Today's briefing
- Physical AI
Nvidia releases Flora, reducing robot training data costs by up to 125 times
Physical AIKorean outletAI TimesOct 23, 2026 - LLM
Mistral AI launches Voxtral 3 for real-time voice agents
LLMKorean outletAI TimesOct 23, 2026 - Open Source & Tools
Anthropic launches Claude Security for enterprise code vulnerability analysis
Open Source & ToolsKorean outletAI TimesOct 23, 2026
Korea Desk
What Korea’s tech press is reporting, summarized in English.
6 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Agents
ScholaEvolve framework enables agents to self-improve by reading academic papers
Researchers from UC Santa Barbara and Microsoft introduced ScholaEvolve, a framework that allows AI agents to identify problems during task execution, search for relevant academic papers, implement verified solutions, and validate improvements. The approach shifts from trial-and-error improvement to research-driven optimization.
AI TimesOct 23, 2026 - Open Source & Tools
Critical remote code execution vulnerability discovered in PHP
A critical security vulnerability was found in PHP, the widely-used web development language, that allows malicious code execution and unauthorized server access without authentication. The flaw affects broad infrastructure, prompting some European governments to preemptively shut down public web services.
Byline Network +2 moreOct 22, 2026
Open Source & Tools
5 stories · 2 on the front page · 1 at the Korea DeskAll stories in this section →Stories in this issue
27 storiesGoogle's Agent Compute Engine isolates each Gemini agent in its own micro virtual machine
Google unveiled Agent Compute Engine, a runtime environment designed for AI agents that assigns each Gemini agent a hardware-isolated micro virtual machine. It expands governance capabilities for managing agents across cloud, hybrid, and on-premises environments.
Nvidia launches Minerva AIOps platform for AI factory observability
Nvidia launched Minerva, an AIOps platform that combines telemetry from infrastructure and AI workloads and applies purpose-built AI models to deliver actionable insights. It helps teams identify problems early and keep AI factories productive.
OpenAI makes its decision model widely available
OpenAI launched its decision model to a broader audience as competition in decision-making AI technology intensifies.
Recursive releases Flint, an on-device coding model built with AI automation
AI startup Recursive open-sourced Flint, an on-device coding model built in under four months using AI automation systems. With 40 billion parameters but using a mixture-of-experts structure to activate 2.5 billion parameters per token, it runs on personal computers while preserving data security.
Nvidia releases Flora, reducing robot training data costs by up to 125 times
Nvidia open-sourced Flora, a model that automatically analyzes human and robot task videos to record actions, timing, success, and failure causes. It reduces high-quality robot training data costs by up to 125 times and accelerates robot foundation model development and physical AI research.
Mistral AI launches Voxtral 3 for real-time voice agents
Mistral AI released Voxtral 3, a new voice AI product family combining speech-to-text and text-to-speech models with an SDK. It supports multilingual voice recognition and natural speech generation for real-time voice agent development, customer service automation, and conversational services.
Anthropic launches Claude Security for enterprise code vulnerability analysis
Anthropic released Claude Security in public beta, a tool that automates code vulnerability detection and remediation tracking. The tool continuously analyzes code changes to identify vulnerabilities and adapts risk prioritization to match enterprise security policies and system environments.
OpenAI agents analyze mouse brain data in 32 minutes
OpenAI teamed up with PerturbAI and Scripps Research to deploy thousands of AI agents that analyzed biological data from millions of mouse cells. The agents completed a task in 32 minutes that would take humans months.
Anthropic launches Claude Duet, an always-on autonomous agent
Anthropic released Claude Duet, a proactive AI agent built on Claude Opus 5.5 that handles tasks independently, even when offline. It connects to over 6,000 services including Gmail, Slack, and Stripe, and requests approval before sensitive actions like payments.
Nvidia expands Cosmos platform with tools to train and deploy world models
Nvidia expanded the Cosmos platform with new training tools, simulation checkpoints, and Cosmos Data Factory blueprints to help developers build world models for physical AI systems. Cosmos models rank in the top five on six different open-model leaderboards.
Formic raises $50 million to scale physical AI in American manufacturing
Formic secured $50 million in funding to expand physical AI deployment across American manufacturing facilities.
1Password enables AI agents to log in without exposing passwords
1Password released a new feature that allows AI agents to log into customer accounts while protecting the underlying credentials. Agents can access accounts without viewing or handling passwords.
Canva releases its first image generation model
Canva made its newly developed image generation model available to all users. The company promises better, more customizable performance at lower costs.
Microsoft AI diagnostic system outperforms doctors on hard cases
Microsoft's new AI diagnostic system beat generalist doctors on difficult cases, according to a study published in Science. Much of the performance improvement derives from the underlying AI models rather than Microsoft's system itself.
Former OpenAI scientist launches Warp Science to automate wet labs with robots
Warp Science trains robots to perform laboratory tasks using human chemists wearing cameras and sensor gloves. The startup aims to automate wet lab operations through robot learning.
ScholaEvolve framework enables agents to self-improve by reading academic papers
Researchers from UC Santa Barbara and Microsoft introduced ScholaEvolve, a framework that allows AI agents to identify problems during task execution, search for relevant academic papers, implement verified solutions, and validate improvements. The approach shifts from trial-and-error improvement to research-driven optimization.
Chinese hackers spied on thousands of companies after breaching ConnectWise
Microsoft attributed the ConnectWise breach to a China-aligned hacker group that used AI agents to rapidly compromise thousands of the company's customers. The attackers leveraged automated tools to accelerate intrusions at scale.
Cohere releases Rerank 5 with passage-level context focus
Cohere released Rerank 5 in Pro and Fast tiers, featuring a 32K context window, passage-level context extraction, and calibrated relevance scores. The Pro version scores 82.2 on the BRIGHT benchmark and costs $0.05 per 1 million tokens.
Dream Labs raises $90 million and launches Dream API
Dream Labs raised $90 million and launched Dream API, powered by its Dream-1 AI model. The model enables robots from different manufacturers to work together and transfer learning across platforms.
IBM releases Granite Speech 5.0 2B for multilingual transcription and translation
IBM released Granite Speech 5.0 2B, an Apache 2.0-licensed speech recognition and translation model. It ranks second on the Open ASR Leaderboard with 5.68 percent average English word error rate, alongside updated Turbo CTC and 4.1 2B Plus variants.
Critical remote code execution vulnerability discovered in PHP
A critical security vulnerability was found in PHP, the widely-used web development language, that allows malicious code execution and unauthorized server access without authentication. The flaw affects broad infrastructure, prompting some European governments to preemptively shut down public web services.
OpenAI launches Decisions API for fast, low-cost decision-making
OpenAI released the Decisions API, which provides fast, focused, and low-cost decision-making capabilities for any application.
NVIDIA open-sources OpenShell for safe autonomous AI agent fleets
NVIDIA released OpenShell, an open-source runtime environment designed to operate fleets of autonomous AI agents safely and privately on GitHub.
Only 13 percent of companies see real returns from AI, Gartner survey finds
A Gartner survey of 2,002 CIOs found that just 13 percent of companies are realizing actual business value from their AI investments. At the same time, 51 percent expect AI to increase operating costs, yet 75 percent of CEOs plan to continue raising technology budgets.
OpenAI, Anthropic, and Nvidia report record lobbying spending
Nvidia, Anthropic, and OpenAI each set records for quarterly lobbying spending in the July-September period, at $2.2 million, $2.1 million, and $1.5 million respectively, according to federal disclosures. The increase reflects the three firms' focus on shaping AI policy.
AI agent hijacks unprotected language models to conduct autonomous cyberattacks
Symantec security researchers discovered an offensive AI agent that hijacks exposed language models on the web to autonomously attack servers, steal databases, and install cryptominers. The agent can run up to 100,000 steps without human intervention, leaving the hijacked model's owner liable for the costs.
Perplexity and NVIDIA introduce JOLT, a joint reinforcement learning method
Perplexity and NVIDIA researchers introduced JOLT, a post-training method that uses one model as both teacher and student. On the MATH benchmark, JOLT matched GRPO's accuracy while processing 6.5 times fewer training tokens.