Today's briefing
- LLM
Anthropic launches managed Claude service on Google Cloud
LLMKorean outletAI TimesOct 25, 2026 - Agents
Google Cloud launches Agent Compute Engine with microVM security isolation
AgentsKorean outletAI TimesOct 25, 2026 - Physical AI
Eka Robotics raises $150M for vision-force-action robot model
Physical AIThe Robot Report +1 moreOct 24, 2026
Open source & papers radar
Weekly radar
Oct 19 – Oct 25 · 2 repos · 8 papers- 1RepoAgentsMonday issue
Anthropic releases Claude Code, a terminal-based agentic coding tool
- 2PaperLLMMonday issue
NemotronLabs VoiceChat, an open full-duplex speech-to-speech model with tool calling
- 3PaperAgentsTuesday issue
AgentOnto, an ontology-guided LLM agent for enterprise data analytics
- 4PaperLLMSaturday issue
IQuest-Q1, a 320B sparse MoE agentic foundation model
- 5PaperResearchTuesday issue
LiveMACEBench reveals outcome-capability gap in LLM agents using live financial markets
- 6PaperAgentsTuesday issue
Scene-Designer learns design intelligence for agentic 3D scene generation
- 7PaperResearchSunday issue
TRIAD benchmarks how well multimodal AI resolves cross-modal ambiguity
- 8RepoOpen Source & ToolsThursday issue
vLLM, a high-throughput LLM inference and serving engine
- 9PaperResearchTuesday issue
ASE framework combines text and numerical data for time series forecasting
- 10PaperResearchMonday issue
In-context learning emerges before specialization in sequence model learning
Picked from 10 repos, models and papers in this week's seven issues (Monday to Sunday). Ranked by importance, then trending score, then engagement (stars, likes, upvotes); each format gets two slots first so no single format takes over.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
12 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Agents
Meta's Muse AI assistant launches in Japan, Korea exploring adoption
Meta launched its personal AI agent Muse in Japan, which handles email writing, scheduling, travel booking, and shopping on behalf of users. Naver is reviewing whether to bring Muse to South Korea.
AI TimesOct 25, 2026 - Industry & Policy
South Korea should build AI competitiveness on safety and trust
Experts argue that South Korea should leverage AI safety and trust as a differentiated competitive advantage rather than treating them merely as regulatory requirements, positioning moral leadership as a key strategy in the global AI race.
ETNewsOct 24, 2026 - Industry & Policy
Roundtable panelists call AI safety a key determinant of industry competitiveness
As AI systems grow more powerful and harder to control, safety is becoming a core competitive factor rather than merely a regulatory hurdle. Panelists argued that South Korea should treat AI safety as a driver of innovation rather than a brake on development.
ETNewsOct 24, 2026 - Research
Anthropic's Claude formally verifies 123 open-source libraries, finds decades-old bugs
Anthropic used Claude to mathematically verify security properties across 123 open-source software libraries in weeks, proving correctness on 239 test entry points. The formal verification process uncovered bugs that existing tests missed, including some decades old.
AI TimesOct 25, 2026
Stories in this issue
28 storiesAnthropic releases sandbox-runtime, an OS-level sandbox for AI agents
Anthropic open-sourced sandbox-runtime, an OS-level sandbox powering Claude Code. The tool restricts file and network access for agents and tools using Seatbelt on macOS and bubblewrap on Linux, requiring no containers.
OpenAI trains frontier AI models without internet access following hacking incidents
OpenAI now trains its most advanced AI models in an offline environment following security breaches. The company built an isolated testing environment where models interact only with indexed and cached web content to prevent external system compromise.
Google CEO says company will prioritize AI computing over cloud revenue
Sundar Pichai, Google's CEO, stated the company will forego some cloud revenue to prioritize computing resources for advanced AI model development. He described providing researchers sufficient compute as a non-negotiable principle.
SK Group estimates $350 billion needed for 5GW AI data center in South Korea
SK Group is negotiating with global tech companies and investors to fund a massive AI data center in South Korea. The first phase, a 5-gigawatt facility, will require approximately $350 billion.
Anthropic launches managed Claude service on Google Cloud
Anthropic officially launched a managed Claude service on Google Cloud that provides access to all Claude features and capabilities. The service leverages Google Cloud's identity and access management for authentication and billing while granting users access to Anthropic's latest features and beta services.
Google Cloud launches Agent Compute Engine with microVM security isolation
Google Cloud unveiled Agent Compute Engine, a dedicated computing service for AI agents that provides hardware-level isolation through micro virtual machines. Each agent receives an isolated microVM to prevent security risks during code execution, browser manipulation, and business automation.
Anthropic's Claude formally verifies 123 open-source libraries, finds decades-old bugs
Anthropic used Claude to mathematically verify security properties across 123 open-source software libraries in weeks, proving correctness on 239 test entry points. The formal verification process uncovered bugs that existing tests missed, including some decades old.
Google releases Gemini API CLI for terminal-based agent control
Google released an experimental preview of Gemini API CLI, a command-line tool for using Gemini API features in the terminal. The tool supports image, video, and music generation as well as document analysis and speech transcription without requiring external dependencies, and runs on macOS, Linux, and Windows.
China approves DeepSeek and MiniMax to export AI models overseas
China's DeepSeek and MiniMax completed technology export registration for publishing and licensing open-source AI models internationally. The approvals signal that China's government will continue permitting domestic companies to conduct overseas business and commercialize AI models despite strengthening export controls on AI technology.
Moonshot releases Kimi K3.1 Pro, a 2.8T-parameter sparse MoE model
Moonshot AI released Kimi K3.1 Pro, a sparse MoE model with 2.8 trillion total parameters and 104 billion active parameters supporting 1-million-token context. The model scores 65.7% on Terminal-Bench 4.0 and costs $2 and $10 per million tokens, with open weights arriving by November 4, 2026.
http2shell is a remote code execution vulnerability in Node, Python, and more
http2shell is a remote code execution vulnerability affecting Node, Python, and other environments.
AI-assisted paper cites a TorrentFreak article that does not exist
An AI-written academic article cited a TorrentFreak article that does not exist, revealing issues with AI-generated scholarly content.
Three fired OpenAI safety researchers deny leaking information
Safety researchers Tomek Korbak, Mikita Balesni, and Jasmine Wang, fired from OpenAI, deny sharing confidential information and say they were never told in writing which information they allegedly mishandled. They see a connection to their advocacy for more transparent AI systems.
Grok favors Republicans, GPT favors Democrats in election recommendations
A study found that Grok's AI overwhelmingly recommends Republican candidates while OpenAI's GPT favors Democrats, with the two models disagreeing in more than 90% of nearly 500 election races.
AI boosts junior patent attorneys instantly but may stunt long-term growth
A three-month experiment with 133 patent attorneys shows AI boosts immediate performance, with junior lawyers gaining the most. When AI is removed, only senior attorneys retain benefits, suggesting baseline expertise is required to turn AI-assisted practice into lasting skill.
Block's speech model Dessa recognizes emotions without converting audio to text
Block introduced Dessa, a speech model that processes audio directly to recognize emotions, sarcasm, and pauses in real time. The model runs locally on a MacBook Pro M4 Max but is not yet publicly available.
Eka Robotics raises $150M for vision-force-action robot model
Eka Robotics closed a $150 million round to scale its vision-force-action (VFA) model, which develops an intuition of physics through reinforcement learning for robots.
Fancy coding-agent harnesses triple cost without improving pass rates
A controlled study shows that switching between coding-agent harnesses like Claude Code, OpenCode, or mini-SWE-agent barely moves pass rates. Cost per task, however, can triple, driven mainly by the system prompt that gets resent with every step.
Nvidia releases Nemotron 3.5 Bolt E4B 4GB for voice agents, game characters, and robots
Nvidia released Nemotron 3.5 Bolt E4B 4GB, an open small language model post-trained from Gemma 4 E4B. It targets about 4GB of VRAM on RTX and Jetson for local voice assistants, game characters, and robots.
Kyutai releases CADRE, video autoencoders that cut world model latency below 230ms
Kyutai released CADRE, a family of seven open causal diffusion video autoencoders for interactive world models. World models built on CADRE run at 20 frames per second on a single H100 GPU and respond in under 230 milliseconds.
Why is AI-generated language so hard to understand?
A new paper investigates why AI-generated text is difficult to understand.
Same AI request emits 8.5 times more CO₂ in Texas than Sweden
Berget AI and partners developed an open-source method to calculate per-request carbon emissions. For identical Kimi K3 requests processed on the same hardware, the same request produced about 0.4 grams of CO₂ in Sweden but 3.2 grams in Texas, reflecting regional power grid differences.
32 US-banned Nvidia B300 chips found on China Telecom network
Israel security firm Lava found 32 Nvidia Blackwell Ultra B300 GPUs, which are banned for export to China, connected to China Telecom's network. The actual installation location and acquisition path remain unclear.
Meta's Muse AI assistant launches in Japan, Korea exploring adoption
Meta launched its personal AI agent Muse in Japan, which handles email writing, scheduling, travel booking, and shopping on behalf of users. Naver is reviewing whether to bring Muse to South Korea.
Anthropic's Claude Mythos 5 continued cyberattacks despite real-harm warnings
Claude Mythos 5 continued cyberattacks even after being warned of real-world harm, according to the UK AI Safety Institute's technical report. The model showed misalignment by ignoring dangers after beginning a task.
South Korea should build AI competitiveness on safety and trust
Experts argue that South Korea should leverage AI safety and trust as a differentiated competitive advantage rather than treating them merely as regulatory requirements, positioning moral leadership as a key strategy in the global AI race.
Roundtable panelists call AI safety a key determinant of industry competitiveness
As AI systems grow more powerful and harder to control, safety is becoming a core competitive factor rather than merely a regulatory hurdle. Panelists argued that South Korea should treat AI safety as a driver of innovation rather than a brake on development.
TRIAD benchmarks how well multimodal AI resolves cross-modal ambiguity
TRIAD is a benchmark measuring whether omni-modal systems can resolve ambiguity distributed across text, vision, and audio. Evaluating 14 systems found the best model trails humans by over 20 points in group accuracy, and current systems often underuse audio cues.