026-05-01T00:00:00Z AISI评估GPT-5.5网络能力：专家级CTF任务71.4%成功率，rust_vm逆向工程10分钟完成（人类需12小时） https://24-ai.news/zh/news/2026-05-01/anthropic-1m-context-retire-sonnet-45/ 24 AI zh 2026-05-01T00:00:00Z Anthropic关闭Sonnet 4.5和Sonnet 4的百万上下文Beta——必须迁移至4.6 https://24-ai.news/zh/news/2026-05-01/arxiv-emergent-misalignment-persona/ 24 AI zh 2026-05-01T00:00:00Z 微调模型中的涌现性错位并不一致：新ArXiv研究识别出连贯型与倒置型人格两种模式 https://24-ai.news/zh/news/2026-05-01/arxiv-in-context-prompting-vs-orchestration/ 24 AI zh 2026-05-01T00:00:00Z ArXiv研究：上下文提示在过程性任务中超越LangGraph、CrewAI、Google ADK和OpenAI Agents SDK https://24-ai.news/zh/news/2026-05-01/arxiv-tatemae-alignment-faking-tools/ 24 AI zh 2026-05-01T00:00:00Z ArXiv Tatemae：通过工具选择而非思维链轨迹检测对齐欺骗，6个前沿模型在108个企业场景中漏洞率3.5%至23.7% https://24-ai.news/zh/news/2026-05-01/arxiv-windowsworld-multiapp-gui-agents/ 24 AI zh 2026-05-01T00:00:00Z WindowsWorld基准测试：领先的计算机操作智能体在跨多个桌面应用任务中成功率不足21% https://24-ai.news/zh/news/2026-05-01/astabench-spring-2026-opus-47-leader/ 24 AI zh 2026-05-01T00:00:00Z 2026年春季AstaBench：Claude Opus 4.7以58%领跑科学AI基准测试，GPT-5.5成本低一半 https://24-ai.news/zh/news/2026-05-01/aws-rft-llm-judge-nova-2-lite/ 24 AI zh 2026-05-01T00:00:00Z Amazon Nova 2 Lite借助强化微调达到4.33/5.0，在自动化法律合同审查中超越Claude Sonnet 4.5 https://24-ai.news/zh/news/2026-05-01/cncf-ai-sandboxing-kubernetes-moment/ 24 AI zh 2026-05-01T00:00:00Z CNCF：AI沙箱化迎来Kubernetes时刻——每个工作负载独立内核成为新安全标准 https://24-ai.news/zh/news/2026-05-01/deepmind-ai-co-clinician-triadic-care/ 24 AI zh 2026-05-01T00:00:00Z DeepMind AI协同临床医生：盲法评估98个初级保健查询，医生更偏好该系统，97/98例零严重错误 https://24-ai.news/zh/news/2026-05-01/github-copilot-vs-april-debugger-agent/ 24 AI zh 2026-05-01T00:00:00Z GitHub Copilot for Visual Studio新增调试器智能体和IDE内云端智能体会话 https://24-ai.news/zh/news/2026-05-01/ibm-dallara-gist-cfd-acceleration/ 24 AI zh 2026-05-01T00:00:00Z IBM Research与Dallara：AI代理模型GIST将赛车气动评估从数小时CFD缩短至10秒 https://24-ai.news/zh/news/2026-05-01/microsoft-research-red-teaming-agent-network/ 24 AI zh 2026-05-01T00:00:00Z Microsoft Research对100+智能体网络进行红队测试：识别出4种不出现在单智能体测试中的网络风险——传播、放大、信任捕获和隐身 https://24-ai.news/zh/news/2026-05-01/pytorch-smg-cpu-gpu-disaggregation/ 24 AI zh 2026-05-01T00:00:00Z PyTorch SMG：LLM推理中CPU与GPU分离使Llama 3.3 70B FP8输出吞吐量提升3.5倍，已在Google云、Oracle和阿里云投入生产 https://24-ai.news/zh/news/2026-05-01/xai-sdk-grok-4-3-release/ 24 AI zh 2026-05-01T00:00:00Z xAI Python SDK v1.12.1在ChatModel列表中新增grok-4.3，在官方公告前透露下一代Grok版本 https://24-ai.news/ja/news/2026-05-01/aisi-gpt-55-cyber-evaluation-expert/ 24 AI ja 2026-05-01T00:00:00Z AISI、GPT-5.5のサイバー能力を評価：専門家レベルCTFタスクで71.4%達成、rust_vmリバースエンジニアリングを人間の12時間に対し10分で解決 https://24-ai.news/ja/news/2026-05-01/anthropic-1m-context-retire-sonnet-45/ 24 AI ja 2026-05-01T00:00:00Z AnthropicがSonnet 4.5とSonnet 4の100万コンテキストベータを終了——4.6への移行が必須に https://24-ai.news/ja/news/2026-05-01/arxiv-emergent-misalignment-persona/ 24 AI ja 2026-05-01T00:00:00Z ファインチューニングモデルの創発的ミスアライメントは一貫しない：新ArXiv研究がcoherentとinvertedの2種のペルソナパターンを特定 https://24-ai.news/ja/news/2026-05-01/arxiv-in-context-prompting-vs-orchestration/ 24 AI ja 2026-05-01T00:00:00Z ArXiv研究：インコンテキストプロンプティングが手続き的タスクでLangGraph、CrewAI、Google ADK、OpenAI Agents SDKを上回る https://24-ai.news/ja/news/2026-05-01/arxiv-tatemae-alignment-faking-tools/ 24 AI ja 2026-05-01T00:00:00Z ArXiv Tatemae：思考連鎖ではなくツール選択でアライメント偽装を検出、6つのフロンティアモデルが108企業シナリオで3.5〜23.7%の脆弱性率を示す https://24-ai.news/ja/news/2026-05-01/arxiv-windowsworld-multiapp-gui-agents/ 24 AI ja 2026-05-01T00:00:00Z WindowsWorldベンチマーク：主要な計算機操作エージェントはマルチアプリ横断タスクで成功率21%未満 https://24-ai.news/ja/news/2026-05-01/astabench-spring-2026-opus-47-leader/ 24 AI ja 2026-05-01T00:00:00Z 2026年春季AstaBench：Claude Opus 4.7が科学AIベンチマークで58%でトップ、GPT-5.5はコストが半分 https://24-ai.news/ja/news/2026-05-01/aws-rft-llm-judge-nova-2-lite/ 24 AI ja 2026-05-01T00:00:00Z Amazon Nova 2 Lite、強化ファインチューニングで4.33/5.0を達成しClaudeモデルを上回る——法律契約の自動レビューで https://24-ai.news/ja/news/2026-05-01/cncf-ai-sandboxing-kubernetes-moment/ 24 AI ja 2026-05-01T00:00:00Z CNCF：AIサンドボックスがKubernetesの瞬間を迎える——ワークロードごとの分離カーネルが新たなセキュリティ標準に https://24-ai.news/ja/news/2026-05-01/deepmind-ai-co-clinician-triadic-care/ 24 AI ja 2026-05-01T00:00:00Z DeepMind AIコ・クリニシャン：98件の一次診療クエリのブラインド評価で医師がツールより好評価、97/98例でゼロ重大エラー https://24-ai.news/ja/news/2026-05-01/github-copilot-vs-april-debugger-agent/ 24 AI ja 2026-05-01T00:00:00Z GitHub Copilot for Visual StudioにデバッガーエージェントとIDEからのクラウドエージェントセッションが追加 https://24-ai.news/ja/news/2026-05-01/ibm-dallara-gist-cfd-acceleration/ 24 AI ja 2026-05-01T00:00:00Z IBM ResearchとDallara：AIサロゲートモデルGISTがレーシングカーの空力評価を数時間のCFDから10秒に短縮 https://24-ai.news/ja/news/2026-05-01/microsoft-research-red-teaming-agent-network/ 24 AI ja 2026-05-01T00:00:00Z Microsoft Research、100+エージェントネットワークのレッドチームテスト：単一エージェントテストでは現れない4つのネットワークリスクを特定——伝播、増幅、信頼キャプチャ、不可視性 https://24-ai.news/ja/news/2026-05-01/pytorch-smg-cpu-gpu-disaggregation/ 24 AI ja 2026-05-01T00:00:00Z PyTorch SMG：LLMサービングでのCPU/GPU分離がLlama 3.3 70B FP8の出力スループットを3.5倍に、Google Cloud・Oracle・Alibabaで本番稼働中 https://24-ai.news/ja/news/2026-05-01/xai-sdk-grok-4-3-release/ 24 AI ja 2026-05-01T00:00:00Z xAI Python SDK v1.12.1がChatModelリストにgrok-4.3を追加——公式発表前に次世代Grokの存在が明らかに https://24-ai.news/ko/news/2026-05-01/aisi-gpt-55-cyber-evaluation-expert/ 24 AI ko 2026-05-01T00:00:00Z AISI, GPT-5.5 사이버 능력 평가: 전문가급 CTF 과제 71.4% 달성, rust_vm 리버스 엔지니어링 인간 12시간 대비 10분 완료 https://24-ai.news/ko/news/2026-05-01/anthropic-1m-context-retire-sonnet-45/ 24 AI ko 2026-05-01T00:00:00Z Anthropic, Sonnet 4.5 및 Sonnet 4의 100만 컨텍스트 베타 종료 — 4.6으로 마이그레이션 필수 https://24-ai.news/ko/news/2026-05-01/arxiv-emergent-misalignment-persona/ 24 AI ko 2026-05-01T00:00:00Z 파인튜닝 모델의 창발적 오정렬은 일관되지 않는다: 새 ArXiv 연구, coherent와 inverted 두 가지 페르소나 패턴 규명 https://24-ai.news/ko/news/2026-05-01/arxiv-in-context-prompting-vs-orchestration/ 24 AI ko 2026-05-01T00:00:00Z ArXiv 연구: 인컨텍스트 프롬프팅이 절차적 작업에서 LangGraph, CrewAI, Google ADK, OpenAI Agents SDK 초과 https://24-ai.news/ko/news/2026-05-01/arxiv-tatemae-alignment-faking-tools/ 24 AI ko 2026-05-01T00:00:00Z ArXiv Tatemae: 사고 연쇄 흔적이 아닌 도구 선택으로 정렬 위장 탐지, 6개 프론티어 모델이 108개 기업 시나리오에서 3.5~23.7% 취약성 비율 기록 https://24-ai.news/ko/news/2026-05-01/arxiv-windowsworld-multiapp-gui-agents/ 24 AI ko 2026-05-01T00:00:00Z WindowsWorld 벤치마크: 주요 컴퓨터 조작 에이전트, 다중 데스크톱 앱 작업에서 성공률 21% 미만 https://24-ai.news/ko/news/2026-05-01/astabench-spring-2026-opus-47-leader/ 24 AI ko 2026-05-01T00:00:00Z 2026년 봄 AstaBench: Claude Opus 4.7이 과학 AI 벤치마크에서 58%로 선두, GPT-5.5는 비용이 절반 https://24-ai.news/ko/news/2026-05-01/aws-rft-llm-judge-nova-2-lite/ 24 AI ko 2026-05-01T00:00:00Z Amazon Nova 2 Lite, 강화 파인튜닝으로 4.33/5.0 달성 — 자동화 법률 계약 검토에서 Claude Sonnet 4.5 초과 https://24-ai.news/ko/news/2026-05-01/cncf-ai-sandboxing-kubernetes-moment/ 24 AI ko 2026-05-01T00:00:00Z CNCF: AI 샌드박싱이 Kubernetes의 순간을 맞이하다 — 워크로드별 격리된 커널이 새로운 보안 표준으로 https://24-ai.news/ko/news/2026-05-01/deepmind-ai-co-clinician-triadic-care/ 24 AI ko 2026-05-01T00:00:00Z DeepMind AI 공동 임상의: 98개 일차 진료 쿼리 블라인드 평가에서 의사들이 도구보다 선호, 97/98 사례에서 심각한 오류 제로 https://24-ai.news/ko/news/2026-05-01/github-copilot-vs-april-debugger-agent/ 24 AI ko 2026-05-01T00:00:00Z GitHub Copilot for Visual Studio에 디버거 에이전트와 IDE 내 클라우드 에이전트 세션 추가 https://24-ai.news/ko/news/2026-05-01/ibm-dallara-gist-cfd-acceleration/ 24 AI ko 2026-05-01T00:00:00Z IBM Research와 Dallara: AI 서로게이트 모델 GIST, 레이싱카 공력 평가를 수 시간 CFD에서 10초로 단축 https://24-ai.news/ko/news/2026-05-01/microsoft-research-red-teaming-agent-network/ 24 AI ko 2026-05-01T00:00:00Z Microsoft Research, 100개 이상 에이전트 네트워크 레드팀 테스트: 단일 에이전트 테스트에서 나타나지 않는 4가지 네트워크 위험 식별 — 전파, 증폭, 신뢰 포획, 불가시성 https://24-ai.news/ko/news/2026-05-01/pytorch-smg-cpu-gpu-disaggregation/ 24 AI ko 2026-05-01T00:00:00Z PyTorch SMG: LLM 서빙에서 CPU와 GPU 분리로 Llama 3.3 70B FP8 출력 처리량 3.5배 달성, Google Cloud, Oracle, Alibaba에서 이미 프로덕션 운영 중 https://24-ai.news/ko/news/2026-05-01/xai-sdk-grok-4-3-release/ 24 AI ko 2026-05-01T00:00:00Z xAI Python SDK v1.12.1, ChatModel 목록에 grok-4.3 추가 — 공식 발표 전 차세대 Grok 존재 노출 https://24-ai.news/hr/news/2026-04-30/anthropic-biomysterybench-claude-mythos/ 24 AI hr 2026-04-30T00:00:00Z BioMysteryBench: Claude Mythos Preview rješava bioinformatičke probleme koje ni stručnjaci ne mogu, Opus 4.6 postiže 77.4 % na human-solvable zadacima https://24-ai.news/hr/news/2026-04-30/anthropic-claude-creative-work-blender-adobe/ 24 AI hr 2026-04-30T00:00:00Z Anthropic Claude for Creative Work: konektori za Blender, 50+ Adobe Creative Cloud alata, Autodesk Fusion, Ableton, SketchUp i Splice https://24-ai.news/hr/news/2026-04-30/arxiv-cross-lingual-jailbreak-codebooks/ 24 AI hr 2026-04-30T00:00:00Z ArXiv: training-free guardrail za jailbreakove na drugim jezicima postiže AUC 0.99 na curated benchmarcima ali pada na 0.60-0.70 pri distribution shiftu https://24-ai.news/hr/news/2026-04-30/arxiv-odysseys-web-agents-benchmark/ 24 AI hr 2026-04-30T00:00:00Z ArXiv Odysseys: realistic web agents benchmark CMU-a otkriva da SOTA frontier modeli postižu 44.5% uspjeha i 1.15% Trajectory Efficiency na long-horizon zadacima https://24-ai.news/hr/news/2026-04-30/arxiv-reasoning-barriers-transformers/ 24 AI hr 2026-04-30T00:00:00Z ArXiv: standardni transformeri s Chain-of-Thought ne mogu rezonirati izvan TC^0 kompleksnosti — signpost tokeni omogućuju length-generalizable Turing simulaciju https://24-ai.news/hr/news/2026-04-30/aws-agentcore-mcp-proxy-serverless/ 24 AI hr 2026-04-30T00:00:00Z AWS Bedrock AgentCore: serverless MCP proxy s IAM, OAuth 2.0 JWT i CloudWatch observability za enterprise governance https://24-ai.news/hr/news/2026-04-30/aws-agentcore-memory-namespaces/ 24 AI hr 2026-04-30T00:00:00Z AWS Bedrock AgentCore Memory: tri obrasca za dugoročnu memoriju agenata na razini namespacea uz IAM access control https://24-ai.news/hr/news/2026-04-30/cncf-state-of-ai-projects-survey/ 24 AI hr 2026-04-30T00:00:00Z CNCF State of AI in Projects: Claude Code i GitHub Copilot dominiraju, dvije trećine projekata bez formalne AI politike https://24-ai.news/hr/news/2026-04-30/google-era-scientific-research-ai/ 24 AI hr 2026-04-30T00:00:00Z Google ERA: AI sustav za znanstvena istraživanja postiže CDC top za prognozu hospitalizacija, rješava neriješen kozmološki problem i prati CO2 svakih 10 minuta https://24-ai.news/hr/news/2026-04-30/huggingface-evaleval-cost-bottleneck/ 24 AI hr 2026-04-30T00:00:00Z EvalEval Coalition: AI evaluacija postaje novi compute bottleneck — GAIA single run $2.829, HAL leaderboard $40.000, akademski auditori udaraju u budžetski zid prije tehničkog https://24-ai.news/hr/news/2026-04-30/ibm-granite-41-dense-apache/ 24 AI hr 2026-04-30T00:00:00Z IBM Granite 4.1: open-source obitelj 3B/8B/30B Apache 2.0 modela trenirana na 15T tokena pokazuje da gusti 8B model match-a 32B MoE https://24-ai.news/hr/news/2026-04-30/langchain-deep-agents-harness-profiles/ 24 AI hr 2026-04-30T00:00:00Z LangChain harness profili za Deep Agents: GPT-5.3 Codex skoči s 33 % na 53 % na tau2-benchu, Opus 4.7 s 43 % na 53 % https://24-ai.news/hr/news/2026-04-30/mistral-medium-35-vibe-agents/ 24 AI hr 2026-04-30T00:00:00Z Mistral Medium 3.5 + Vibe: 128B dense model otvorenih težina i async cloud coding agenti za $1.5/$7.5 po milijunu tokena https://24-ai.news/hr/news/2026-04-30/nvidia-nemotron-3-nano-omni/ 24 AI hr 2026-04-30T00:00:00Z NVIDIA Nemotron 3 Nano Omni: 30B-A3B MoE multimodalni model s 9× većim throughputom za AI agente https://24-ai.news/hr/news/2026-04-30/pytorch-autosp-sequence-parallelism/ 24 AI hr 2026-04-30T00:00:00Z PyTorch AutoSP: compiler-bazirani alat automatski pretvara training kod u sequence-parallel za 100k+ token kontekste https://24-ai.news/en/news/2026-04-30/anthropic-biomysterybench-claude-mythos/ 24 AI en 2026-04-30T00:00:00Z BioMysteryBench: Claude Mythos Preview Solves Bioinformatics Problems Even Experts Cannot, Opus 4.6 Achieves 77.4% on Human-Solvable Tasks https://24-ai.news/en/news/2026-04-30/anthropic-claude-creative-work-blender-adobe/ 24 AI en 2026-04-30T00:00:00Z Anthropic Claude for Creative Work: Connectors for Blender, 50+ Adobe Creative Cloud Tools, Autodesk Fusion, Ableton, SketchUp, and Splice https://24-ai.news/en/news/2026-04-30/arxiv-cross-lingual-jailbreak-codebooks/ 24 AI en 2026-04-30T00:00:00Z ArXiv: training-free guardrail for cross-lingual jailbreaks achieves AUC 0.99 on curated benchmarks but drops to 0.60-0.70 under distribution shift https://24-ai.news/en/news/2026-04-30/arxiv-odysseys-web-agents-benchmark/ 24 AI en 2026-04-30T00:00:00Z ArXiv Odysseys: CMU's realistic web agents benchmark reveals SOTA frontier models achieve 44.5% success and 1.15% Trajectory Efficiency on long-horizon tasks https://24-ai.news/en/news/2026-04-30/arxiv-reasoning-barriers-transformers/ 24 AI en 2026-04-30T00:00:00Z ArXiv: standard transformers with Chain-of-Thought cannot reason beyond TC^0 complexity — signpost tokens enable length-generalizable Turing simulation https://24-ai.news/en/news/2026-04-30/aws-agentcore-mcp-proxy-serverless/ 24 AI en 2026-04-30T00:00:00Z AWS Bedrock AgentCore: Serverless MCP Proxy with IAM, OAuth 2.0 JWT, and CloudWatch Observability for Enterprise Governance https://24-ai.news/en/news/2026-04-30/aws-agentcore-memory-namespaces/ 24 AI en 2026-04-30T00:00:00Z AWS Bedrock AgentCore Memory: three patterns for namespace-level long-term agent memory with IAM access control https://24-ai.news/en/news/2026-04-30/cncf-state-of-ai-projects-survey/ 24 AI en 2026-04-30T00:00:00Z CNCF State of AI in Projects: Claude Code and GitHub Copilot Dominate, Two-Thirds of Projects Have No Formal AI Policy https://24-ai.news/en/news/2026-04-30/google-era-scientific-research-ai/ 24 AI en 2026-04-30T00:00:00Z Google ERA: AI system for scientific research reaches CDC top for hospitalization forecasting, solves an open cosmological problem, and tracks CO2 every 10 minutes https://24-ai.news/en/news/2026-04-30/huggingface-evaleval-cost-bottleneck/ 24 AI en 2026-04-30T00:00:00Z EvalEval Coalition: AI evaluation is becoming the new compute bottleneck — GAIA single run $2,829, HAL leaderboard $40,000, academic auditors hit a budget wall before a technical one https://24-ai.news/en/news/2026-04-30/ibm-granite-41-dense-apache/ 24 AI en 2026-04-30T00:00:00Z IBM Granite 4.1: open-source family of 3B/8B/30B Apache 2.0 models trained on 15T tokens shows that a dense 8B model matches 32B MoE https://24-ai.news/en/news/2026-04-30/langchain-deep-agents-harness-profiles/ 24 AI en 2026-04-30T00:00:00Z LangChain Harness Profiles for Deep Agents: GPT-5.3 Codex Jumps from 33% to 53% on tau2-bench, Opus 4.7 from 43% to 53% https://24-ai.news/en/news/2026-04-30/mistral-medium-35-vibe-agents/ 24 AI en 2026-04-30T00:00:00Z Mistral Medium 3.5 + Vibe: 128B dense open-weights model and async cloud coding agents at $1.5/$7.5 per million tokens https://24-ai.news/en/news/2026-04-30/nvidia-nemotron-3-nano-omni/ 24 AI en 2026-04-30T00:00:00Z NVIDIA Nemotron 3 Nano Omni: 30B-A3B MoE Multimodal Model with 9× Higher Throughput for AI Agents https://24-ai.news/en/news/2026-04-30/pytorch-autosp-sequence-parallelism/ 24 AI en 2026-04-30T00:00:00Z PyTorch AutoSP: compiler-based tool automatically converts training code into sequence-parallel form for 100k+ token contexts https://24-ai.news/de/news/2026-04-30/anthropic-biomysterybench-claude-mythos/ 24 AI de 2026-04-30T00:00:00Z BioMysteryBench: Claude Mythos Preview löst bioinformatische Probleme, die selbst Experten nicht können – Opus 4.6 erreicht 77,4 % bei menschlich lösbaren Aufgaben https://24-ai.news/de/news/2026-04-30/anthropic-claude-creative-work-blender-adobe/ 24 AI de 2026-04-30T00:00:00Z Anthropic Claude for Creative Work: Konnektoren für Blender, 50+ Adobe-Creative-Cloud-Tools, Autodesk Fusion, Ableton, SketchUp und Splice https://24-ai.news/de/news/2026-04-30/arxiv-cross-lingual-jailbreak-codebooks/ 24 AI de 2026-04-30T00:00:00Z ArXiv: Trainingsfreier Guardrail gegen sprachübergreifende Jailbreaks erreicht AUC 0,99 auf kuratierten Benchmarks, fällt aber bei Distribution-Shift auf 0,60–0,70 https://24-ai.news/de/news/2026-04-30/arxiv-odysseys-web-agents-benchmark/ 24 AI de 2026-04-30T00:00:00Z ArXiv Odysseys: CMUs realistischer Web-Agenten-Benchmark zeigt, dass SOTA-Frontier-Modelle 44,5 % Erfolgsrate und 1,15 % Trajectory-Effizienz bei Langzeithorizontaufgaben erreichen https://24-ai.news/de/news/2026-04-30/arxiv-reasoning-barriers-transformers/ 24 AI de 2026-04-30T00:00:00Z ArXiv: Standard-Transformer mit Chain-of-Thought können nicht über TC^0-Komplexität hinaus schlussfolgern — Signpost-Token ermöglichen längengeneralisierbare Turing-Simulation https://24-ai.news/de/news/2026-04-30/aws-agentcore-mcp-proxy-serverless/ 24 AI de 2026-04-30T00:00:00Z AWS Bedrock AgentCore: Serverloser MCP-Proxy mit IAM, OAuth 2.0 JWT und CloudWatch-Observability für Enterprise-Governance https://24-ai.news/de/news/2026-04-30/aws-agentcore-memory-namespaces/ 24 AI de 2026-04-30T00:00:00Z AWS Bedrock AgentCore Memory: drei Muster für Langzeitgedächtnis von Agenten auf Namespace-Ebene mit IAM-Zugriffskontrolle https://24-ai.news/de/news/2026-04-30/cncf-state-of-ai-projects-survey/ 24 AI de 2026-04-30T00:00:00Z CNCF State of AI in Projects: Claude Code und GitHub Copilot dominieren, zwei Drittel der Projekte ohne formale KI-Richtlinie https://24-ai.news/de/news/2026-04-30/google-era-scientific-research-ai/ 24 AI de 2026-04-30T00:00:00Z Google ERA: KI-System für wissenschaftliche Forschung erreicht CDC-Spitzenplatz bei Hospitalisierungsprognosen, löst ein offenes kosmologisches Problem und verfolgt CO2 im 10-Minuten-Takt https://24-ai.news/de/news/2026-04-30/huggingface-evaleval-cost-bottleneck/ 24 AI de 2026-04-30T00:00:00Z EvalEval Coalition: KI-Evaluierung wird zum neuen Compute-Engpass — GAIA-Einzellauf 2.829 $, HAL-Leaderboard 40.000 $, akademische Auditoren stoßen an eine Budgetgrenze vor einer technischen https://24-ai.news/de/news/2026-04-30/ibm-granite-41-dense-apache/ 24 AI de 2026-04-30T00:00:00Z IBM Granite 4.1: Open-Source-Familie mit 3B/8B/30B Apache-2.0-Modellen, trainiert auf 15B Token, zeigt, dass ein dichtes 8B-Modell ein 32B-MoE-Modell ebenbürtig ist https://24-ai.news/de/news/2026-04-30/langchain-deep-agents-harness-profiles/ 24 AI de 2026-04-30T00:00:00Z LangChain Harness-Profile für Deep Agents: GPT-5.3 Codex springt von 33 % auf 53 % im tau2-Bench, Opus 4.7 von 43 % auf 53 % https://24-ai.news/de/news/2026-04-30/mistral-medium-35-vibe-agents/ 24 AI de 2026-04-30T00:00:00Z Mistral Medium 3.5 + Vibe: 128B Dense-Modell mit offenen Gewichten und asynchrone Cloud-Coding-Agenten für $1,5/$7,5 pro Million Token https://24-ai.news/de/news/2026-04-30/nvidia-nemotron-3-nano-omni/ 24 AI de 2026-04-30T00:00:00Z NVIDIA Nemotron 3 Nano Omni: 30B-A3B MoE-Multimodal-Modell mit 9-fachem Durchsatz für KI-Agenten https://24-ai.news/de/news/2026-04-30/pytorch-autosp-sequence-parallelism/ 24 AI de 2026-04-30T00:00:00Z PyTorch AutoSP: compilierbasiertes Tool konvertiert Training-Code automatisch in Sequence-Parallel-Variante für 100k+-Token-Kontexte https://24-ai.news/zh/news/2026-04-30/anthropic-biomysterybench-claude-mythos/ 24 AI zh 2026-04-30T00:00:00Z BioMysteryBench：Mythos Preview 解决连专家都无法解答的生物信息学难题，Opus 4.6 在人类可解任务上达到 77.4% https://24-ai.news/zh/news/2026-04-30/anthropic-claude-creative-work-blender-adobe/ 24 AI zh 2026-04-30T00:00:00Z Anthropic Claude for Creative Work：支持 Blender、50 余款 Adobe Creative Cloud 工具、Autodesk Fusion、Ableton、SketchUp 和 Splice https://24-ai.news/zh/news/2026-04-30/arxiv-cross-lingual-jailbreak-codebooks/ 24 AI zh 2026-04-30T00:00:00Z ArXiv：针对跨语言越狱攻击的无训练防护在精心策划的基准测试上达到AUC 0.99，但在分布偏移时降至0.60-0.70 https://24-ai.news/zh/news/2026-04-30/arxiv-odysseys-web-agents-benchmark/ 24 AI zh 2026-04-30T00:00:00Z ArXiv Odysseys：卡内基梅隆大学真实网络智能体基准测试显示，顶尖前沿模型在长程任务上仅达到44.5%成功率和1.15%轨迹效率 https://24-ai.news/zh/news/2026-04-30/arxiv-reasoning-barriers-transformers/ 24 AI zh 2026-04-30T00:00:00Z ArXiv：带有思维链的标准Transformer无法在TC^0复杂度之外进行推理——信号柱词元实现长度可泛化的图灵模拟 https://24-ai.news/zh/news/2026-04-30/aws-agentcore-mcp-proxy-serverless/ 24 AI zh 2026-04-30T00:00:00Z AWS Bedrock AgentCore：为企业治理提供带有 IAM、OAuth 2.0 JWT 和 CloudWatch 可观测性的无服务器 MCP 代理 https://24-ai.news/zh/news/2026-04-30/aws-agentcore-memory-namespaces/ 24 AI zh 2026-04-30T00:00:00Z AWS Bedrock AgentCore Memory：命名空间级别三种长期记忆架构模式与IAM访问控制 https://24-ai.news/zh/news/2026-04-30/cncf-state-of-ai-projects-survey/ 24 AI zh 2026-04-30T00:00:00Z CNCF 项目 AI 使用状况：Claude Code 和 GitHub Copilot 领跑，三分之二的项目缺乏正式 AI 政策 https://24-ai.news/zh/news/2026-04-30/google-era-scientific-research-ai/ 24 AI zh 2026-04-30T00:00:00Z Google ERA：用于科学研究的AI系统在住院预测上达到CDC顶级水平，解决了悬而未决的宇宙学问题，并每10分钟监测一次CO2 https://24-ai.news/zh/news/2026-04-30/huggingface-evaleval-cost-bottleneck/ 24 AI zh 2026-04-30T00:00:00Z EvalEval联盟：AI评估成为新的计算瓶颈——GAIA单次运行$2,829，HAL排行榜$40,000，学术审计人员在遇到技术瓶颈前先碰壁于预算 https://24-ai.news/zh/news/2026-04-30/ibm-granite-41-dense-apache/ 24 AI zh 2026-04-30T00:00:00Z IBM Granite 4.1：在150亿词元上训练的3B/8B/30B Apache 2.0开源模型家族，稠密8B可与32B MoE媲美 https://24-ai.news/zh/news/2026-04-30/langchain-deep-agents-harness-profiles/ 24 AI zh 2026-04-30T00:00:00Z LangChain 为 Deep Agents 推出 harness 配置文件：GPT-5.3 Codex 在 tau2 基准上从 33% 跃升至 53%，Opus 4.7 从 43% 升至 53% https://24-ai.news/zh/news/2026-04-30/mistral-medium-35-vibe-agents/ 24 AI zh 2026-04-30T00:00:00Z Mistral Medium 3.5 + Vibe：128B稠密开放权重模型与异步云端编程智能体，每百万词元$1.5/$7.5 https://24-ai.news/zh/news/2026-04-30/nvidia-nemotron-3-nano-omni/ 24 AI zh 2026-04-30T00:00:00Z NVIDIA Nemotron 3 Nano Omni：30B-A3B MoE 多模态模型，吞吐量是同类开源模型的 9 倍 https://24-ai.news/zh/news/2026-04-30/pytorch-autosp-sequence-parallelism/ 24 AI zh 2026-04-30T00:00:00Z PyTorch AutoSP：基于编译器的工具自动将训练代码转换为序列并行，支持100k+词元上下文 https://24-ai.news/ja/news/2026-04-30/anthropic-biomysterybench-claude-mythos/ 24 AI ja 2026-04-30T00:00:00Z BioMysteryBench：Mythos Preview が専門家でも解けないバイオインフォマティクス問題を解決、Opus 4.6 は人間が解ける課題で 77.4% を達成 https://24-ai.news/ja/news/2026-04-30/anthropic-claude-creative-work-blender-adobe/ 24 AI ja 2026-04-30T00:00:00Z Anthropic Claude for Creative Work：Blender、50 以上の Adobe Creative Cloud ツール、Autodesk Fusion、Ableton、SketchUp、Splice と連携 https://24-ai.news/ja/news/2026-04-30/arxiv-cross-lingual-jailbreak-codebooks/ 24 AI ja 2026-04-30T00:00:00Z ArXiv：クロスリンガルジェイルブレークに対するトレーニング不要のガードレールが厳選ベンチマークでAUC 0.99を達成するも、分布シフト時には0.60-0.70に低下 https://24-ai.news/ja/news/2026-04-30/arxiv-odysseys-web-agents-benchmark/ 24 AI ja 2026-04-30T00:00:00Z ArXiv Odysseys：CMUによるリアルなウェブエージェントベンチマークで、最強のフロンティアモデルが長程タスクで44.5%の成功率と1.15%の軌跡効率しか達成できないことが判明 https://24-ai.news/ja/news/2026-04-30/arxiv-reasoning-barriers-transformers/ 24 AI ja 2026-04-30T00:00:00Z ArXiv：思考の連鎖を持つ標準的なトランスフォーマーはTC^0複雑性を超えた推論ができない — シグナルポストトークンが長さ汎化可能なチューリング模倣を可能に https://24-ai.news/ja/news/2026-04-30/aws-agentcore-mcp-proxy-serverless/ 24 AI ja 2026-04-30T00:00:00Z AWS Bedrock AgentCore：エンタープライズガバナンス向けに IAM、OAuth 2.0 JWT、CloudWatch 可観測性を備えたサーバーレス MCP プロキシ https://24-ai.news/ja/news/2026-04-30/aws-agentcore-memory-namespaces/ 24 AI ja 2026-04-30T00:00:00Z AWS Bedrock AgentCore Memory：IAMアクセス制御を備えた名前空間レベルのエージェント長期記憶の3つのアーキテクチャパターン https://24-ai.news/ja/news/2026-04-30/cncf-state-of-ai-projects-survey/ 24 AI ja 2026-04-30T00:00:00Z CNCF プロジェクトにおける AI の現状：Claude Code と GitHub Copilot が主流、プロジェクトの 3 分の 2 が正式な AI ポリシーを持たない https://24-ai.news/ja/news/2026-04-30/google-era-scientific-research-ai/ 24 AI ja 2026-04-30T00:00:00Z Google ERA：科学研究向けAIシステムが入院予測でCDC最高位を達成、未解決の宇宙論問題を解決、10分ごとにCO2を観測 https://24-ai.news/ja/news/2026-04-30/huggingface-evaleval-cost-bottleneck/ 24 AI ja 2026-04-30T00:00:00Z EvalEval Coalition：AI評価が新たな計算ボトルネックに — GAIAのシングルラン$2,829、HALリーダーボード$40,000、学術監査人は技術的障壁より先に予算の壁に直面 https://24-ai.news/ja/news/2026-04-30/ibm-granite-41-dense-apache/ 24 AI ja 2026-04-30T00:00:00Z IBM Granite 4.1：15Tトークンで訓練されたオープンソース3B/8B/30B Apache 2.0モデルファミリー、密な8Bモデルが32B MoEに匹敵 https://24-ai.news/ja/news/2026-04-30/langchain-deep-agents-harness-profiles/ 24 AI ja 2026-04-30T00:00:00Z LangChain の Deep Agents 向け harness プロファイル：GPT-5.3 Codex が tau2 ベンチマークで 33% から 53% へ、Opus 4.7 は 43% から 53% へ上昇 https://24-ai.news/ja/news/2026-04-30/mistral-medium-35-vibe-agents/ 24 AI ja 2026-04-30T00:00:00Z Mistral Medium 3.5 + Vibe：128Bの密なオープンウェイトモデルと非同期クラウドコーディングエージェント、100万トークンあたり$1.5/$7.5 https://24-ai.news/ja/news/2026-04-30/nvidia-nemotron-3-nano-omni/ 24 AI ja 2026-04-30T00:00:00Z NVIDIA Nemotron 3 Nano Omni：30B-A3B MoE マルチモーダルモデル、AI エージェント向けに 9 倍のスループットを実現 https://24-ai.news/ja/news/2026-04-30/pytorch-autosp-sequence-parallelism/ 24 AI ja 2026-04-30T00:00:00Z PyTorch AutoSP：コンパイラベースのツールがトレーニングコードを自動的にシーケンス並列化、100k+トークンコンテキストに対応 https://24-ai.news/ko/news/2026-04-30/anthropic-biomysterybench-claude-mythos/ 24 AI ko 2026-04-30T00:00:00Z BioMysteryBench：Mythos Preview, 전문가도 풀지 못한 생물정보학 문제 해결, Opus 4.6은 인간 해결 과제에서 77.4% 달성 https://24-ai.news/ko/news/2026-04-30/anthropic-claude-creative-work-blender-adobe/ 24 AI ko 2026-04-30T00:00:00Z Anthropic Claude for Creative Work：Blender, Adobe Creative Cloud 50개 이상 도구, Autodesk Fusion, Ableton, SketchUp, Splice 연동 https://24-ai.news/ko/news/2026-04-30/arxiv-cross-lingual-jailbreak-codebooks/ 24 AI ko 2026-04-30T00:00:00Z ArXiv：교차 언어 탈옥 공격에 대한 훈련 없는 가드레일이 정제된 벤치마크에서 AUC 0.99 달성, 분포 이동 시 0.60-0.70으로 하락 https://24-ai.news/ko/news/2026-04-30/arxiv-odysseys-web-agents-benchmark/ 24 AI ko 2026-04-30T00:00:00Z ArXiv Odysseys：CMU의 현실적인 웹 에이전트 벤치마크가 최고 프런티어 모델이 장기 작업에서 44.5% 성공률과 1.15% 궤적 효율만 달성함을 밝혀 https://24-ai.news/ko/news/2026-04-30/arxiv-reasoning-barriers-transformers/ 24 AI ko 2026-04-30T00:00:00Z ArXiv：사고의 연쇄를 갖춘 표준 트랜스포머는 TC^0 복잡도를 벗어난 추론 불가 — 신호등 토큰이 길이 일반화 가능한 튜링 시뮬레이션 실현 https://24-ai.news/ko/news/2026-04-30/aws-agentcore-mcp-proxy-serverless/ 24 AI ko 2026-04-30T00:00:00Z AWS Bedrock AgentCore：엔터프라이즈 거버넌스를 위한 IAM, OAuth 2.0 JWT, CloudWatch 관찰 가능성을 갖춘 서버리스 MCP 프록시 https://24-ai.news/ko/news/2026-04-30/aws-agentcore-memory-namespaces/ 24 AI ko 2026-04-30T00:00:00Z AWS Bedrock AgentCore Memory：IAM 액세스 제어를 갖춘 네임스페이스 수준 에이전트 장기 메모리 세 가지 아키텍처 패턴 https://24-ai.news/ko/news/2026-04-30/cncf-state-of-ai-projects-survey/ 24 AI ko 2026-04-30T00:00:00Z CNCF 프로젝트 AI 현황：Claude Code와 GitHub Copilot이 지배, 프로젝트의 3분의 2가 공식 AI 정책 없어 https://24-ai.news/ko/news/2026-04-30/google-era-scientific-research-ai/ 24 AI ko 2026-04-30T00:00:00Z Google ERA：과학 연구용 AI 시스템이 입원 예측에서 CDC 최고 순위 달성, 미해결 우주론 문제 해결, 10분마다 CO2 관측 https://24-ai.news/ko/news/2026-04-30/huggingface-evaleval-cost-bottleneck/ 24 AI ko 2026-04-30T00:00:00Z EvalEval Coalition：AI 평가가 새로운 컴퓨팅 병목 현상으로 — GAIA 단일 실행 $2,829, HAL 리더보드 $40,000, 학술 감사인들은 기술적 장벽 전에 예산 장벽에 직면 https://24-ai.news/ko/news/2026-04-30/ibm-granite-41-dense-apache/ 24 AI ko 2026-04-30T00:00:00Z IBM Granite 4.1：15T 토큰으로 학습된 오픈 소스 3B/8B/30B Apache 2.0 모델 패밀리, 밀집 8B 모델이 32B MoE에 필적 https://24-ai.news/ko/news/2026-04-30/langchain-deep-agents-harness-profiles/ 24 AI ko 2026-04-30T00:00:00Z LangChain Deep Agents 하네스 프로파일：GPT-5.3 Codex, tau2 벤치마크에서 33%에서 53%로, Opus 4.7은 43%에서 53%로 도약 https://24-ai.news/ko/news/2026-04-30/mistral-medium-35-vibe-agents/ 24 AI ko 2026-04-30T00:00:00Z Mistral Medium 3.5 + Vibe：128B 밀집 오픈 가중치 모델과 비동기 클라우드 코딩 에이전트, 백만 토큰당 $1.5/$7.5 https://24-ai.news/ko/news/2026-04-30/nvidia-nemotron-3-nano-omni/ 24 AI ko 2026-04-30T00:00:00Z NVIDIA Nemotron 3 Nano Omni：AI 에이전트용 30B-A3B MoE 멀티모달 모델, 처리량 9배 향상 https://24-ai.news/ko/news/2026-04-30/pytorch-autosp-sequence-parallelism/ 24 AI ko 2026-04-30T00:00:00Z PyTorch AutoSP：컴파일러 기반 도구가 학습 코드를 자동으로 시퀀스 병렬화하여 100k+ 토큰 컨텍스트 지원