https://24-ai.news/hr/news/2026-05-23/amd-gluon-gemm-kernels-mi355/ 24 AI hr 2026-05-23T00:00:00Z AMD: Gluon block-level model omogućuje GEMM kernele s 5.255 TFLOPS MXFP4 na Instinct MI355 https://24-ai.news/hr/news/2026-05-23/anthropic-glasswing-vulnerability-discovery/ 24 AI hr 2026-05-23T00:00:00Z Anthropic: Project Glasswing pronašao 10.000 visokorizičnih ranjivosti u prvom mjesecu rada s Claude Mythos Preview https://24-ai.news/hr/news/2026-05-23/arxiv-compiling-workflows-llm-weights/ 24 AI hr 2026-05-23T00:00:00Z arXiv:2605.22502: Kompiliranjem agentskih workflowa u LLM težine postiže se near-frontier kvaliteta uz 100 puta nižu cijenu https://24-ai.news/hr/news/2026-05-23/arxiv-cusp-ai-cannot-predict-breakthroughs/ 24 AI hr 2026-05-23T00:00:00Z arXiv:2605.22681: CUSP benchmark pokazuje da frontier modeli ne mogu pouzdano predviđati znanstvene proboje https://24-ai.news/hr/news/2026-05-23/arxiv-formal-proof-search-erdos-problems/ 24 AI hr 2026-05-23T00:00:00Z arXiv:2605.22763: AI agent s Lean verifikacijom rješava 9 otvorenih Erdősovih problema i 44 OEIS konjekture https://24-ai.news/hr/news/2026-05-23/arxiv-lcguard-multi-agent-kv-cache-security/ 24 AI hr 2026-05-23T00:00:00Z arXiv:2605.22786: LCGuard štiti dijeljeni KV cache između agenata u multi-agent sustavima od curenja podataka https://24-ai.news/hr/news/2026-05-23/arxiv-meta-soft-kv-cache-compression/ 24 AI hr 2026-05-23T00:00:00Z arXiv:2605.22337: Meta-Soft uvodi kompresiju KV cachea kroz composable meta-tokene i naučljive ortogonalne baze https://24-ai.news/hr/news/2026-05-23/arxiv-moss-self-evolving-agents-source-code/ 24 AI hr 2026-05-23T00:00:00Z arXiv:2605.22794: MOSS pokazuje agente koji se sami unapređuju mijenjajući vlastiti izvorni kod https://24-ai.news/hr/news/2026-05-23/arxiv-terminalworld-llm-agent-terminal-benchmark/ 24 AI hr 2026-05-23T00:00:00Z arXiv:2605.22535: TerminalWorld benchmark mjeri LLM agente na stvarnim Linux terminal zadacima bez simulacije https://24-ai.news/hr/news/2026-05-23/arxiv-workstreambench-spreadsheet-agents/ 24 AI hr 2026-05-23T00:00:00Z arXiv:2605.22664: WorkstreamBench testira LLM agente na end-to-end spreadsheet zadacima u financijama i frontier modeli padaju https://24-ai.news/hr/news/2026-05-23/claude-code-2-1-149-usage-security/ 24 AI hr 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.149 donosi per-category breakdown u /usage i zatvara PowerShell permission bypass https://24-ai.news/hr/news/2026-05-23/claude-code-2-1-150-infra-patch/ 24 AI hr 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.150 — interni infrastrukturni patch bez korisničkih promjena https://24-ai.news/hr/news/2026-05-23/github-gartner-magic-quadrant-leader-2026/ 24 AI hr 2026-05-23T00:00:00Z GitHub: Gartner Magic Quadrant 2026 — GitHub Copilot Leader treću godinu zaredom za Enterprise AI Coding Agents https://24-ai.news/hr/news/2026-05-23/kedro-1-2-0-langgraph-agentic-pipelines/ 24 AI hr 2026-05-23T00:00:00Z Kedro: verzija 1.2.0 donosi @experimental dekorator i LangGraph agentic starter za GenAI pipeline-e https://24-ai.news/hr/news/2026-05-23/npm-staged-publishing-install-controls/ 24 AI hr 2026-05-23T00:00:00Z GitHub: npm 11.15.0 uvodi staged publishing i tri nova install-time --allow flaga za supply chain hardening https://24-ai.news/en/news/2026-05-23/amd-gluon-gemm-kernels-mi355/ 24 AI en 2026-05-23T00:00:00Z AMD: Gluon block-level model enables GEMM kernels with 5.255 TFLOPS MXFP4 on Instinct MI355 https://24-ai.news/en/news/2026-05-23/anthropic-glasswing-vulnerability-discovery/ 24 AI en 2026-05-23T00:00:00Z Anthropic: Project Glasswing found 10,000 high-risk vulnerabilities in its first month using Claude Mythos Preview https://24-ai.news/en/news/2026-05-23/arxiv-compiling-workflows-llm-weights/ 24 AI en 2026-05-23T00:00:00Z arXiv:2605.22502: Compiling agentic workflows into LLM weights achieves near-frontier quality at 100× lower cost https://24-ai.news/en/news/2026-05-23/arxiv-cusp-ai-cannot-predict-breakthroughs/ 24 AI en 2026-05-23T00:00:00Z arXiv:2605.22681: CUSP benchmark shows frontier models cannot reliably predict scientific breakthroughs https://24-ai.news/en/news/2026-05-23/arxiv-formal-proof-search-erdos-problems/ 24 AI en 2026-05-23T00:00:00Z arXiv:2605.22763: AI agent with Lean verification solves 9 open Erdős problems and 44 OEIS conjectures https://24-ai.news/en/news/2026-05-23/arxiv-lcguard-multi-agent-kv-cache-security/ 24 AI en 2026-05-23T00:00:00Z arXiv:2605.22786: LCGuard protects shared KV cache between agents in multi-agent systems from data leakage https://24-ai.news/en/news/2026-05-23/arxiv-meta-soft-kv-cache-compression/ 24 AI en 2026-05-23T00:00:00Z arXiv:2605.22337: Meta-Soft introduces KV cache compression via composable meta-tokens and learnable orthogonal bases https://24-ai.news/en/news/2026-05-23/arxiv-moss-self-evolving-agents-source-code/ 24 AI en 2026-05-23T00:00:00Z arXiv:2605.22794: MOSS shows agents that self-improve by rewriting their own source code https://24-ai.news/en/news/2026-05-23/arxiv-terminalworld-llm-agent-terminal-benchmark/ 24 AI en 2026-05-23T00:00:00Z arXiv:2605.22535: TerminalWorld benchmark measures LLM agents on real Linux terminal tasks without simulation https://24-ai.news/en/news/2026-05-23/arxiv-workstreambench-spreadsheet-agents/ 24 AI en 2026-05-23T00:00:00Z arXiv:2605.22664: WorkstreamBench tests LLM agents on end-to-end spreadsheet tasks in finance — and frontier models fail https://24-ai.news/en/news/2026-05-23/claude-code-2-1-149-usage-security/ 24 AI en 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.149 brings per-category breakdown in /usage and closes PowerShell permission bypass https://24-ai.news/en/news/2026-05-23/claude-code-2-1-150-infra-patch/ 24 AI en 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.150 — internal infrastructure patch with no user-facing changes https://24-ai.news/en/news/2026-05-23/github-gartner-magic-quadrant-leader-2026/ 24 AI en 2026-05-23T00:00:00Z GitHub: Gartner Magic Quadrant 2026 — GitHub Copilot Leader for the third consecutive year in Enterprise AI Coding Agents https://24-ai.news/en/news/2026-05-23/kedro-1-2-0-langgraph-agentic-pipelines/ 24 AI en 2026-05-23T00:00:00Z Kedro: version 1.2.0 brings the @experimental decorator and a LangGraph agentic starter for GenAI pipelines https://24-ai.news/en/news/2026-05-23/npm-staged-publishing-install-controls/ 24 AI en 2026-05-23T00:00:00Z GitHub: npm 11.15.0 introduces staged publishing and three new install-time --allow flags for supply chain hardening https://24-ai.news/de/news/2026-05-23/amd-gluon-gemm-kernels-mi355/ 24 AI de 2026-05-23T00:00:00Z AMD: Gluon Block-Level-Modell ermöglicht GEMM-Kernels mit 5,255 TFLOPS MXFP4 auf Instinct MI355 https://24-ai.news/de/news/2026-05-23/anthropic-glasswing-vulnerability-discovery/ 24 AI de 2026-05-23T00:00:00Z Anthropic: Project Glasswing findet 10.000 hochriskante Schwachstellen im ersten Monat mit Claude Mythos Preview https://24-ai.news/de/news/2026-05-23/arxiv-compiling-workflows-llm-weights/ 24 AI de 2026-05-23T00:00:00Z arXiv:2605.22502: Kompilierung agentischer Workflows in LLM-Gewichte liefert Near-Frontier-Qualität bei 100× niedrigeren Kosten https://24-ai.news/de/news/2026-05-23/arxiv-cusp-ai-cannot-predict-breakthroughs/ 24 AI de 2026-05-23T00:00:00Z arXiv:2605.22681: CUSP-Benchmark zeigt, dass Frontier-Modelle wissenschaftliche Durchbrüche nicht zuverlässig vorhersagen können https://24-ai.news/de/news/2026-05-23/arxiv-formal-proof-search-erdos-problems/ 24 AI de 2026-05-23T00:00:00Z arXiv:2605.22763: KI-Agent mit Lean-Verifikation löst 9 offene Erdős-Probleme und 44 OEIS-Vermutungen https://24-ai.news/de/news/2026-05-23/arxiv-lcguard-multi-agent-kv-cache-security/ 24 AI de 2026-05-23T00:00:00Z arXiv:2605.22786: LCGuard schützt geteilten KV-Cache zwischen Agenten in Multi-Agenten-Systemen vor Datenlecks https://24-ai.news/de/news/2026-05-23/arxiv-meta-soft-kv-cache-compression/ 24 AI de 2026-05-23T00:00:00Z arXiv:2605.22337: Meta-Soft führt KV-Cache-Komprimierung mit komponierbaren Meta-Token und lernbaren orthogonalen Basen ein https://24-ai.news/de/news/2026-05-23/arxiv-moss-self-evolving-agents-source-code/ 24 AI de 2026-05-23T00:00:00Z arXiv:2605.22794: MOSS zeigt Agenten, die sich durch Umschreiben ihres eigenen Quellcodes verbessern https://24-ai.news/de/news/2026-05-23/arxiv-terminalworld-llm-agent-terminal-benchmark/ 24 AI de 2026-05-23T00:00:00Z arXiv:2605.22535: TerminalWorld-Benchmark misst KI-Agenten auf echten Linux-Terminal-Aufgaben ohne Simulation https://24-ai.news/de/news/2026-05-23/arxiv-workstreambench-spreadsheet-agents/ 24 AI de 2026-05-23T00:00:00Z arXiv:2605.22664: WorkstreamBench testet KI-Agenten auf End-to-End-Tabellenkalkulationsaufgaben im Finanzbereich — und Frontier-Modelle scheitern https://24-ai.news/de/news/2026-05-23/claude-code-2-1-149-usage-security/ 24 AI de 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.149 bringt kategorienweise /usage-Aufschlüsselung und schließt PowerShell-Permission-Bypass https://24-ai.news/de/news/2026-05-23/claude-code-2-1-150-infra-patch/ 24 AI de 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.150 — interner Infrastruktur-Patch ohne nutzerseitige Änderungen https://24-ai.news/de/news/2026-05-23/github-gartner-magic-quadrant-leader-2026/ 24 AI de 2026-05-23T00:00:00Z GitHub: Gartner Magic Quadrant 2026 — GitHub Copilot zum dritten Mal Leader bei Enterprise AI Coding Agents https://24-ai.news/de/news/2026-05-23/kedro-1-2-0-langgraph-agentic-pipelines/ 24 AI de 2026-05-23T00:00:00Z Kedro: Version 1.2.0 bringt @experimental-Dekorator und LangGraph Agentic Starter für GenAI-Pipelines https://24-ai.news/de/news/2026-05-23/npm-staged-publishing-install-controls/ 24 AI de 2026-05-23T00:00:00Z GitHub: npm 11.15.0 führt Staged Publishing und drei neue Install-Zeit --allow-Flags für Supply-Chain-Hardening ein https://24-ai.news/zh/news/2026-05-23/amd-gluon-gemm-kernels-mi355/ 24 AI zh 2026-05-23T00:00:00Z AMD：Gluon块级模型在Instinct MI355上实现MXFP4 5.255 TFLOPS GEMM内核 https://24-ai.news/zh/news/2026-05-23/anthropic-glasswing-vulnerability-discovery/ 24 AI zh 2026-05-23T00:00:00Z Anthropic：Project Glasswing首月发现超过10,000个高风险漏洞 https://24-ai.news/zh/news/2026-05-23/arxiv-compiling-workflows-llm-weights/ 24 AI zh 2026-05-23T00:00:00Z arXiv:2605.22502：将智能体工作流编译进LLM权重，以百分之一的成本实现接近前沿的质量 https://24-ai.news/zh/news/2026-05-23/arxiv-cusp-ai-cannot-predict-breakthroughs/ 24 AI zh 2026-05-23T00:00:00Z arXiv:2605.22681：CUSP基准测试显示前沿模型无法可靠预测科学突破 https://24-ai.news/zh/news/2026-05-23/arxiv-formal-proof-search-erdos-problems/ 24 AI zh 2026-05-23T00:00:00Z arXiv:2605.22763：AI智能体结合Lean验证解决9个开放Erdős问题和44个OEIS猜想 https://24-ai.news/zh/news/2026-05-23/arxiv-lcguard-multi-agent-kv-cache-security/ 24 AI zh 2026-05-23T00:00:00Z arXiv:2605.22786：LCGuard保护多智能体系统中共享KV缓存，防止智能体间数据泄露 https://24-ai.news/zh/news/2026-05-23/arxiv-meta-soft-kv-cache-compression/ 24 AI zh 2026-05-23T00:00:00Z arXiv:2605.22337：Meta-Soft通过可组合元令牌和可学习正交基引入KV缓存压缩 https://24-ai.news/zh/news/2026-05-23/arxiv-moss-self-evolving-agents-source-code/ 24 AI zh 2026-05-23T00:00:00Z arXiv:2605.22794：MOSS展示通过修改自身源代码进行自我改进的智能体 https://24-ai.news/zh/news/2026-05-23/arxiv-terminalworld-llm-agent-terminal-benchmark/ 24 AI zh 2026-05-23T00:00:00Z arXiv:2605.22535：TerminalWorld基准测试在真实Linux终端任务中评估LLM智能体，无需模拟环境 https://24-ai.news/zh/news/2026-05-23/arxiv-workstreambench-spreadsheet-agents/ 24 AI zh 2026-05-23T00:00:00Z arXiv:2605.22664：WorkstreamBench测试LLM智能体处理金融端到端电子表格任务，前沿模型表现不佳 https://24-ai.news/zh/news/2026-05-23/claude-code-2-1-149-usage-security/ 24 AI zh 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.149带来/usage分类明细并修复PowerShell权限绕过 https://24-ai.news/zh/news/2026-05-23/claude-code-2-1-150-infra-patch/ 24 AI zh 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.150 — 内部基础设施补丁，无用户可见更改 https://24-ai.news/zh/news/2026-05-23/github-gartner-magic-quadrant-leader-2026/ 24 AI zh 2026-05-23T00:00:00Z GitHub：2026年Gartner魔力象限——GitHub Copilot连续第三年荣获企业AI编码代理领导者 https://24-ai.news/zh/news/2026-05-23/kedro-1-2-0-langgraph-agentic-pipelines/ 24 AI zh 2026-05-23T00:00:00Z Kedro：1.2.0版本带来@experimental装饰器和LangGraph智能体starter用于GenAI管道 https://24-ai.news/zh/news/2026-05-23/npm-staged-publishing-install-controls/ 24 AI zh 2026-05-23T00:00:00Z GitHub：npm 11.15.0引入分阶段发布和三个新的安装时--allow标志用于供应链加固 https://24-ai.news/ja/news/2026-05-23/amd-gluon-gemm-kernels-mi355/ 24 AI ja 2026-05-23T00:00:00Z AMD：GluonブロックレベルモデルでInstinct MI355のMXFP4 5.255 TFLOPSを実現 https://24-ai.news/ja/news/2026-05-23/anthropic-glasswing-vulnerability-discovery/ 24 AI ja 2026-05-23T00:00:00Z Anthropic：Project Glasswingが初月に1万件超の高リスク脆弱性を発見 https://24-ai.news/ja/news/2026-05-23/arxiv-compiling-workflows-llm-weights/ 24 AI ja 2026-05-23T00:00:00Z arXiv:2605.22502：エージェントワークフローをLLM重みにコンパイルして100分の1のコストでフロンティア近似品質を実現 https://24-ai.news/ja/news/2026-05-23/arxiv-cusp-ai-cannot-predict-breakthroughs/ 24 AI ja 2026-05-23T00:00:00Z arXiv:2605.22681：CUSPベンチマークがフロンティアモデルは科学的ブレークスルーを信頼できるほど予測できないことを示す https://24-ai.news/ja/news/2026-05-23/arxiv-formal-proof-search-erdos-problems/ 24 AI ja 2026-05-23T00:00:00Z arXiv:2605.22763：AI エージェントがLean検証で9つの開放Erdős問題と44のOEIS予想を解決 https://24-ai.news/ja/news/2026-05-23/arxiv-lcguard-multi-agent-kv-cache-security/ 24 AI ja 2026-05-23T00:00:00Z arXiv:2605.22786：LCGuardがマルチエージェントシステムにおける共有KVキャッシュをエージェント間のデータ漏洩から保護します https://24-ai.news/ja/news/2026-05-23/arxiv-meta-soft-kv-cache-compression/ 24 AI ja 2026-05-23T00:00:00Z arXiv:2605.22337：Meta-Softが組合せ可能なメタトークンと学習可能な直交基でKVキャッシュ圧縮を実現 https://24-ai.news/ja/news/2026-05-23/arxiv-moss-self-evolving-agents-source-code/ 24 AI ja 2026-05-23T00:00:00Z arXiv:2605.22794：MOSSが自身のソースコードを修正して自己改善するエージェントを実証 https://24-ai.news/ja/news/2026-05-23/arxiv-terminalworld-llm-agent-terminal-benchmark/ 24 AI ja 2026-05-23T00:00:00Z arXiv:2605.22535：TerminalWorldベンチマークがシミュレーションなしで実際のLinuxターミナルタスクにおけるLLMエージェントを評価します https://24-ai.news/ja/news/2026-05-23/arxiv-workstreambench-spreadsheet-agents/ 24 AI ja 2026-05-23T00:00:00Z arXiv:2605.22664：WorkstreamBenchが金融分野のエンドツーエンドスプレッドシートタスクでLLMエージェントをテストし、フロンティアモデルが苦戦しています https://24-ai.news/ja/news/2026-05-23/claude-code-2-1-149-usage-security/ 24 AI ja 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.149が/usageカテゴリ内訳とPowerShellパーミッションバイパス修正を提供 https://24-ai.news/ja/news/2026-05-23/claude-code-2-1-150-infra-patch/ 24 AI ja 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.150 — ユーザー向けの変更なしの内部インフラストラクチャパッチ https://24-ai.news/ja/news/2026-05-23/github-gartner-magic-quadrant-leader-2026/ 24 AI ja 2026-05-23T00:00:00Z GitHub：2026年Gartnerマジッククアドラント——GitHub Copilotが企業AIコーディングエージェントで3年連続リーダーに https://24-ai.news/ja/news/2026-05-23/kedro-1-2-0-langgraph-agentic-pipelines/ 24 AI ja 2026-05-23T00:00:00Z Kedro：バージョン1.2.0が@experimentalデコレータとLangGraphエージェントstarterをGenAIパイプライン向けに提供 https://24-ai.news/ja/news/2026-05-23/npm-staged-publishing-install-controls/ 24 AI ja 2026-05-23T00:00:00Z GitHub：npm 11.15.0がサプライチェーン強化のためのステージング公開と3つの新しいインストール時--allowフラグを導入 https://24-ai.news/ko/news/2026-05-23/amd-gluon-gemm-kernels-mi355/ 24 AI ko 2026-05-23T00:00:00Z AMD：Gluon 블록 수준 모델로 Instinct MI355에서 MXFP4 5.255 TFLOPS GEMM 커널 구현 https://24-ai.news/ko/news/2026-05-23/anthropic-glasswing-vulnerability-discovery/ 24 AI ko 2026-05-23T00:00:00Z Anthropic：Project Glasswing이 첫 달에 10,000개 이상의 고위험 취약점 발견 https://24-ai.news/ko/news/2026-05-23/arxiv-compiling-workflows-llm-weights/ 24 AI ko 2026-05-23T00:00:00Z arXiv:2605.22502：에이전트 워크플로를 LLM 가중치로 컴파일하여 100분의 1 비용으로 프런티어 근접 품질 달성 https://24-ai.news/ko/news/2026-05-23/arxiv-cusp-ai-cannot-predict-breakthroughs/ 24 AI ko 2026-05-23T00:00:00Z arXiv:2605.22681：CUSP 벤치마크, 프런티어 모델이 과학적 돌파구를 신뢰할 수 있게 예측하지 못함을 입증 https://24-ai.news/ko/news/2026-05-23/arxiv-formal-proof-search-erdos-problems/ 24 AI ko 2026-05-23T00:00:00Z arXiv:2605.22763：AI 에이전트가 Lean 검증으로 9개의 개방 Erdős 문제와 44개의 OEIS 추측 해결 https://24-ai.news/ko/news/2026-05-23/arxiv-lcguard-multi-agent-kv-cache-security/ 24 AI ko 2026-05-23T00:00:00Z arXiv:2605.22786：LCGuard, 멀티에이전트 시스템에서 공유 KV 캐시를 에이전트 간 데이터 유출로부터 보호합니다 https://24-ai.news/ko/news/2026-05-23/arxiv-meta-soft-kv-cache-compression/ 24 AI ko 2026-05-23T00:00:00Z arXiv:2605.22337：Meta-Soft, 조합 가능한 메타 토큰과 학습 가능한 직교 기저로 KV 캐시 압축 도입 https://24-ai.news/ko/news/2026-05-23/arxiv-moss-self-evolving-agents-source-code/ 24 AI ko 2026-05-23T00:00:00Z arXiv:2605.22794：MOSS, 자신의 소스 코드를 수정하여 자기 개선하는 에이전트 시연 https://24-ai.news/ko/news/2026-05-23/arxiv-terminalworld-llm-agent-terminal-benchmark/ 24 AI ko 2026-05-23T00:00:00Z arXiv:2605.22535：TerminalWorld 벤치마크, 시뮬레이션 없이 실제 Linux 터미널 작업에서 LLM 에이전트를 평가합니다 https://24-ai.news/ko/news/2026-05-23/arxiv-workstreambench-spreadsheet-agents/ 24 AI ko 2026-05-23T00:00:00Z arXiv:2605.22664：WorkstreamBench, 금융 분야 엔드투엔드 스프레드시트 작업에서 LLM 에이전트를 테스트하고 프론티어 모델은 부진한 결과를 보입니다 https://24-ai.news/ko/news/2026-05-23/claude-code-2-1-149-usage-security/ 24 AI ko 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.149, /usage 카테고리 분류와 PowerShell 권한 우회 수정 제공 https://24-ai.news/ko/news/2026-05-23/claude-code-2-1-150-infra-patch/ 24 AI ko 2026-05-23T00:00:00Z Anthropic Claude Code v2.1.150 — 사용자 변경 사항 없는 내부 인프라 패치 https://24-ai.news/ko/news/2026-05-23/github-gartner-magic-quadrant-leader-2026/ 24 AI ko 2026-05-23T00:00:00Z GitHub：2026년 Gartner 매직 쿼드런트——GitHub Copilot이 기업용 AI 코딩 에이전트에서 3년 연속 리더 선정 https://24-ai.news/ko/news/2026-05-23/kedro-1-2-0-langgraph-agentic-pipelines/ 24 AI ko 2026-05-23T00:00:00Z Kedro：버전 1.2.0, GenAI 파이프라인을 위한 @experimental 데코레이터와 LangGraph 에이전트 starter 제공 https://24-ai.news/ko/news/2026-05-23/npm-staged-publishing-install-controls/ 24 AI ko 2026-05-23T00:00:00Z GitHub：npm 11.15.0, 공급망 강화를 위한 단계적 게시와 세 가지 새로운 설치 시 --allow 플래그 도입 https://24-ai.news/hr/news/2026-05-22/aisi-ai-oversight-erosion-report-2026/ 24 AI hr 2026-05-22T00:00:00Z UK AI Safety Institute: Nadzor nad naprednim AI sustavima postaje sve teži — 20+ putanja degradacije identificirano https://24-ai.news/hr/news/2026-05-22/arxiv-pals-power-aware-llm-serving-moe-energy/ 24 AI hr 2026-05-22T00:00:00Z arXiv:2605.21427: PALS — power-aware LLM serving za MoE modele postiže +26.3% energetske efikasnosti i 4-7× manje QoS povreda https://24-ai.news/hr/news/2026-05-22/arxiv-persona-vectors-sycophancy-steering-llm-alignment/ 24 AI hr 2026-05-22T00:00:00Z arXiv:2605.21006: Off-the-shelf persona vektori postižu 68-98% učinkovitosti targetiranog sycophancy steeringa u LLM modelima https://24-ai.news/hr/news/2026-05-22/aws-nova-act-hipaa-eligible-healthcare-agents/ 24 AI hr 2026-05-22T00:00:00Z AWS: Nova Act dobio HIPAA-eligibility — agentska automatizacija ePHI za zdravstvene radne tokove https://24-ai.news/hr/news/2026-05-22/claude-code-workflow-tool-multi-agent-orchestration/ 24 AI hr 2026-05-22T00:00:00Z Anthropic: Claude Code v2.1.147 uvodi Workflow tool za deterministički multi-agent orchestration https://24-ai.news/hr/news/2026-05-22/cncf-fluid-llm-cold-start-30s-kubernetes/ 24 AI hr 2026-05-22T00:00:00Z CNCF: NetEase Games postigao 30-sekundni LLM cold start na Kubernetesu preko Fluid prefetching slojeva https://24-ai.news/hr/news/2026-05-22/flux-erase-object-removal-scene-reconstruction/ 24 AI hr 2026-05-22T00:00:00Z Black Forest Labs: FLUX Erase nadmašuje GPT Image-2 (68.5%) i Finegrain (63.2%) u uklanjanju objekata bez prompta https://24-ai.news/hr/news/2026-05-22/langchain-agent-streams-typed-channels-langgraph/ 24 AI hr 2026-05-22T00:00:00Z LangChain: Od token streama do agent streama — tipizirani kanali zamjenjuju klasično streamanje za multi-agent UI https://24-ai.news/hr/news/2026-05-22/microsoft-magenticlite-fara15-small-model-agents/ 24 AI hr 2026-05-22T00:00:00Z Microsoft Research: MagenticLite + Fara1.5 (4B/9B/27B) — agentic AI optimiziran za male modele postiže SOTA https://24-ai.news/hr/news/2026-05-22/microsoft-vega-zk-proofs-digital-identity-ai-agents/ 24 AI hr 2026-05-22T00:00:00Z Microsoft Research: Vega — ZK proofs za digitalni identitet, 92ms generiranje i 70% brži ponovljeni dokazi https://24-ai.news/hr/news/2026-05-22/oecd-collective-ai-security-foundations-g7/ 24 AI hr 2026-05-22T00:00:00Z OECD AI: Kolektivna AI sigurnost zahtijeva G7 koordinaciju — prompt injection, agent sigurnost i model poisoning prioriteti https://24-ai.news/hr/news/2026-05-22/openai-codex-enterprise-scale-4m-users/ 24 AI hr 2026-05-22T00:00:00Z OpenAI: Codex skaliranje na enterprise — 4 milijuna tjednih aktivnih korisnika i Codex Labs program https://24-ai.news/en/news/2026-05-22/aisi-ai-oversight-erosion-report-2026/ 24 AI en 2026-05-22T00:00:00Z UK AI Safety Institute: Overseeing advanced AI systems is becoming harder — 20+ degradation pathways identified https://24-ai.news/en/news/2026-05-22/arxiv-pals-power-aware-llm-serving-moe-energy/ 24 AI en 2026-05-22T00:00:00Z arXiv:2605.21427: PALS — power-aware LLM serving for MoE models achieves +26.3% energy efficiency and 4-7× fewer QoS violations https://24-ai.news/en/news/2026-05-22/arxiv-persona-vectors-sycophancy-steering-llm-alignment/ 24 AI en 2026-05-22T00:00:00Z arXiv:2605.21006: Off-the-shelf persona vectors achieve 68-98% effectiveness of targeted sycophancy steering in LLM models https://24-ai.news/en/news/2026-05-22/aws-nova-act-hipaa-eligible-healthcare-agents/ 24 AI en 2026-05-22T00:00:00Z AWS: Nova Act receives HIPAA eligibility — agentic ePHI automation for healthcare workflows https://24-ai.news/en/news/2026-05-22/claude-code-workflow-tool-multi-agent-orchestration/ 24 AI en 2026-05-22T00:00:00Z Anthropic: Claude Code v2.1.147 introduces Workflow tool for deterministic multi-agent orchestration https://24-ai.news/en/news/2026-05-22/cncf-fluid-llm-cold-start-30s-kubernetes/ 24 AI en 2026-05-22T00:00:00Z CNCF: NetEase Games achieves 30-second LLM cold start on Kubernetes via Fluid prefetching layers https://24-ai.news/en/news/2026-05-22/flux-erase-object-removal-scene-reconstruction/ 24 AI en 2026-05-22T00:00:00Z Black Forest Labs: FLUX Erase outperforms GPT Image-2 (68.5%) and Finegrain (63.2%) in prompt-free object removal https://24-ai.news/en/news/2026-05-22/langchain-agent-streams-typed-channels-langgraph/ 24 AI en 2026-05-22T00:00:00Z LangChain: From token streams to agent streams — typed channels replace classic streaming for multi-agent UI https://24-ai.news/en/news/2026-05-22/microsoft-magenticlite-fara15-small-model-agents/ 24 AI en 2026-05-22T00:00:00Z Microsoft Research: MagenticLite + Fara1.5 (4B/9B/27B) — agentic AI optimised for small models achieves SOTA https://24-ai.news/en/news/2026-05-22/microsoft-vega-zk-proofs-digital-identity-ai-agents/ 24 AI en 2026-05-22T00:00:00Z Microsoft Research: Vega — ZK proofs for digital identity, 92ms generation and 70% faster repeated proofs https://24-ai.news/en/news/2026-05-22/oecd-collective-ai-security-foundations-g7/ 24 AI en 2026-05-22T00:00:00Z OECD AI: Collective AI security requires G7 coordination — prompt injection, agent security, and model poisoning as priorities https://24-ai.news/en/news/2026-05-22/openai-codex-enterprise-scale-4m-users/ 24 AI en 2026-05-22T00:00:00Z OpenAI: Codex scaling to enterprise — 4 million weekly active users and the Codex Labs program https://24-ai.news/de/news/2026-05-22/aisi-ai-oversight-erosion-report-2026/ 24 AI de 2026-05-22T00:00:00Z UK AI Safety Institute: Aufsicht über fortgeschrittene KI-Systeme wird schwieriger — 20+ Degradierungspfade identifiziert https://24-ai.news/de/news/2026-05-22/arxiv-pals-power-aware-llm-serving-moe-energy/ 24 AI de 2026-05-22T00:00:00Z arXiv:2605.21427: PALS — energiebewusstes LLM-Serving für MoE-Modelle erreicht +26,3 % Energieeffizienz und 4-7× weniger QoS-Verstöße https://24-ai.news/de/news/2026-05-22/arxiv-persona-vectors-sycophancy-steering-llm-alignment/ 24 AI de 2026-05-22T00:00:00Z arXiv:2605.21006: Off-the-shelf-Persona-Vektoren erreichen 68-98 % der Wirksamkeit von gezieltem Sycophancy-Steering in LLM-Modellen https://24-ai.news/de/news/2026-05-22/aws-nova-act-hipaa-eligible-healthcare-agents/ 24 AI de 2026-05-22T00:00:00Z AWS: Nova Act erhält HIPAA-Eligibility — agentische ePHI-Automatisierung für Gesundheits-Workflows https://24-ai.news/de/news/2026-05-22/claude-code-workflow-tool-multi-agent-orchestration/ 24 AI de 2026-05-22T00:00:00Z Anthropic: Claude Code v2.1.147 führt Workflow-Tool für deterministisches Multi-Agent-Orchestrierung ein https://24-ai.news/de/news/2026-05-22/cncf-fluid-llm-cold-start-30s-kubernetes/ 24 AI de 2026-05-22T00:00:00Z CNCF: NetEase Games erreicht 30-Sekunden-LLM-Cold-Start auf Kubernetes über Fluid-Prefetching-Schichten https://24-ai.news/de/news/2026-05-22/flux-erase-object-removal-scene-reconstruction/ 24 AI de 2026-05-22T00:00:00Z Black Forest Labs: FLUX Erase übertrifft GPT Image-2 (68,5 %) und Finegrain (63,2 %) bei der promptlosen Objektentfernung https://24-ai.news/de/news/2026-05-22/langchain-agent-streams-typed-channels-langgraph/ 24 AI de 2026-05-22T00:00:00Z LangChain: Von Token-Streams zu Agent-Streams — typisierte Kanäle ersetzen klassisches Streaming für Multi-Agent-UIs https://24-ai.news/de/news/2026-05-22/microsoft-magenticlite-fara15-small-model-agents/ 24 AI de 2026-05-22T00:00:00Z Microsoft Research: MagenticLite + Fara1.5 (4B/9B/27B) — agentische KI für kleine Modelle erreicht SOTA https://24-ai.news/de/news/2026-05-22/microsoft-vega-zk-proofs-digital-identity-ai-agents/ 24 AI de 2026-05-22T00:00:00Z Microsoft Research: Vega — ZK-Proofs für digitale Identität, 92ms Generierung und 70 % schnellere wiederholte Nachweise https://24-ai.news/de/news/2026-05-22/oecd-collective-ai-security-foundations-g7/ 24 AI de 2026-05-22T00:00:00Z OECD AI: Kollektive KI-Sicherheit erfordert G7-Koordinierung — Prompt Injection, Agentensicherheit und Model Poisoning als Prioritäten https://24-ai.news/de/news/2026-05-22/openai-codex-enterprise-scale-4m-users/ 24 AI de 2026-05-22T00:00:00Z OpenAI: Codex-Skalierung auf Enterprise — 4 Millionen wöchentlich aktive Nutzer und das Codex Labs-Programm https://24-ai.news/zh/news/2026-05-22/aisi-ai-oversight-erosion-report-2026/ 24 AI zh 2026-05-22T00:00:00Z UK AI Safety Institute：先进AI系统的监管愈发困难 — 识别出20余条退化路径 https://24-ai.news/zh/news/2026-05-22/arxiv-pals-power-aware-llm-serving-moe-energy/ 24 AI zh 2026-05-22T00:00:00Z arXiv:2605.21427：PALS — MoE模型的功耗感知LLM服务实现+26.3%能效提升与4-7倍QoS违规减少 https://24-ai.news/zh/news/2026-05-22/arxiv-persona-vectors-sycophancy-steering-llm-alignment/ 24 AI zh 2026-05-22T00:00:00Z arXiv:2605.21006：现成角色向量以68-98%效率实现LLM模型中靶向谄媚引导 https://24-ai.news/zh/news/2026-05-22/aws-nova-act-hipaa-eligible-healthcare-agents/ 24 AI zh 2026-05-22T00:00:00Z AWS：Nova Act 获得HIPAA合规资格 — 面向医疗工作流的ePHI智能体自动化 https://24-ai.news/zh/news/2026-05-22/claude-code-workflow-tool-multi-agent-orchestration/ 24 AI zh 2026-05-22T00:00:00Z Anthropic：Claude Code v2.1.147 引入Workflow工具实现确定性多智能体编排 https://24-ai.news/zh/news/2026-05-22/cncf-fluid-llm-cold-start-30s-kubernetes/ 24 AI zh 2026-05-22T00:00:00Z CNCF：网易游戏通过Fluid预取层在Kubernetes上实现30秒LLM冷启动 https://24-ai.news/zh/news/2026-05-22/flux-erase-object-removal-scene-reconstruction/ 24 AI zh 2026-05-22T00:00:00Z Black Forest Labs：FLUX Erase在无提示词物体移除中超越GPT Image-2（68.5%）和Finegrain（63.2%） https://24-ai.news/zh/news/2026-05-22/langchain-agent-streams-typed-channels-langgraph/ 24 AI zh 2026-05-22T00:00:00Z LangChain：从令牌流到智能体流 — 类型化通道取代多智能体UI的经典流式传输 https://24-ai.news/zh/news/2026-05-22/microsoft-magenticlite-fara15-small-model-agents/ 24 AI zh 2026-05-22T00:00:00Z Microsoft Research：MagenticLite + Fara1.5（4B/9B/27B）— 专为小型模型优化的智能体AI达到SOTA https://24-ai.news/zh/news/2026-05-22/microsoft-vega-zk-proofs-digital-identity-ai-agents/ 24 AI zh 2026-05-22T00:00:00Z Microsoft Research：Vega — 数字身份的零知识证明，92毫秒生成与70%更快的重复证明 https://24-ai.news/zh/news/2026-05-22/oecd-collective-ai-security-foundations-g7/ 24 AI zh 2026-05-22T00:00:00Z OECD AI：集体AI安全需要G7协调 — 提示注入、智能体安全和模型投毒为优先事项 https://24-ai.news/zh/news/2026-05-22/openai-codex-enterprise-scale-4m-users/ 24 AI zh 2026-05-22T00:00:00Z OpenAI：Codex企业级扩展 — 400万周活跃用户与Codex Labs计划 https://24-ai.news/ja/news/2026-05-22/aisi-ai-oversight-erosion-report-2026/ 24 AI ja 2026-05-22T00:00:00Z UK AI Safety Institute：先進AIシステムの監視がますます困難に — 20以上の劣化経路を特定 https://24-ai.news/ja/news/2026-05-22/arxiv-pals-power-aware-llm-serving-moe-energy/ 24 AI ja 2026-05-22T00:00:00Z arXiv:2605.21427：PALS — MoEモデルの電力対応LLMサービングで+26.3%エネルギー効率と4-7倍QoS違反削減を達成 https://24-ai.news/ja/news/2026-05-22/arxiv-persona-vectors-sycophancy-steering-llm-alignment/ 24 AI ja 2026-05-22T00:00:00Z arXiv:2605.21006：既製ペルソナベクトルがLLMモデルの的を絞った追従ステアリングの68-98%の効果を達成 https://24-ai.news/ja/news/2026-05-22/aws-nova-act-hipaa-eligible-healthcare-agents/ 24 AI ja 2026-05-22T00:00:00Z AWS：Nova ActがHIPAA適格を取得 — 医療ワークフロー向けePHIエージェント自動化が解禁 https://24-ai.news/ja/news/2026-05-22/claude-code-workflow-tool-multi-agent-orchestration/ 24 AI ja 2026-05-22T00:00:00Z Anthropic：Claude Code v2.1.147が確定性マルチエージェントオーケストレーションのWorkflowツールを導入 https://24-ai.news/ja/news/2026-05-22/cncf-fluid-llm-cold-start-30s-kubernetes/ 24 AI ja 2026-05-22T00:00:00Z CNCF：NetEase GamesがFluidプリフェッチレイヤーでKubernetes上の30秒LLMコールドスタートを達成 https://24-ai.news/ja/news/2026-05-22/flux-erase-object-removal-scene-reconstruction/ 24 AI ja 2026-05-22T00:00:00Z Black Forest Labs：FLUX Eraseがプロンプトなしの物体削除でGPT Image-2（68.5%）とFinegrain（63.2%）を上回る https://24-ai.news/ja/news/2026-05-22/langchain-agent-streams-typed-channels-langgraph/ 24 AI ja 2026-05-22T00:00:00Z LangChain：トークンストリームからエージェントストリームへ — 型付きチャネルがマルチエージェントUIの古典的ストリーミングに取って代わる https://24-ai.news/ja/news/2026-05-22/microsoft-magenticlite-fara15-small-model-agents/ 24 AI ja 2026-05-22T00:00:00Z Microsoft Research：MagenticLite + Fara1.5（4B/9B/27B）— 小型モデル向けに最適化したエージェントAIがSOTAを達成 https://24-ai.news/ja/news/2026-05-22/microsoft-vega-zk-proofs-digital-identity-ai-agents/ 24 AI ja 2026-05-22T00:00:00Z Microsoft Research：Vega — デジタルアイデンティティ向けZKプルーフ、92ms生成と70%高速な再証明を実現 https://24-ai.news/ja/news/2026-05-22/oecd-collective-ai-security-foundations-g7/ 24 AI ja 2026-05-22T00:00:00Z OECD AI：集合的AIセキュリティにはG7調整が必要 — プロンプトインジェクション、エージェントセキュリティ、モデルポイズニングが優先事項 https://24-ai.news/ja/news/2026-05-22/openai-codex-enterprise-scale-4m-users/ 24 AI ja 2026-05-22T00:00:00Z OpenAI：Codexのエンタープライズ展開 — 週間アクティブユーザー400万人とCodex Labsプログラム https://24-ai.news/ko/news/2026-05-22/aisi-ai-oversight-erosion-report-2026/ 24 AI ko 2026-05-22T00:00:00Z UK AI Safety Institute：고급 AI 시스템 감시가 점점 더 어려워진다 — 20개 이상의 열화 경로 확인 https://24-ai.news/ko/news/2026-05-22/arxiv-pals-power-aware-llm-serving-moe-energy/ 24 AI ko 2026-05-22T00:00:00Z arXiv:2605.21427：PALS — MoE 모델 전력 인식 LLM 서빙으로 +26.3% 에너지 효율 및 4-7배 QoS 위반 감소 달성 https://24-ai.news/ko/news/2026-05-22/arxiv-persona-vectors-sycophancy-steering-llm-alignment/ 24 AI ko 2026-05-22T00:00:00Z arXiv:2605.21006：기성 페르소나 벡터가 LLM 모델 표적 아첨 스티어링의 68-98% 효율 달성 https://24-ai.news/ko/news/2026-05-22/aws-nova-act-hipaa-eligible-healthcare-agents/ 24 AI ko 2026-05-22T00:00:00Z AWS：Nova Act HIPAA 적격 획득 — 의료 워크플로를 위한 ePHI 에이전트 자동화 해금 https://24-ai.news/ko/news/2026-05-22/claude-code-workflow-tool-multi-agent-orchestration/ 24 AI ko 2026-05-22T00:00:00Z Anthropic：Claude Code v2.1.147, 결정론적 다중 에이전트 오케스트레이션을 위한 Workflow 도구 도입 https://24-ai.news/ko/news/2026-05-22/cncf-fluid-llm-cold-start-30s-kubernetes/ 24 AI ko 2026-05-22T00:00:00Z CNCF：NetEase Games, Fluid 프리페칭 레이어로 Kubernetes LLM 콜드 스타트 30초 달성 https://24-ai.news/ko/news/2026-05-22/flux-erase-object-removal-scene-reconstruction/ 24 AI ko 2026-05-22T00:00:00Z Black Forest Labs：FLUX Erase, 프롬프트 없는 객체 제거에서 GPT Image-2（68.5%）와 Finegrain（63.2%）을 능가 https://24-ai.news/ko/news/2026-05-22/langchain-agent-streams-typed-channels-langgraph/ 24 AI ko 2026-05-22T00:00:00Z LangChain：토큰 스트림에서 에이전트 스트림으로 — 타입화 채널이 다중 에이전트 UI의 고전적 스트리밍을 대체 https://24-ai.news/ko/news/2026-05-22/microsoft-magenticlite-fara15-small-model-agents/ 24 AI ko 2026-05-22T00:00:00Z Microsoft Research：MagenticLite + Fara1.5（4B/9B/27B）— 소형 모델에 최적화된 에이전트 AI가 SOTA 달성 https://24-ai.news/ko/news/2026-05-22/microsoft-vega-zk-proofs-digital-identity-ai-agents/ 24 AI ko 2026-05-22T00:00:00Z Microsoft Research：Vega — 디지털 신원을 위한 ZK 증명, 92ms 생성과 70% 빠른 반복 증명 실현 https://24-ai.news/ko/news/2026-05-22/oecd-collective-ai-security-foundations-g7/ 24 AI ko 2026-05-22T00:00:00Z OECD AI：집체 AI 보안에 G7 조율 필요 — 프롬프트 인젝션·에이전트 보안·모델 오염이 우선 과제 https://24-ai.news/ko/news/2026-05-22/openai-codex-enterprise-scale-4m-users/ 24 AI ko 2026-05-22T00:00:00Z OpenAI：Codex 엔터프라이즈 확장 — 주간 활성 사용자 400만 명과 Codex Labs 프로그램 https://24-ai.news/hr/news/2026-05-21/amd-rocm-7-13-mi350p-virtualization-therock/ 24 AI hr 2026-05-21T00:00:00Z AMD: ROCm 7.13 donosi MI350P GPU, multi-VF virtualizaciju i TheRock pakiranje https://24-ai.news/hr/news/2026-05-21/anthropic-mcp-tunnels-self-hosted-sandboxes/ 24 AI hr 2026-05-21T00:00:00Z Anthropic: MCP Tunnels i self-hosted sandboxes za Claude Managed Agents https://24-ai.news/hr/news/2026-05-21/arxiv-icml-structured-reasoning-math-beyond-code/ 24 AI hr 2026-05-21T00:00:00Z arXiv:2605.19762: ICML 2026 paper tvrdi da kod ne poboljšava matematičko zaključivanje LLM-ova https://24-ai.news/hr/news/2026-05-21/aws-sagemaker-openai-compatible-api-endpoints/ 24 AI hr 2026-05-21T00:00:00Z AWS: SageMaker AI dobiva OpenAI-kompatibilne API endpointe https://24-ai.news/hr/news/2026-05-21/deepmind-co-scientist-multi-agent-research-partner/ 24 AI hr 2026-05-21T00:00:00Z Google DeepMind: Co-Scientist multi-agent AI partner za znanstveno istraživanje https://24-ai.news/hr/news/2026-05-21/deepmind-weathernext-hurricane-melissa-cat5/ 24 AI hr 2026-05-21T00:00:00Z Google DeepMind: WeatherNext predvidio uragan Melissa kao Kategorija 5 pet dana unaprijed https://24-ai.news/hr/news/2026-05-21/eu-ai-act-high-risk-classification-guidelines/ 24 AI hr 2026-05-21T00:00:00Z EU AI Office: nacrt smjernica za klasifikaciju visokorizičnih AI sustava https://24-ai.news/hr/news/2026-05-21/github-vscode-extension-internal-repo-breach/ 24 AI hr 2026-05-21T00:00:00Z GitHub: trovana VS Code ekstenzija probila ~3.800 internih repozitorija https://24-ai.news/hr/news/2026-05-21/google-io-2026-antigravity-2-spark-platform/ 24 AI hr 2026-05-21T00:00:00Z Google: I/O 2026 round 2 — Antigravity 2.0, Gemini Spark i Universal Cart https://24-ai.news/hr/news/2026-05-21/langchain-deep-agents-interpreter-quickjs-runtime/ 24 AI hr 2026-05-21T00:00:00Z LangChain: Deep Agents dobivaju QuickJS interpreters za kod između tool poziva https://24-ai.news/hr/news/2026-05-21/openai-model-disproves-unit-distance-conjecture/ 24 AI hr 2026-05-21T00:00:00Z OpenAI: AI model opovrgnuo 80-godišnju konjekturu u diskretnoj geometriji https://24-ai.news/hr/news/2026-05-21/stability-stable-audio-3-open-weight-six-min/ 24 AI hr 2026-05-21T00:00:00Z Stability AI: Stable Audio 3.0 s open-weight modelima i generacijom do 6 minuta https://24-ai.news/en/news/2026-05-21/amd-rocm-7-13-mi350p-virtualization-therock/ 24 AI en 2026-05-21T00:00:00Z AMD: ROCm 7.13 brings MI350P GPU, multi-VF virtualisation and TheRock packaging https://24-ai.news/en/news/2026-05-21/anthropic-mcp-tunnels-self-hosted-sandboxes/ 24 AI en 2026-05-21T00:00:00Z Anthropic: MCP Tunnels and self-hosted sandboxes for Claude Managed Agents https://24-ai.news/en/news/2026-05-21/arxiv-icml-structured-reasoning-math-beyond-code/ 24 AI en 2026-05-21T00:00:00Z arXiv:2605.19762: ICML 2026 paper claims code does not improve LLM mathematical reasoning https://24-ai.news/en/news/2026-05-21/aws-sagemaker-openai-compatible-api-endpoints/ 24 AI en 2026-05-21T00:00:00Z AWS: SageMaker AI gains OpenAI-compatible API endpoints https://24-ai.news/en/news/2026-05-21/deepmind-co-scientist-multi-agent-research-partner/ 24 AI en 2026-05-21T00:00:00Z Google DeepMind: Co-Scientist multi-agent AI partner for scientific research https://24-ai.news/en/news/2026-05-21/deepmind-weathernext-hurricane-melissa-cat5/ 24 AI en 2026-05-21T00:00:00Z Google DeepMind: WeatherNext predicted Hurricane Melissa as Category 5 five days ahead https://24-ai.news/en/news/2026-05-21/eu-ai-act-high-risk-classification-guidelines/ 24 AI en 2026-05-21T00:00:00Z EU AI Office: draft guidelines for classifying high-risk AI systems https://24-ai.news/en/news/2026-05-21/github-vscode-extension-internal-repo-breach/ 24 AI en 2026-05-21T00:00:00Z GitHub: malicious VS Code extension breached ~3,800 internal repositories https://24-ai.news/en/news/2026-05-21/google-io-2026-antigravity-2-spark-platform/ 24 AI en 2026-05-21T00:00:00Z Google: I/O 2026 round 2 — Antigravity 2.0, Gemini Spark and Universal Cart https://24-ai.news/en/news/2026-05-21/langchain-deep-agents-interpreter-quickjs-runtime/ 24 AI en 2026-05-21T00:00:00Z LangChain: Deep Agents get QuickJS interpreters for code between tool calls https://24-ai.news/en/news/2026-05-21/openai-model-disproves-unit-distance-conjecture/ 24 AI en 2026-05-21T00:00:00Z OpenAI: AI model disproves 80-year-old conjecture in discrete geometry https://24-ai.news/en/news/2026-05-21/stability-stable-audio-3-open-weight-six-min/ 24 AI en 2026-05-21T00:00:00Z Stability AI: Stable Audio 3.0 with open-weight models and generation up to 6 minutes https://24-ai.news/de/news/2026-05-21/amd-rocm-7-13-mi350p-virtualization-therock/ 24 AI de 2026-05-21T00:00:00Z AMD: ROCm 7.13 bringt MI350P-GPU, Multi-VF-Virtualisierung und TheRock-Paketierung https://24-ai.news/de/news/2026-05-21/anthropic-mcp-tunnels-self-hosted-sandboxes/ 24 AI de 2026-05-21T00:00:00Z Anthropic: MCP Tunnels und Self-hosted Sandboxes für Claude Managed Agents https://24-ai.news/de/news/2026-05-21/arxiv-icml-structured-reasoning-math-beyond-code/ 24 AI de 2026-05-21T00:00:00Z arXiv:2605.19762: ICML-2026-Paper behauptet, Code verbessert nicht das Mathe-Reasoning von LLMs https://24-ai.news/de/news/2026-05-21/aws-sagemaker-openai-compatible-api-endpoints/ 24 AI de 2026-05-21T00:00:00Z AWS: SageMaker AI erhält OpenAI-kompatible API-Endpunkte https://24-ai.news/de/news/2026-05-21/deepmind-co-scientist-multi-agent-research-partner/ 24 AI de 2026-05-21T00:00:00Z Google DeepMind: Co-Scientist Multi-Agent-KI-Partner für die Wissenschaft https://24-ai.news/de/news/2026-05-21/deepmind-weathernext-hurricane-melissa-cat5/ 24 AI de 2026-05-21T00:00:00Z Google DeepMind: WeatherNext sagte Hurrikan Melissa als Kategorie 5 fünf Tage voraus https://24-ai.news/de/news/2026-05-21/eu-ai-act-high-risk-classification-guidelines/ 24 AI de 2026-05-21T00:00:00Z EU AI Office: Entwurf von Leitlinien zur Hochrisiko-KI-Klassifizierung https://24-ai.news/de/news/2026-05-21/github-vscode-extension-internal-repo-breach/ 24 AI de 2026-05-21T00:00:00Z GitHub: Schädliche VS-Code-Erweiterung kompromittierte ~3.800 interne Repos https://24-ai.news/de/news/2026-05-21/google-io-2026-antigravity-2-spark-platform/ 24 AI de 2026-05-21T00:00:00Z Google: I/O 2026 Runde 2 — Antigravity 2.0, Gemini Spark und Universal Cart https://24-ai.news/de/news/2026-05-21/langchain-deep-agents-interpreter-quickjs-runtime/ 24 AI de 2026-05-21T00:00:00Z LangChain: Deep Agents erhalten QuickJS-Interpreter für Code zwischen Tool-Aufrufen https://24-ai.news/de/news/2026-05-21/openai-model-disproves-unit-distance-conjecture/ 24 AI de 2026-05-21T00:00:00Z OpenAI: KI-Modell widerlegt 80 Jahre alte Vermutung in diskreter Geometrie https://24-ai.news/de/news/2026-05-21/stability-stable-audio-3-open-weight-six-min/ 24 AI de 2026-05-21T00:00:00Z Stability AI: Stable Audio 3.0 mit Open-Weight-Modellen und 6-Minuten-Generierung https://24-ai.news/zh/news/2026-05-21/amd-rocm-7-13-mi350p-virtualization-therock/ 24 AI zh 2026-05-21T00:00:00Z AMD：ROCm 7.13带来MI350P GPU、多VF虚拟化与TheRock模块化打包 https://24-ai.news/zh/news/2026-05-21/anthropic-mcp-tunnels-self-hosted-sandboxes/ 24 AI zh 2026-05-21T00:00:00Z Anthropic：Claude Managed Agents推出MCP Tunnels与自托管沙盒 https://24-ai.news/zh/news/2026-05-21/arxiv-icml-structured-reasoning-math-beyond-code/ 24 AI zh 2026-05-21T00:00:00Z arXiv:2605.19762: ICML 2026论文声称代码并不能提升LLM的数学推理能力 https://24-ai.news/zh/news/2026-05-21/aws-sagemaker-openai-compatible-api-endpoints/ 24 AI zh 2026-05-21T00:00:00Z AWS：SageMaker AI新增OpenAI兼容API端点 https://24-ai.news/zh/news/2026-05-21/deepmind-co-scientist-multi-agent-research-partner/ 24 AI zh 2026-05-21T00:00:00Z Google DeepMind：Co-Scientist——面向科学研究的多代理AI合作伙伴 https://24-ai.news/zh/news/2026-05-21/deepmind-weathernext-hurricane-melissa-cat5/ 24 AI zh 2026-05-21T00:00:00Z Google DeepMind：WeatherNext提前5天预测梅利莎飓风达到5级 https://24-ai.news/zh/news/2026-05-21/eu-ai-act-high-risk-classification-guidelines/ 24 AI zh 2026-05-21T00:00:00Z EU AI办公室：高风险AI系统分类指南草案公开征询意见 https://24-ai.news/zh/news/2026-05-21/github-vscode-extension-internal-repo-breach/ 24 AI zh 2026-05-21T00:00:00Z GitHub：恶意VS Code扩展入侵约3800个内部代码仓库 https://24-ai.news/zh/news/2026-05-21/google-io-2026-antigravity-2-spark-platform/ 24 AI zh 2026-05-21T00:00:00Z Google：I/O 2026第二波——Antigravity 2.0、Gemini Spark与Universal Cart https://24-ai.news/zh/news/2026-05-21/langchain-deep-agents-interpreter-quickjs-runtime/ 24 AI zh 2026-05-21T00:00:00Z LangChain：Deep Agents引入QuickJS解释器，可在工具调用之间执行代码 https://24-ai.news/zh/news/2026-05-21/openai-model-disproves-unit-distance-conjecture/ 24 AI zh 2026-05-21T00:00:00Z OpenAI：AI模型推翻了离散几何学中存在80年的猜想 https://24-ai.news/zh/news/2026-05-21/stability-stable-audio-3-open-weight-six-min/ 24 AI zh 2026-05-21T00:00:00Z Stability AI：Stable Audio 3.0发布开放权重模型，支持最长6分钟生成 https://24-ai.news/ja/news/2026-05-21/amd-rocm-7-13-mi350p-virtualization-therock/ 24 AI ja 2026-05-21T00:00:00Z AMD：ROCm 7.13がMI350P GPU、マルチVF仮想化、TheRockモジュラーパッケージを導入 https://24-ai.news/ja/news/2026-05-21/anthropic-mcp-tunnels-self-hosted-sandboxes/ 24 AI ja 2026-05-21T00:00:00Z Anthropic：Claude Managed AgentsにMCP Tunnelsとセルフホストサンドボックスが登場 https://24-ai.news/ja/news/2026-05-21/arxiv-icml-structured-reasoning-math-beyond-code/ 24 AI ja 2026-05-21T00:00:00Z arXiv:2605.19762: ICML 2026論文——コードはLLMの数学的推論を改善しないと主張 https://24-ai.news/ja/news/2026-05-21/aws-sagemaker-openai-compatible-api-endpoints/ 24 AI ja 2026-05-21T00:00:00Z AWS：SageMaker AIにOpenAI互換APIエンドポイントが登場 https://24-ai.news/ja/news/2026-05-21/deepmind-co-scientist-multi-agent-research-partner/ 24 AI ja 2026-05-21T00:00:00Z Google DeepMind：Co-Scientist——科学研究を加速するマルチエージェントAIパートナー https://24-ai.news/ja/news/2026-05-21/deepmind-weathernext-hurricane-melissa-cat5/ 24 AI ja 2026-05-21T00:00:00Z Google DeepMind：WeatherNextがハリケーン「メリッサ」のカテゴリー5到達を5日前に予測 https://24-ai.news/ja/news/2026-05-21/eu-ai-act-high-risk-classification-guidelines/ 24 AI ja 2026-05-21T00:00:00Z EU AI Office：高リスクAIシステム分類ガイドライン草案のパブリックコンサルテーション開始 https://24-ai.news/ja/news/2026-05-21/github-vscode-extension-internal-repo-breach/ 24 AI ja 2026-05-21T00:00:00Z GitHub：悪意あるVS Code拡張機能により内部リポジトリ約3,800件が侵害される https://24-ai.news/ja/news/2026-05-21/google-io-2026-antigravity-2-spark-platform/ 24 AI ja 2026-05-21T00:00:00Z Google：I/O 2026 第2弾——Antigravity 2.0、Gemini Spark、Universal Cart発表 https://24-ai.news/ja/news/2026-05-21/langchain-deep-agents-interpreter-quickjs-runtime/ 24 AI ja 2026-05-21T00:00:00Z LangChain：Deep AgentsにツールコールをまたぐコードのためのQuickJSインタープリターが登場 https://24-ai.news/ja/news/2026-05-21/openai-model-disproves-unit-distance-conjecture/ 24 AI ja 2026-05-21T00:00:00Z OpenAI：AIモデルが80年来の離散幾何学の予想を反証 https://24-ai.news/ja/news/2026-05-21/stability-stable-audio-3-open-weight-six-min/ 24 AI ja 2026-05-21T00:00:00Z Stability AI：Stable Audio 3.0がオープンウェイトモデルと最大6分の生成機能を引っさげて登場 https://24-ai.news/ko/news/2026-05-21/amd-rocm-7-13-mi350p-virtualization-therock/ 24 AI ko 2026-05-21T00:00:00Z AMD: ROCm 7.13, MI350P GPU·멀티 VF 가상화·TheRock 모듈형 패키징 도입 https://24-ai.news/ko/news/2026-05-21/anthropic-mcp-tunnels-self-hosted-sandboxes/ 24 AI ko 2026-05-21T00:00:00Z Anthropic: Claude Managed Agents에 MCP Tunnels와 셀프 호스팅 샌드박스 도입 https://24-ai.news/ko/news/2026-05-21/arxiv-icml-structured-reasoning-math-beyond-code/ 24 AI ko 2026-05-21T00:00:00Z arXiv:2605.19762: ICML 2026 논문——코드가 LLM의 수학적 추론을 향상시키지 않는다고 주장 https://24-ai.news/ko/news/2026-05-21/aws-sagemaker-openai-compatible-api-endpoints/ 24 AI ko 2026-05-21T00:00:00Z AWS: SageMaker AI에 OpenAI 호환 API 엔드포인트 도입 https://24-ai.news/ko/news/2026-05-21/deepmind-co-scientist-multi-agent-research-partner/ 24 AI ko 2026-05-21T00:00:00Z Google DeepMind: Co-Scientist——과학 연구를 위한 멀티 에이전트 AI 파트너 https://24-ai.news/ko/news/2026-05-21/deepmind-weathernext-hurricane-melissa-cat5/ 24 AI ko 2026-05-21T00:00:00Z Google DeepMind: WeatherNext, 허리케인 멜리사의 5등급 도달을 5일 전에 예측 https://24-ai.news/ko/news/2026-05-21/eu-ai-act-high-risk-classification-guidelines/ 24 AI ko 2026-05-21T00:00:00Z EU AI 사무국: 고위험 AI 시스템 분류 가이드라인 초안 공개 의견 수렴 시작 https://24-ai.news/ko/news/2026-05-21/github-vscode-extension-internal-repo-breach/ 24 AI ko 2026-05-21T00:00:00Z GitHub: 악성 VS Code 확장이 내부 저장소 약 3,800개를 침해하다 https://24-ai.news/ko/news/2026-05-21/google-io-2026-antigravity-2-spark-platform/ 24 AI ko 2026-05-21T00:00:00Z Google: I/O 2026 2차 발표——Antigravity 2.0, Gemini Spark, Universal Cart https://24-ai.news/ko/news/2026-05-21/langchain-deep-agents-interpreter-quickjs-runtime/ 24 AI ko 2026-05-21T00:00:00Z LangChain: Deep Agents에 도구 호출 사이 코드 실행을 위한 QuickJS 인터프리터 도입 https://24-ai.news/ko/news/2026-05-21/openai-model-disproves-unit-distance-conjecture/ 24 AI ko 2026-05-21T00:00:00Z OpenAI: AI 모델이 이산기하학의 80년 된 추측을 반증하다 https://24-ai.news/ko/news/2026-05-21/stability-stable-audio-3-open-weight-six-min/ 24 AI ko 2026-05-21T00:00:00Z Stability AI: Stable Audio 3.0, 개방 가중치 모델과 최대 6분 생성 지원 출시