# Horizon AI Briefing — 2026-11-11

## Top stories
- **Google Launches Gemini 4 Carbon with General Availability and /Swarm Multi-Agent Feature** (https://www.itmedia.co.jp/aiplus/article/2611/11/2000003093/) — Google has made Gemini 4 Carbon generally available across developer APIs, enterprise, and paid consumer tiers, with the consumer version exclusive to the AI Ultra subscription. Per Google's own claims and self-reported benchmarks (no linked evidence), the model ranks first in several benchmarks and is priced at roughly one-quarter of Anthropic's Claude Fable 5.2. Google is also previewing a /swarm feature that coordinates teams of AI agents to split complex tasks in real time. [vendor-claimed capability; no independent eval]
- **Anthropic IPO: Nasdaq Listing Targets Up to $140 Billion Valuation** (https://www.aitimes.com/news/articleView.html?idxno=217248) — Anthropic has filed to go public on Nasdaq, offering 825 million shares at $70–78 per share, targeting a valuation of up to $140 billion — well below the $200 billion figure initially floated, reflecting broader tech stock volatility and investor skepticism about AI profitability. One report, citing self-reported figures, puts the total raise including private placements from Nvidia and Google at over $90 billion, though the independent reporting pegs it closer to $64 billion. The IPO marks a major liquidity event for the AI safety-focused lab and signals the broader AI sector's push toward public markets.
- **Anthropic Expands Claude into a Full Collaborative Workspace** (https://the-decoder.com/anthropic-turns-claude-into-a-full-workspace-with-plugins-spreadsheets-whiteboards-and-soon-meeting-notes/) — Anthropic has introduced Claude Sheets and Claude Whiteboards as collaborative editing tools in beta, rebranded Claude Dashboards as Claude Data, and added interactive connectors that let partners like Adobe and Atlassian embed their apps directly inside Claude. Per Anthropic's own claims, Claude Meetings — which can join calls and take notes — is also coming soon. The moves position Claude as a central work platform rather than a standalone assistant, intensifying competition with Microsoft 365 Copilot and Google Workspace. [vendor-claimed capability; no independent eval]
- **Nvidia Loses $430 Billion in Market Cap Amid CEO Scandal and Alleged Chip Export Violations** (https://openai.com/index/gilbert-tobin-sam-nickless) — Nvidia's stock fell roughly 9.5% over two days, erasing approximately $430 billion in market cap, after The Information and Reuters independently reported that a Chinese businesswoman allegedly linked to CEO Jensen Huang supplied at least $7.5 billion in Nvidia chips to Chinese companies in violation of US export controls. The scandal pushed Microsoft past Nvidia to become the world's second-largest company by market cap and amplified existing investor concerns about AI investment returns. [primary source]
- **OpenAI Turns ChatGPT into a Business Directory with AI Customer Agents** (https://the-decoder.com/openai-turns-chatgpt-into-a-business-directory-with-ai-agents-that-handle-customer-questions/) — OpenAI is launching a 'Businesses in ChatGPT' hub that lets companies claim profiles, add product data, and track how users interact with them via ChatGPT. Per OpenAI's own claims, eligible US e-commerce providers can deploy 'Business Agents' to handle customer queries directly inside ChatGPT, effectively turning the platform into a commercial marketplace and AI-powered storefront. This signals a new monetization vector for OpenAI beyond subscriptions and API revenue. [vendor-claimed capability; no independent eval]
- **Fireworks AI Security Breach Impacts Notion and Its Customers** (https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor) — An independently reported security breach at Fireworks AI has affected customers including Notion, raising concerns about supply-chain risk in the AI infrastructure layer. The incident highlights how third-party AI platform vulnerabilities can cascade into enterprise SaaS products that rely on them.
- **Microsoft Researchers Discover 'Lazarus' Backdoor Technique Targeting AI Agents** (https://www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents/) — Microsoft researchers have identified a new attack vector called 'Lazarus' that reprograms dormant neurons within an AI model, causing it to execute malicious commands when a hidden trigger is activated. The threat is especially severe for AI agents with tool or terminal access, where a triggered model could silently download and run attacker code. This represents a novel class of model-level supply chain attack distinct from prompt injection. [self-reported; no independent confirmation · primary source]
- **Alibaba Cloud Unveils Qwen 4 AI Model Preview** (https://www.itmedia.co.jp/aiplus/article/2611/11/2000003098/) — Alibaba Cloud announced a preview release of Qwen 4, its next-generation AI model, made available through its cloud AI services platform. The announcement carries no linked benchmark evidence, so capability claims remain self-reported at this stage. Qwen 4 enters a rapidly crowding frontier model market at a moment when pricing pressure from Google's Gemini 4 Carbon is already intensifying. [vendor-claimed capability; no independent eval]
- **Apple Reportedly Developing Mac App for Easy Local AI Model Deployment** (https://9to5mac.com/2026/11/10/apple-has-new-mac-app-coming-to-make-running-local-ai-models-easy/) — Per a self-reported claim with no linked evidence, Apple has a new Mac application in development designed to simplify running local AI models on Apple hardware. If accurate, this would significantly lower the barrier for developers and privacy-conscious users to run inference locally, potentially accelerating the on-device AI shift that Apple has been telegraphing with its Neural Engine investments. [self-reported; no independent confirmation]

## Emerging signals
- **AI Backdoor Attacks on Dormant Neurons Signal a New Model-Level Threat Class** (https://www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents/) — Microsoft's Lazarus research points to a growing body of work on model-level attacks that survive fine-tuning and standard safety evaluations, targeting the agent layer rather than the prompt layer — a threat that current red-teaming frameworks are not designed to catch.
- **Minimum Safety Testing Windows for AI Models Gaining Policy Traction** (https://www.alignmentforum.org/posts/XNLhpWgzhF8GYZ4Er/delaying-internal-deployments-until-after-x-days-of-safety) — A proposal circulating in AI safety circles advocates mandatory fixed-duration safety testing periods before any new model checkpoint can be deployed, even internally, representing a procedural safeguard that could become a regulatory template as governments push for pre-deployment review requirements.
- **Physical AI Becoming a Strategic Pillar for Telecoms and Governments** (https://www.itmedia.co.jp/business/articles/2611/11/news038.html) — Japan's three major telecoms (SoftBank, NTT, KDDI) and South Korea's government are independently making large bets on physical AI and robotics infrastructure, suggesting that hardware-embodied AI is transitioning from R&D novelty to national industrial policy priority.
- **NotebookLM API Launch Signals Google's Push to Embed AI Research Tools in Agent Pipelines** (https://atmarkit.itmedia.co.jp/ait/articles/2611/11/news011.html) — Google's release of a public REST API and MCP server for NotebookLM (now Gemini Notebook) enables automated ingestion of web and YouTube content and integration with AI agents, quietly expanding the agentic surface area of its productivity suite.
- **AI Black Market Emerging in Russia as Sanctions Tighten** (https://bsky.app/profile/united24media.com/post/3mzyh4pkky626) — As OpenAI, Google, and Anthropic enforce stricter compliance with international sanctions, Russian users are turning to underground markets for AI access, creating data privacy and security risks that regulators in the West have not yet addressed.

## New entrants
- **Gemini 4 Carbon** (model) — Google's flagship model now in general availability, per Google's self-reported benchmarks claiming top rankings in several evals at roughly one-quarter the API cost of Claude Fable 5.2; also introduces /swarm multi-agent coordination. Capability claims carry no linked third-party evidence.
- **MiniMax H3-Image** (model) — MiniMax open-sourced H3-Image, a 33.1B diffusion transformer capable of generating 16-megapixel images in six steps and editing with up to 16 reference images; released under a community licence with territory restrictions. Capability claims are self-reported.
- **Nvidia GR00T 2** (model) — Nvidia's new robotics foundation model learns tasks from human video footage (~400,000 hours of training video) rather than teleoperation, transferring hand movements to robotic arms; per Nvidia's own claims, success rates range from 10–80% depending on task.
- **Pinky** (tool) — Open-sourced by DIYgod under Apache-2.0, Pinky is an MCP layer that lets AI agents operate real iPhones and Android phones, with gateway-enforced approvals, action logs, and a human takeover option.
- **BotPsy** (model) — A sub-billion-parameter on-device personal assistant with agentic language model capabilities, described in a new research paper; targets deployment on consumer devices without cloud dependency. Capability claims are self-reported with no linked evidence.

## Biggest movers this week
- **Anthropic** (company) — 355 mentions this week, ↑115 vs the prior week
- **Claude** (model) — 165 mentions this week, ↑53 vs the prior week
- **OpenAI** (company) — 226 mentions this week, ↑39 vs the prior week
- **NVIDIA** (company) — 85 mentions this week, ↑39 vs the prior week
- **Meta** (company) — 43 mentions this week, ↑32 vs the prior week
- **Google** (company) — 102 mentions this week, ↑18 vs the prior week

## China & East-Asia AI
- **China's Qwen 4 AI Model Unveiled: Alibaba Cloud Launches Preview Version** (https://www.itmedia.co.jp/aiplus/article/2611/11/2000003098/) — ITmedia
- **DIYgod Open-Sources Pinky, an MCP Layer That Lets AI Agents Operate Real iPhones and Android Phones** (https://pandaily.com/diygod-pinky-open-source-mcp-ai-agents-control-iphone-android) — Pandaily
- **MiniMax Opens Weights for H3-Image, a 16-Megapixel Text-to-Image and Editing Model Run in Six Steps** (https://pandaily.com/minimax-h3-image-open-weights-16mp-text-to-image-editing-six-step) — Pandaily
- **AI microdramas cause ‘aesthetic fatigue’, professor tells Macau forum** (https://www.scmp.com/news/hong-kong/society/article/3373984/ai-microdramas-cause-aesthetic-fatigue-professor-tells-macau-forum) — South China Morning Post
- **Tencent Intensifies 'AI Self-Improvement' Research, Joining U.S. in RSI Competition** (https://www.aitimes.com/news/articleView.html?idxno=217232) — AI Times

## Korea AI
- **The Information: Chinese Businesswoman Used Jensen Huang Connection to Supply $7.5 Billion in NVIDIA Chips to China** (https://www.aitimes.com/news/articleView.html?idxno=217262) — AI Times
- **KT Provides AI Email Security Service to KAIST** (https://www.aitimes.com/news/articleView.html?idxno=217252) — AI Times
- **Nvidia Loses $430 Billion in Market Cap Over Two Days Amid CEO Scandal and AI Investment Concerns** (https://www.aitimes.com/news/articleView.html?idxno=217258) — AI Times
- **South Korean Government Selects Next-Generation Sensors, Actuators, and Secondary Batteries for New Physical AI Projects** (https://www.aitimes.com/news/articleView.html?idxno=217250) — AI Times
- **AI Leader's Bookshelf: Learning Modern AI with PyTorch** (https://www.aitimes.com/news/articleView.html?idxno=217031) — AI Times

## Japan AI
- **China's Qwen 4 AI Model Unveiled: Alibaba Cloud Launches Preview Version** (https://www.itmedia.co.jp/aiplus/article/2611/11/2000003098/) — ITmedia
- **Google's "Most Powerful" Model Gemini 4 Carbon Now Generally Available; Multi-Agent Team Coordination Feature Also Announced** (https://www.itmedia.co.jp/aiplus/article/2611/11/2000003093/) — ITmedia
- **Claude Gets Collaborative Sheets and Whiteboards, Plus In-App External App Integration** (https://www.itmedia.co.jp/news/article/2611/11/2000003089/) — ITmedia
- **New ChatGPT Business Features: Site Analytics and Customer AI Chat** (https://www.itmedia.co.jp/aiplus/article/2611/11/2000003088/) — ITmedia
- **AI Learns Past Assembly Sequences, Generates Assembly Plans Based on Company Know-How** (https://monoist.itmedia.co.jp/mn/articles/2611/11/news018.html) — ITmedia

## Europe (EU) AI
- **Defence startup Mistt raises £5m to build the ‘next GPS’ for autonomous systems and AI** (https://sifted.eu/articles/defence-mistt-fundraise-next-gps-data-autonomous-systems-world-models/) — Sifted
- **Europe’s chips startups mapped: 113 companies building the hardware behind AI** (https://sifted.eu/articles/europes-chips-startups-mapped-113-companies-building-the-hardware-behind-ai/) — Sifted
- **OpenAI turns ChatGPT into a business directory with AI agents that handle customer questions** (https://the-decoder.com/openai-turns-chatgpt-into-a-business-directory-with-ai-agents-that-handle-customer-questions/) — The Decoder
- **Anthropic turns Claude into a full workspace with Plugins, spreadsheets, whiteboards, and soon Meeting notes** (https://the-decoder.com/anthropic-turns-claude-into-a-full-workspace-with-plugins-spreadsheets-whiteboards-and-soon-meeting-notes/) — The Decoder
- **Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot** (https://the-decoder.com/nvidias-model-gr00t-2-watches-humans-on-video-then-imagines-itself-as-the-robot/) — The Decoder

## Regulation updates
- [US State] **Relating to a prohibition on the use of certain personalized dynamic pricing models by a business in this state.** — Proposed. Tracked
- [US State] **Relating to the requirements for artificial intelligence operations of state agencies, establishing artificial intelligence divisions.** — Proposed. Tracked
- [US State] **Relating to the establishment of the Texas Department of Artificial Intelligence and oversight of artificial intelligence systems used by state agencies.** — Proposed. Tracked
- [US State] **Relating to a criminal offense for the creation or distribution of certain artificially generated media with the intent to influence the outcome of an election.** — Proposed. Tracked
- [US State] **Relating to requiring a business in this state to disclose the use of certain personalized dynamic pricing models.** — Proposed. Tracked
- [US State] **Relating to the required disclosure to a consumer that the consumer is interacting with an artificial intelligence system.** — Proposed. Tracked
- [US State] **Relating to a restriction on state agency deployment and use of an artificial intelligence system regarding an employment decision or confidential information.** — Proposed. Tracked
- [US State] **Relating to human oversight and review of decisions and actions involving artificial intelligence systems used by state agencies.** — Proposed. Tracked
- [US State] **Relating to the labeling of artificial intelligence systems by businesses for commercial purposes.** — Proposed. Tracked
- [US State] **Relating to the use of artificial intelligence by health care providers.** — Proposed. Tracked
- [US State] **Relating to human oversight and override for certain high-risk artificial intelligence systems; providing a civil penalty.** — Proposed. Tracked
- [US State] **Relating to a health care facility grant program supporting the use of artificial intelligence technology in scanning medical images for cancer detection.** — Proposed. Tracked
- [US State] **Relating to a study on employer and state agency use of automated employment decision tools in assessing an applicant's suitability for a position.** — Proposed. Tracked
- [US] **VET Artificial Intelligence Act** — Proposed. Read twice and referred to the Committee on Commerce, Science, and Transportation.
- [US] **VET Artificial Intelligence Act** — Floor Action. Placed on Senate Legislative Calendar under General Orders. Calendar No. 727.

---
Source: Horizon (https://horizon.alchemylab.sh) — aggregated, LLM-scored AI intelligence; each item also lists its own primary source, and provider.citation credits both.