Weekly AI Briefing — Week ending 17 October 2026

A weekly roundup of the Signal items we published this week (11–17 October 2026).

• White House orders federal agencies to say "Super Intelligence" instead of "AI" — Executive Order 14434 directs the executive branch to replace "AI" with "Super Intelligence" across official documents and communications. The White House (Executive Office of the President)
• Anthropic's Claude Haiku 5.5 cuts small-model running costs by about 75% — Anthropic launched Claude Haiku 5.5, its cheapest and fastest small model, priced roughly 75% below Haiku 4.5 for high-volume, cost-sensitive tasks. Anthropic (official newsroom)
• Google DeepMind's EmbeddingGemma 2 puts image, audio and video in one on-device model — EmbeddingGemma 2 is a 740M-parameter open model (Apache 2.0) unifying text, code, images, video and audio in one embedding space for on-device use. Google DeepMind (official blog)
• Nvidia and Microsoft push AI agents onto Windows PCs with RTX Spark — At a Microsoft event, Jensen Huang and Satya Nadella unveiled RTX Spark and NVIDIA DGX Station for Windows to run AI agents locally on PCs. NVIDIA (official blog)
• Meta extends its AI agent Muse to run small-business chores — Meta launched Muse for Small Business, adding connectors to tools like Shopify, Stripe and QuickBooks so its personal AI agent can handle business tasks. Meta (official newsroom)
• China's standards committee rolls out AI Safety Governance Framework 3.0 — At National Cybersecurity Week 2026, China's cybersecurity standardization committee released AI Safety Governance Framework 3.0, revising risk categories and safeguards. Cyberspace Administration of China (CAC)
• China releases a 120GB Chinese-language corpus plus AI safety datasets — Chinese authorities released Internet Corpus 4.0 (120GB) alongside cybersecurity and AIGC-detection datasets to support model training and safety work. Cyberspace Administration of China (CAC)
• China's generative-AI filing list tops 1,100 services after 124 new entries — China added 124 generative-AI services to its national filing list for July and August 2026, taking the cumulative total to 1,112. Cyberspace Administration of China (CAC)
• China's internet regulator names cases, flagging unlabeled AI-generated content — China's CAC published enforcement cases across cybersecurity and data protection, including a firm punished for failing to label AI-generated content. Cyberspace Administration of China (CAC)
• Three on-device generative-AI services filed in China, including 'YOYO Claw' — China's regulator announced filing for three new phone on-device generative-AI services, including one named YOYO Claw. Cyberspace Administration of China (CAC)
• Google's Gemini 4 Argon reaches cyber defenders first, with a 1M-token output limit — Google's newest frontier model, Gemini 4 Argon, rolls out to trusted cyber defenders first and lifts the output limit to 1M tokens at $2 and $10 per million. Google (official blog)
• Anthropic pledges $150M to the federal Genesis Mission, opening Claude to 15-plus agencies — Anthropic will commit $150 million over three years to the Genesis Mission and extend Claude, Claude Code and API credits to more than 15 U.S. agencies. Anthropic (official newsroom)
• Anthropic opens a $100M academy to train 10,000 enterprise AI engineers by 2027 — Backed by $100 million, Anthropic's Claude Frontier Academy aims to train 10,000 Frontier Deployed Engineers by the end of 2027, starting with big consultancies. Anthropic (official newsroom)
• Microsoft's RetroChimera AI plans how to build molecules, with GSK and Novartis on board — A Microsoft Research AI model called RetroChimera, detailed in Nature, maps synthesis routes for target molecules and adapts to GSK's proprietary chemistry data. Microsoft (official newsroom)
• Anthropic rewrites its usage rules as Claude takes on longer, more independent work — Anthropic's 2026 Usage Policy overhaul adds a deceptive-activity section, reworks elections and surveillance rules, and takes effect November 12. Anthropic (official newsroom)
• DeepSeek retires its flagship as a smaller model, V4.1-Flash, outruns it — DeepSeek released V4.1-Flash on Sept 10, a 552B-parameter model it says beats its own V4-Pro flagship on benchmarks and cost, and it is phasing the larger model out. DeepSeek (official newsroom)
• China's software ministry sets 2028 targets to rebuild the industry around AI agents — MIIT's new AI + Software plan targets 20,000 firms, 100 agent-software benchmarks and a shift from people writing code to human-machine teams by 2028. Ministry of Industry and Information Technology (MIIT), China
• China benchmarks AI's role in cyber defence across eight scenarios — A CAC-guided testing exercise drew 251 teams from 151 organisations to measure AI in cyber defence, from AIGC image detection to spotting rogue AI-agent behaviour. Cyberspace Administration of China (CAC)
• China's cabinet puts the computing-power network at the centre of its AI build-out — At a State Council executive meeting, China called the computing-power network the foundational support for AI and ordered closer coordination of compute, the power grid and fibre. National Data Administration (NDA), China
• China's clean-up of 'AI application chaos' clears 5.6m pieces of content in its second phase — CAC's second-phase campaign against AI misuse removed over 5.6m pieces of content, punished 49,000 accounts and disposed of 2,400-plus apps and sites. Cyberspace Administration of China (CAC)
• Google opens its AI watermark detector to everyone — Google DeepMind has opened its SynthID Detector to everyone worldwide, letting anyone check whether an image, video or audio file carries Google's AI watermark. Google DeepMind
• Google Cloud puts a universal agent at the centre of workplace tasks — Announced at Gemini at Work 2026, Google Cloud's new Gemini agent plans and finishes work across a company's systems from a single prompt box. Google Cloud
• Microsoft's Maia 200 enters production as Maia 300 targets 2027 — Microsoft says its Azure Maia 200 chip is now serving production workloads, and it will scale the Maia family to multi-gigawatt deployments from 2027. Microsoft
• Nvidia pledges $1 billion to a US super-intelligence science push — Nvidia says it will commit $1 billion over five years to US super-intelligence and quantum research, tied to the White House's Genesis Mission. NVIDIA
• Anthropic splits access to its cyber capabilities into three tiers — Anthropic has expanded its Cyber Verification Program into three tiers, giving vetted security teams stronger cyber capabilities with fewer safeguards. Anthropic
• China folds a dedicated 'AI+' action into its new national productivity doctrine — A Central Committee and State Council document sets 19 tasks for 'new-quality productive forces' and hands AI its own chapter: application bases, on-device AI and safety oversight. State Council of China (via Xinhua)
• Why Beijing put 'AI+' at the heart of its productivity plan — in its own words — The NDRC's Q&A on the new quality-productivity Opinions gives three reasons it singles out a nationwide 'AI+' action, from strategy and industry to safety. National Development and Reform Commission (NDRC), China
• Tencent's WorkBuddy wants to end the copy-paste loop between you and AI — Tencent's new agent edits the files you already have open — Word, Excel, PowerPoint or Tencent Docs — and pulls in outside apps through the MCP standard. Tencent (official newsroom)
• Huawei pitches an 'AI-centric' broadband network to carry 100,000-GPU clusters — At UBBF 2026 Huawei laid out a network redesign for giant AI clusters: millisecond compute access, 99.999% carrier availability and quantum-safe gear at lower build cost. Huawei (official newsroom)
• Huawei Cloud opens its embodied-AI stack to outside developers — Huawei Cloud closed its 'Xingyao' developer contest with 27 awards while showcasing five AI platforms, including CloudRobo for embodied robots. Huawei Cloud (official newsroom)
• Anthropic opens its fifth Asia-Pacific office in Singapore, paired with a government AI-safety pact — Anthropic opened a Singapore office and signed an MOU with the country's digital ministry to cooperate on AI safety, security and model evaluation. Anthropic
• Google DeepMind and Singapore build a scam detector that never leaves the phone — Google DeepMind and GovTech Singapore fine-tuned a privacy-preserving scam-detection model that runs entirely on-device, flagging manipulative chats in real time. Google DeepMind
• Nvidia pledges $250 million in compute to UK agentic-safety research — Nvidia pledged a five-year, $250 million compute donation to UK agentic safety and security research, delivered through UK-based cloud partners. NVIDIA
• Microsoft's HASTE turns satellite images into disaster damage maps within hours — Microsoft's AI for Good team uses the open-source HASTE satellite-analysis tool to map disaster damage in hours, aiding responders after 23 disasters since 2023. Microsoft
• Meta's Muse Wonder models chase "personal superintelligence" with persistent agents — Meta introduced Muse Wonder and Muse Wonder Full, frontier models for personal and professional agent tasks, adding Active Memory and Active Focus features. Meta
• Shanghai's 'quantum city' gets its first government AI agents, and SenseTime built three of them — Shanghai's planning authority released the first batch of 'quantum city' government agents. SenseTime's infrastructure arm says it built and delivered three of them, for flood dispatch, spatial insight and planning review. SenseTime (official newsroom)
• SenseTime's coding agent helped a blind teacher build study tools for visually impaired students — A blind teacher in Shanxi used SenseTime's Raccoon Work agent to build 'Light Bridge', an accessibility assistant that turns unwieldy textbooks into screen-reader-friendly study material. SenseTime says it won silver at its own challenge. SenseTime (official newsroom)
• MiniMax courts Japanese IP owners with a rights-aware video model — At a Tokyo conference with the KAGAMIAI association, MiniMax launched 'H3 IP Edition' and pitched a licensing pipeline it calls Kagami Gate, arguing AI video should pay rights holders rather than simply scrape their work. MiniMax (official newsroom)
• Alibaba's QwenWork turns its workplace agent on China's online merchants — Alibaba launched 'QwenWork E-Commerce' in invite-only beta, running its workplace AI agent across a merchant's planning, marketing, content and cross-platform advertising, with one dashboard that pulls store data together. Alibaba Cloud (official blog, via Alizila)
• Alibaba's Qoder now grades the agent's workspace, not just its answers — Qoder launched 'Better Harness' in beta, a tool that inspects the whole environment a coding agent works in — goals, permissions, verification, delivery — and turns each task's mistakes into reusable fixes. Alibaba Cloud (official blog)
• The White House bets on K-12 schools to build 'super intelligence' readiness — The White House's 2027 Presidential SI Challenge invites K-12 students and schools to work with 'super intelligence,' adding a new Seal of SI Readiness for participating schools. The White House
• Anthropic: politics is 5.75% of US Claude chats, but vote advice is just 0.01% — Anthropic says politics and policy talk makes up 5.75% of US Claude.ai conversations and direct voting advice just 0.01%, as it adds official election data ahead of the midterms. Anthropic
• Nvidia's cuPotential lets AI agents simulate atoms inside 'AI-defined' chip fabs — Nvidia's new cuPotential library lets AI agents simulate atomic and molecular interactions, part of a CUDA-X push to automate 'AI-defined' chip factories. NVIDIA
• Meta opens its Facebook ranking algorithm to plain-language user instructions — Meta is letting Facebook and Reels users describe in their own words what they want to see, feeding freeform text straight into its ranking systems in real time. Meta
• Google's Gemini 4 Argon clears all 12 simulated cyberattacks, with new 'activation probe' guardrails — As Gemini 4 Argon rolls out broadly, Google details three layers of real-time safety controls and says the model solved all 12 simulated cyberattack scenarios. Google
• Huawei pitches a core network built for AI agents, not just phones — Huawei's AgenticCore won a NetworkX 2026 award for reworking the mobile core so AI agents, not only people, are first-class network users. Huawei (official newsroom)
• Alibaba graded Qwen3.8-Max against Claude and GPT-5.6 on one agent harness — Alibaba ran Qwen3.8-Max and three rival models through the same coding-agent harness on five long tasks - it lost to Claude, but was the most GPU-thrifty. Alibaba Cloud (official blog)
• Asked to add features, Alibaba's coding agent re-found five old CVEs — Alibaba gave Qoder Agent five ordinary feature requests from real projects, then ran its new security scan - it flagged the same flaws that had shipped as past CVEs. Alibaba Cloud (official blog)
• SenseTime says its Raccoon office agent held China's No. 2 slot for two months — SenseTime cites a third-party China AI-app ranking where its Raccoon office agent stayed No. 2 on visit growth and moved up on usage time in September. SenseTime (official newsroom)
• Ant Group runs a company-wide AI agent platform entirely inside its own network — Alibaba Cloud details how Ant Group deployed QwenWork for tens of thousands of staff as a private, Kubernetes-based agent foundation behind its own security boundary. Alibaba Cloud (official blog)
• Anthropic gives METR evaluators the same access its own safety team gets — Anthropic says it now grants a METR team employee-like access to assess its models, and is working with SecureBio and Accenture as well. Anthropic (official newsroom)
• Nvidia opens a 120B agent model plus two edge models for on-device agents — Nvidia's open Nemotron 3.5 Super VL targets always-on agents, joined by Bolt 74B and Bolt E4B models for local hardware. NVIDIA (official blog)
• DeepMind's Antigravity agent surfaces a new antimicrobial peptide candidate — Using its Antigravity platform and ProtNLM, DeepMind found a novel antimicrobial peptide that is now headed to wet-lab validation. Google DeepMind (official blog)
• Bar-Ilan University's course-grounded AI agent cut students' study time by hours — A Copilot Studio agent grounded in course materials is expanding to 100+ courses after a 2,400-student pilot at Israel's Bar-Ilan University. Microsoft (official newsroom)
• Mercedes-Benz and Nvidia extend their bet on AI-driven automated driving — Mercedes-Benz and Nvidia will jointly build next-gen automated driving systems, pairing MB.OS with Nvidia's Hyperion and Halos safety stack. NVIDIA (official newsroom)
• Alibaba's Qoder team opens a managed runtime for long-running 'personal agents' — Alibaba's Qoder team introduced Cloud Agents, a managed platform that runs long-horizon agent tasks in the cloud and returns progress in real time. Alibaba Cloud (official blog)
• Huawei says its AI WAN has reached 60-plus carriers, pitching security and lossless compute — Huawei says its AI WAN solution is now in commercial deployment with more than 60 carriers, bundling in-network security, lossless compute and L4 autonomous-network features. Huawei (official newsroom)
• Qoder's new Fast tier cuts first-response time from eight seconds to three — Alibaba's Qoder introduced a Fast tier that trims AI-coder first-response time from 8s to 3s and lowers credit cost, aimed at command execution and small edits. Alibaba Cloud (official blog)
• Huawei tops an OSS market-share report and bets on multi-agent network operations — Huawei says Appledore Research ranked its ADN solution first in global OSS software share, as it pitches agent-to-agent coordination between network systems. Huawei (official newsroom)
• SenseTime lands on Fortune and BCG's 'Future 120' after its first profitable half — SenseTime was named to Fortune and BCG's 2026 Fortune Future 120 list, one of 11 Chinese companies, after reporting its first profit under IFRS. SenseTime (official newsroom)
• NIST, NSA and CISA set ten practices for containing AI agent incidents — Three U.S. agencies issued voluntary guidance for handling incidents when AI agents under training take uninstructed actions. NIST (official newsroom)
• Nvidia puts 128GB and a local agent runtime into a Windows laptop — Nvidia's RTX Spark laptops ship with up to 128GB unified memory and 1 petaflop of AI compute, pitching local agents as a PC feature. NVIDIA (official newsroom)
• Mercedes' F1 strategists moved race simulations from overnight to trackside — Mercedes-AMG F1's race strategist says Azure Batch cut full-race simulations from about an hour to seconds, run trackside. Microsoft (official newsroom)
• Adobe's creative apps get a local AI speed boost on Nvidia's new chip — Adobe optimized Photoshop, Premiere and Substance 3D for RTX Spark, claiming up to 2x faster AI work, with Frame.io gains. NVIDIA (official newsroom)
• Google expands AI early-warning tools as it publishes a crisis-resilience review — Google says its flood, cyclone and fire models reach billions of people; it published a white paper on crisis resilience. Google (official blog)