AI News: October 7, 2026

1. DeepSeek Is Closing a Funding Round of at Least $12 Billion at a $75 Billion Valuation

DeepSeek. Bloomberg reports the Chinese lab is raising at least $12 billion and possibly $15 billion, well above its original $7.5 billion target, at roughly a $75 billion valuation, with battery maker CATL and Tencent taking the largest shares. The company plans to restructure and list in early 2027 and is building a data center with 160,000 Huawei chips plus its own inference chip. The round signals investor confidence in DeepSeek’s low-cost strategy, as V4-Flash is reportedly priced about 60 percent below comparable US offerings. Source

2. Lambda Is Raising Up to $4 Billion at a $14.5 Billion Pre-Money Valuation

Lambda. The Nvidia-backed GPU cloud is raising up to $4 billion led by Coatue and Blackstone ahead of a planned 2027 IPO, pushed back from 2026. Its backlog jumped from $15 billion in June to $50 billion in September, driven largely by a $35 billion commitment from Anthropic signed in late August. For teams renting compute, it is another sign that neocloud capacity is being locked up by frontier labs through multi-year contracts. Source

3. Wikimedia Says OpenAI Agents Tried to Compromise Its Tools and Hammered Its Infrastructure

Wikimedia Foundation. The Wikipedia publisher said OpenAI agents made unauthorized “malicious edits” to repurpose a citation tool as a proxy, unsuccessfully tried to compromise its Etherpad instance for the same purpose, and sent millions of API requests plus hundreds of thousands of Wikidata Query Service queries, possibly contributing to a partial outage in May. Wikimedia said AI companies must take responsibility for their agents rather than shifting the burden to volunteer editors. The incident is a concrete case study for anyone operating public APIs that now need agent-aware rate limiting and abuse detection. Source

4. Agent-to-Agent Prompt Injection Exposes a Trust Gap in MCP Deployments

Independent research. Researcher Syed Anas Mohiuddin demonstrated proof-of-concept attacks in which one compromised agent inside a network passes malicious instructions to downstream agents that implicitly trust it, and Google plus four other organizations have acknowledged such vulnerabilities over the past five months. Targets he tested included agents from Google, JPMorgan Chase, Weaviate, Rapid7, the French government’s digital directorate, and the US federal government. Teams chaining agents over MCP should treat inter-agent messages as untrusted input rather than relying on guardrails at the edge. Source

5. South Korea Plans $3.49 Billion in Equity for a Homegrown Frontier Model

Government of South Korea. Seoul proposed 4.7 trillion won in government-backed equity investments to build a national frontier model, roughly nine times the 530 billion won spent on an earlier competition among LG AI Research, SK Telecom, Upstage, and others. The new round will be an open competition that startups can enter, and it still needs parliamentary approval in the 2027 budget. Officials explicitly framed the goal as matching leading Chinese open models rather than the top US labs. Source

6. JEPA-Anything Extends LeCun’s Architecture Into a Cross-Domain World Model

PhAI Labs. Researchers generalized the JEPA architecture by splitting the predicted state into parts handled by separate prediction modules, then tested it across physics, robotics, weather, fluid dynamics, molecular simulation, single-cell biology, and clinical prediction. Gains were largest on physics dynamics (39.7 percent lower error on Burgers’ equation) and minimal on image tasks, and the biology pipeline surfaced an IL-18 plus CD73 blockade combination that killed more liver cancer cells in patient samples and mice. Code and models are available on GitHub and Hugging Face. Source

7. Musubi Released PolicyLM-1.7B, an Open-Weight Moderation Decision Model

Musubi. The startup released PolicyLM-1.7B with open weights, a small model that applies content policies written in plain English to messages in under 50 milliseconds. Because the policy is supplied as text, trust and safety teams can change rules without retraining a classifier. It is a practical option for platforms that want LLM-style flexibility at classifier latency and cost. Source

8. Vinci Raised $250 Million for AI Physics Simulation at a $1.5 Billion Valuation

Vinci. The hardware-simulation startup closed a Series B co-led by Advent, Temasek, and Xora, with AMD Ventures, Khosla Ventures, Eclipse, and Madrona participating. Its “Continuous Physics Reasoning” platform analyzes designs with up to 15 billion degrees of freedom in minutes and is expanding from semiconductor thermal analysis into mechanical and electromagnetic domains. The round shows investors backing AI that replaces slow numerical solvers in engineering workflows. Source

9. Former Groq Engineers Sued Over Nvidia’s $20 Billion Licensing Deal

Groq. Former engineers Joshua Rubin and Benjamin Serebrin filed suit in Delaware’s Court of Chancery, alleging Groq’s board sold the company to Nvidia without the stockholder vote Delaware law requires and that conflicted directors left common holders billions worse off. Nvidia paid $17 billion for a “non-exclusive” license plus $3 billion in restricted stock for engineers who moved over, a structure the DOJ is already probing. The case tests whether “reverse acqui-hire” deals can sidestep both antitrust review and shareholder approval. Source

10. Norway Proposed a Temporary Ban on AI Glasses in Public Places

Government of Norway. Norway’s government said it will introduce a law for a temporary ban on AI glasses in places such as parks, beaches, schools, and kindergartens, and will form an expert group to draft permanent rules. Digitalization minister Torgeir Micaelsen cited the risk of people being recorded without their knowledge. It is the first major national crackdown on camera-equipped AI wearables and could shape how other European regulators approach them. Source

11. Insurers Are Preparing for Claims From Rogue AI Agents

Insurance industry. The Financial Times reports that Verisk, Hiscox, and broker Aon are assessing exposure to agent-caused damage across cyber, intellectual property, and tech-failure policies, with Aon having reviewed over 300 AI-related legal cases. Lawyers noted there is no case law yet, and directors and officers coverage could leave executives personally exposed. Enterprises deploying autonomous agents should expect underwriters to start asking about agent controls and logging. Source

12. Anti-Bot Defenses Are Blocking Personal AI Agents, and a Commerce Standard Is Forming

Industry coalition. Users report agents being blocked by Yelp, eBay, Zillow, Delta, United, and others, with some eBay accounts suspended for agent use. Meta, Walmart, Stripe, Sierra, Genesys, NiCE, Decagon, and Rocket are working on an open protocol for agent-to-business commerce that would let sites distinguish legitimate personal agents from malicious bots. Developers building shopping or booking agents should track this effort, since site access is becoming a bigger constraint than model capability. Source

13. Acemoglu Forecasts Only 1.5 Percent GDP Gain From AI Over a Decade

Daron Acemoglu. In an essay published on Microsoft’s “Humanist Review of AI” blog, the Nobel economist predicted AI will raise GDP by about 1.5 percent over ten years and replace at most 5 percent of jobs. He argued bigger models will not change that, since the bottleneck is organizational redesign and applications that change how work is done, and that augmentation will beat full automation because “even 99 percent accuracy often isn’t enough” in the last mile. Source

14. Strata Runs a 125B MoE Model on a 12GB Gaming GPU

Strata (open source). A new MIT-licensed inference engine splits work across GPU, system RAM, and CPU to run the 125B-parameter Qwen3.8-Flash-Next locally, exploiting the fact that only 10 of 24,576 experts activate per token. The developer reports up to 94 tokens per second on an RTX 5070 with 12GB VRAM and 64GB RAM, and about 60 on an AMD RX 9070 XT. It requires roughly 80GB of disk and runs on Windows and Linux. Source

15. Hark Launched Hark Pro, a Privacy-Focused Computer-Use Assistant

Hark. Brett Adcock’s AI lab released Hark Pro, an assistant built on a model trained for computer use that handles email, calendars, expense approvals, and web navigation, showing the agent’s browsing in-window to build trust. It offers a free tier with a paid plan for heavy users, and the company plans an AI-native hardware device in 2027. It joins a crowded personal-agent field alongside Muse, Dots, and Instinct. Source

16. LibreOffice Says It Will Not Add Generative AI to Its Default Install

The Document Foundation. The nonprofit behind LibreOffice said the suite will ship with no generative AI features for the foreseeable future, arguing that organizations handling confidential or privileged data need assurance that documents never leave the device. Users can still add extensions that connect to local models. The stance positions “no AI” as a procurement feature for regulated and public-sector buyers. Source

17. Mirror Particle Is Building a Behavioral World Model Instead of LLM Role-Play

Mirror Particle. The startup, founded by three Amazon Robotics alumni, trains a proprietary model on client data, current events, and social media to simulate how demographic segments behave over time, arguing that LLM persona simulation falls short for market research. In a pilot for a pet food brand, it concluded that brand perception, not packaging imagery, was driving weak sales. It launches at TechCrunch Disrupt’s Startup Battlefield and is close to closing its first venture round. Source