Today's briefing
- LLM
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights image model
LLMThe Decoder +1 moreOct 16, 2026 - Agents
NVIDIA releases Nemotron 3.5 models for specialized agents and edge deployment
AgentsNVIDIA Blog +1 moreOct 16, 2026 - Physical AI
Gemini 4 Argon controls robots and generates simulations from images
Physical AIThe DecoderOct 16, 2026
Open source & papers radar
- PaperResearcharXiv 2610.19128Upvotes: 24
CLASHarness benchmarks multi-agent conflict resolution in computer-use systems
CLASHarness introduces a controllable framework for generating verifiable multi-agent conflicts in computer-use tasks, with a benchmark of 600 multi-round tasks. Evaluating 15 frontier models revealed that agents struggle to resolve conflicts when sub-agent outputs contradict the environment, and coordination protocols remain a universal bottleneck.
- PaperResearcharXiv 2610.13623Upvotes: 11
APEX-Harm reveals high vulnerability of frontier AI agents to script injection attacks
APEX-Harm benchmarks prompt and script injection attacks against agents performing finance, consulting, and legal tasks in large shared workspaces with 115 to 328 files per task. Evaluating 12 frontier models showed that task-specific script replacement and attacker-provided generic script attacks achieve 81–84% and 73% success rates respectively.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
10 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Agents
Kakao builds Modu's AI, an open platform connecting external agents and tools
Kakao is developing Modu's AI, an open AI platform on KakaoTalk where 34 companies and institutions will provide specialized agents and tools. Kakao plans to launch the platform by year-end with a target of 5 million monthly active users, and will gradually introduce personalized one-to-one agents starting in 2027.
Byline NetworkOct 15, 2026 - Industry & Policy
Apple introduces HomeView with Siri AI; partners with LG Electronics
Apple unveiled HomeView, a 7-inch display device with Siri AI priced at $299, launching November 13. The company also announced Apple Home Collection, a bundle of cameras, doorbells, door locks, and thermostats designed with LG Electronics and Schneider Electric, available December.
Byline NetworkOct 14, 2026 - LLM
Google officially launches Gemini 4 Argon
Google officially launched Gemini 4 Argon, a frontier model it had released with limited access two weeks earlier. Google said the model outperforms competing models including GPT-6 Astra and Claude Opus 5.5.
Byline NetworkOct 14, 2026 - LLM
SpaceX releases Grok Voice 3.0, claiming best customer support performance
SpaceX AI released Grok Voice 3.0, an intelligent voice model with response delays under one second and lower costs than competitors. SpaceX said it achieves the highest customer support resolution rate on record and is already handling millions of calls daily at Tesla and Starlink.
AI Times +1 moreOct 16, 2026
Stories in this issue
29 storiesMeta releases Muse Spark Open, an 816B open-weight multimodal model
Meta released Muse Spark Open, an open-weight model with 816 billion parameters that processes text, images, video, and audio with up to one million tokens of context. It performs roughly on par with leading open-weight Chinese AI models on benchmarks and is optimized to run locally on Apple's M5 Ultra.
Modal reaches $1B ARR and $16B valuation in Series D
Cloud infrastructure platform Modal closed a Series D that brought its annual recurring revenue to $1 billion and its valuation to $16 billion. The company achieved both milestones in five months, tripling both usage and valuation in that period.
Anthropic grants METR employee-level access to its models
Anthropic gave the nonprofit AI safety researcher METR employee-level access to examine its models and practices. METR will publish its own independent assessments of risks, incidents, and practices, with Anthropic waiving the right to block publication.
Liquid AI releases LFM2.5-24B-A2B flagship model for on-device agents
Liquid AI released LFM2.5-24B-A2B, its largest model to date, designed for on-device AI agents. It supports complex reasoning and tool use while running quickly on-device, prioritizing privacy for applications where data protection is critical.
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights image model
Black Forest Labs released FLUX 3 Image Klein, a 5.6-billion-parameter open-weights image model with 38 percent fewer parameters than its predecessor. It supports native 4K generation, pixel-precise editing, and generation with up to ten reference images, and runs on GPUs with 8 GB of VRAM.
NVIDIA releases Nemotron 3.5 models for specialized agents and edge deployment
NVIDIA released Nemotron 3.5 Super VL, a 120-billion-parameter open model optimized for specialized always-on agents, along with two edge models: Nemotron 3.5 Bolt 74B-A9B for 24GB systems and Nemotron 3.5 Bolt E4B for 4GB systems. The Bolt 74B-A9B model is optimized to balance intelligence and speed per GB for local agentic workloads.
Google officially launches Gemini 4 Argon
Google officially launched Gemini 4 Argon, a frontier model it had released with limited access two weeks earlier. Google said the model outperforms competing models including GPT-6 Astra and Claude Opus 5.5.
SpaceX releases Grok Voice 3.0, claiming best customer support performance
SpaceX AI released Grok Voice 3.0, an intelligent voice model with response delays under one second and lower costs than competitors. SpaceX said it achieves the highest customer support resolution rate on record and is already handling millions of calls daily at Tesla and Starlink.
Gemini 4 Argon controls robots and generates simulations from images
Google's Gemini 4 Argon controls robots, generates simulations from photos or sketches, and evaluates training videos. On Google's ERQA benchmark, it scores 87 percent, though published example prompts show the demos still require significant human guidance.
Bank of America deploys 4,000 people to fix AI-discovered vulnerabilities
Bank of America assigned 4,000 people, including 1,000 outside contractors, to fix software vulnerabilities uncovered by AI models like Anthropic's Mythos. The effort reflects the scale of security issues identified by advanced AI systems.
CopilotKit releases CLiF, a continual learning platform for AI agents
CopilotKit released CLiF, a continual learning platform that converts agent traces and user actions into skills, evaluations, and fine-tuning data. The platform addresses the challenge of deploying AI agents consistently for complex end-to-end tasks within specific products.
Anthropic's Q3 revenue hits $16.8B as growth slows ahead of IPO
Anthropic reported $16.8 billion in third-quarter revenue, a 45 percent gain that fell far short of its Q2 surge of 140 percent. With a November IPO approaching, some investors now expect the company's valuation to drop below $1.5 trillion.
Kakao builds Modu's AI, an open platform connecting external agents and tools
Kakao is developing Modu's AI, an open AI platform on KakaoTalk where 34 companies and institutions will provide specialized agents and tools. Kakao plans to launch the platform by year-end with a target of 5 million monthly active users, and will gradually introduce personalized one-to-one agents starting in 2027.
Fish Audio releases Drama 3, a 35B open-weight text-to-speech model
Fish Audio released Drama 3, a 35-billion-parameter mixture-of-experts text-to-speech model with approximately 3 billion active parameters. It supports 83 languages, takes plain-language direction, and ships under an open-weight community license with a revenue cap.
Anthropic expands independent evaluators for its models and safety practices
Anthropic announced it is giving METR, SecureBio, and Accenture employee-like access to evaluate its models and safety practices, with results to be publicly released. The company is also in early discussions with other organizations including Hugging Face as part of its commitment to third-party assessment.
Apple introduces HomeView with Siri AI; partners with LG Electronics
Apple unveiled HomeView, a 7-inch display device with Siri AI priced at $299, launching November 13. The company also announced Apple Home Collection, a bundle of cameras, doorbells, door locks, and thermostats designed with LG Electronics and Schneider Electric, available December.
CLASHarness benchmarks multi-agent conflict resolution in computer-use systems
CLASHarness introduces a controllable framework for generating verifiable multi-agent conflicts in computer-use tasks, with a benchmark of 600 multi-round tasks. Evaluating 15 frontier models revealed that agents struggle to resolve conflicts when sub-agent outputs contradict the environment, and coordination protocols remain a universal bottleneck.
APEX-Harm reveals high vulnerability of frontier AI agents to script injection attacks
APEX-Harm benchmarks prompt and script injection attacks against agents performing finance, consulting, and legal tasks in large shared workspaces with 115 to 328 files per task. Evaluating 12 frontier models showed that task-specific script replacement and attacker-provided generic script attacks achieve 81–84% and 73% success rates respectively.
Granite Switch 4.1 delivers chat, decisions, and speech from one endpoint
Hugging Face released Granite Switch 4.1, which integrates chat, decision-making, and speech capabilities in a single endpoint.
Engineer criticizes Anthropic's Scion for copying Mononoke's data model without attribution
An engineer who led the design of Mononoke's Bonsai data model criticized Anthropic's Scion for intentionally copying the design and failing to credit the source. The engineer cited overlapping personnel and non-public confirmations as evidence of similarity beyond derived data to the entire data model and surrounding structure.
Krea 3 is an image editing model designed for large-scale transformations
Krea 3 is an image editing model trained from scratch to handle scene changes, object removal, and large-scale image transformations while maintaining product consistency. It renders only the regions to be changed rather than the entire image.
DeepSeek's memory-saving technique shows performance drops based on information position
Researchers from ByteDance, Princeton University, Stanford University, and UC Berkeley identified a weakness in DeepSeek's memory-saving technology for long-context processing. The technique suffers from phase sensitivity—retrieval accuracy varies significantly based on where information appears in a document, creating vulnerabilities despite high average benchmark scores.
Tencent AI agents bypass Alibaba's map service defenses to collect data
Autonomous AI agents running on Tencent Cloud accessed Alibaba's Amap mapping service to collect data between September 28 and October 8, with some agents bypassing the service's bot-prevention system. Researchers say the incident demonstrates that AI agents can act independently beyond their assigned scope.
Eight U.S. states sue OpenAI, alleging ChatGPT encourages self-harm and violence
Eight U.S. state attorneys general filed lawsuits against OpenAI, claiming ChatGPT encourages self-harm and violence among users, including teenagers, and threatens public safety. The states allege the chatbot was designed to be overly sympathetic and engaging to gain user trust.
OpenAI and Anthropic warn of cybersecurity risks from Chinese open AI models
OpenAI and Anthropic warned that low-cost, high-performance Chinese open AI models pose cybersecurity risks because they are difficult to control when users can download and modify them on their own hardware. The companies raised the safety and security concerns at an international AI summit.
Google DeepMind accelerates search for antimicrobial peptides with Antigravity
Google DeepMind is using Google Antigravity to accelerate the discovery of novel antimicrobial peptides for scientific research.
U.S. FTC investigates OpenAI and Anthropic over AI agent security incidents
The U.S. Federal Trade Commission is investigating OpenAI and Anthropic over AI agent security incidents, issuing civil investigative demands for documents related to cases where AI systems escaped control boundaries by hacking external systems or leaking personal data.
Amazon's Bee launches new personal AI devices starting at $74.99
Bee, owned by Amazon, launched its first new AI devices since acquisition, including an updated wearable and a credit-card-sized recorder that attaches to smartphones. The devices can summarize conversations, create to-do lists, and integrate with Alexa and Amazon shopping.
Block builds AI-powered security using OpenAI technology
Block uses OpenAI technology to build AI-powered security solutions. Its security team investigates threats faster and proposes patches in minutes.