Daily News · 11 min read

AI News: October 20, 2026

security ai-agents policy funding hardware inference copyright ai-safety evals open-weight video-generation agentic-commerce compute deepseek openai anthropic
Listen

1. One Hacker Used AI Agents to Breach Websites of More Than 100 Companies for About $5,000

NorForge / Reuters. A Reuters investigation published October 19 found that a single Chinese-speaking operator delegated most of a hacking campaign to AI agents, which attempted to break into more than 1,200 sites (mainly e-commerce sites in the US, Canada, Australia, and Japan) starting September 25 and succeeded at more than 100 companies. NorForge researcher Snorre Fagerland found the operation through exposed deployments of Cairn, an open-source agent platform, and estimated about $5,000 in token spend, or roughly $50 per successful intrusion, with a typical break-in taking about 23 hours from first scan to access; the tooling reportedly included DeepSeek models and Claude Code. Deere, SharkNinja, and Cummins were named among the targets, and the extent of data theft is still unknown. Source

2. DHS Orders DeepSeek Removed From Federal Civilian IT Systems

US Department of Homeland Security. DHS is barring civilian federal agencies from buying or using IT products and services from Hangzhou DeepSeek Artificial Intelligence and its affiliates, using the Federal Acquisition Supply Chain Security Act after a Federal Acquisition Security Council review; the notice was set for publication in the Federal Register on October 20. Agencies must exclude DeepSeek from future procurements and remove deployed versions, and exceptions require a request 60 days in advance with a “compelling justification” such as a mission-critical need. Secretary Markwayne Mullin cited China-based data storage, user data collection, and the risk that prompts and responses could reach the Chinese government; the order does not cover state governments or private companies. Source

3. OpenAI Reportedly Seeks About $30 Billion at Over $1.4 Trillion Before Anthropic’s IPO

OpenAI (report). Reuters reports that OpenAI is trying to accelerate a roughly $30 billion funding round that could value it above $1.4 trillion, with SoftBank Group, MGX, and other existing backers in talks, and with meetings held with long-only and sovereign wealth funds while Anthropic hosted its own investor day. Sources say OpenAI told investors its annualized revenue reached $50 billion and could hit $70 billion by year-end, and the timing is partly meant to close before Anthropic’s expected mid-November public debut makes its financials comparable. The report says Anthropic’s quarterly revenue rose to $16.8 billion in Q3 from $11.6 billion in Q2. Source

4. Nscale and Anthropic Reportedly Near a Second $45 Billion Compute Deal

Nscale / Anthropic (report). Bloomberg reports that Nscale and Anthropic are close to announcing another compute agreement worth about $45 billion, a little over a month after the two announced a deal of the same size; sources say terms could still change, and neither company commented. The talks come as both companies pursue IPOs and as some investors reportedly push back on Nscale’s proposed $30 billion valuation. Source

5. Executives and Analysts Forecast a Neocloud Shakeout, With Nvidia Most Exposed

Industry analysis. Citing The Information and Bloomberg, The Decoder reports that executives expect only seven to ten of the hundreds of new GPU cloud providers to survive, Gartner predicts 80% of independent neoclouds will cease independent operations by 2030, and Bernstein estimates only 35% of announced North American data center capacity will get built. CoreWeave lost about a quarter of its market value in a week, Microsoft pulled out of a planned UK lease of 23,000 Nvidia chips from Nscale, and Firmus withdrew its IPO. Neoclouds and other non-hyperscaler customers accounted for $40 billion, or 45%, of Nvidia’s data center revenue in the July quarter, and Nvidia has revised a loan guarantee offer after larger neoclouds rejected its first terms. Source

6. Tenstorrent Raises $1.5 Billion at a $6.2 Billion Valuation for Inference Chips

Tenstorrent. The Jim Keller-led chip startup raised a $1.5 billion round led by Trace Capital and Innovation X, with Cadence Ventures, Eclipse, Fidelity, SanDisk, and T. Rowe Price participating, more than doubling its $2.6 billion valuation from late 2024. Keller said the raise followed its Galaxy servers demonstrating they could run DeepSeek models, and the money will go toward scaling chips for serving open-weight models and video generation. Tenstorrent reportedly turned down acquisition interest from Intel and Qualcomm at valuations of up to $5 billion about five months ago. Source

7. Town Raises $108 Million at a Unicorn Valuation Four Months After Launch

Town. The San Francisco startup, founded by former Plaid CTO Jean-Denis Greze, raised a $108 million Series B co-led by Forerunner Ventures and Menlo Ventures, with Andreessen Horowitz, Serena Ventures, and First Round participating, bringing total funding to at least $160 million and its valuation above $1 billion. Town gives each employee an assistant (“Townie”) connected to email, calendars, and other tools; Townies can share approved information with each other and now with people outside the company. Town says it reached millions of dollars in annual recurring revenue within weeks of its public launch, and it also announced a shared workflow that pushes new business information to every employee’s assistant. Source

Postmedia, Torstar, The Globe and Mail, The Canadian Press. The publishers filed a statement of claim on October 6 in an Ontario court, reported October 19, alleging Anthropic copied their content without a license to train Claude and continues to scrape it in response to user prompts. The suit also alleges breach of website terms of service and circumvention of technical protections against copying, and says Anthropic raised significant capital “all without obtaining a valid licence from any of the News Media Companies.” The allegations have not been proven, and Anthropic had not responded to requests for comment. Source

9. METR Gets an Embedded Team and “Employee-Like” Access Inside Anthropic

METR / Anthropic. Anthropic has given the nonprofit evaluator METR “employee-like access” to assess model risks, with a METR team now embedded in Anthropic’s office with access to its hardware and internal systems, according to Ars Technica. METR has previously evaluated models for OpenAI, Anthropic, Google, and Meta, and greater third-party evaluator access is one of the main concrete steps labs have proposed in response to concerns about catastrophic risks from more capable systems. Source

10. Hugging Face’s Policy Chief Says Embedded Evaluators Won’t Be an Easy Fix

Hugging Face. Chief policy officer Irene Solaiman said at TechCrunch Disrupt that AI benchmarking is still immature, results are hard to reproduce and compare, and evaluations “start to get ineffective relatively quickly,” so evaluator qualifications need much more standardization before embedded third-party observers can be relied on. She argued that reproducibility depends on access to the model itself, which open weights provide. The article notes that agentic and harness evaluations are especially costly, and that Anthropic’s choice of Accenture as its first embedded evaluator drew criticism. Source

11. MiniMax M3.1 Flash Scores 42 on the AA Index at $0.15 per Million Input Tokens

MiniMax. The new reasoning model takes text and images, supports up to 1 million tokens of context, and costs $0.15 input, $0.60 output, and $0.015 cached input per million tokens, about $0.22 per task on the Artificial Analysis suite; MiniMax says it runs about three times faster than M3. It scores 42 on the Artificial Analysis Intelligence Index (up from 29 for M3), behind DeepSeek V4.1 Flash 1017 (45), and leads DeepSeek on Humanity’s Last Exam (48% vs. 43%) and AA-Omniscience (56% vs. 43%), but its hallucination rate is 93% because it rarely declines to answer. It is available via API and the MiniMax Code app, with open weights planned for October 23. Source

12. Kling AI 4.0 Generates 30-Second Videos With Stereo Audio and Multilingual Lip Sync

Kling AI. Version 4.0 generates up to 30 seconds per pass, accepts up to ten keyframe images and 8,000-token prompts, and supports up to 15 reference assets (images, videos, storyboards, depth maps) per generation, plus a Creative Recreation mode that uses an existing video as a template. It adds two-channel audio and lip sync in more than ten languages, and is accessible via MCP and CLI for agent workflows; a Flash variant is capped at 720p and 20 seconds. 4K HDR output, two-minute extension, and API access are listed as coming soon, and API pricing has not been announced. Source

Stripe. Citing Upstarts Media, The Decoder reports that agent transactions through Stripe’s Link wallet grew 38-fold and payment volume 20-fold in September, though Stripe would not disclose absolute figures and the growth starts from a small base. About 200 agents now use Link (up from about 70), including Meta’s Muse and Grok Bot, with most spending around $100 per transaction; retail accounted for 36% of transactions and travel 28% of volume. Link issues a separate 16-digit card number per agent and supports user-set spending limits. Source

14. Laplace Intelligence Red-Teams Autonomous Vehicles and Robots in Simulation

Laplace Intelligence. The University of Michigan spinout, founded by Howie Sun and Henry Liu, rebuilds high-fidelity simulations from real map and traffic data and uses AI agents to generate rare corner-case scenarios that expose failures before deployment; its simulator uses a post-trained version of Nvidia’s Cosmos. Customers include autonomous trucking firm Bot Auto and a major Japanese automaker, plus a free smart-city traffic pilot in San Jose with Nvidia and T-Mobile. The company says it is profitable and plans to raise funds; its main rival, Applied Intuition, is valued at $15 billion. Source

15. Mayor Mamdani Says Washington Has Abdicated on AI and Urges New Yorkers to Push Back

New York City. Speaking at the AI Now Institute’s People’s AI Assembly, Mayor Zohran Mamdani criticized the federal government’s “complete abdication” on AI regulation, pointing to the administration’s Superintelligence Task Force as largely enabling industry self-regulation, and said rapid AI rollout is not inevitable. He cited his September order banning AI use in elementary and middle schools and a summer proposal requiring disclosure of AI-generated real estate listing photos and videos. He called for a different federal framework under a future administration. Source

16. Court Keeps Accused OpenAI Spy Detained After FBI Cites “Saudi Gov Docs” Folder

US v. Alfozan. A court on October 19 granted the government’s request to keep former OpenAI engineer Abdulrahman Alfozan detained; he is accused of bugging a top-secret OpenAI conference room to steal the lab’s secrets. Prosecutors argued he may have acted for personal business gain or to benefit Saudi Arabia, citing WhatsApp chats with a member of the Saudi royal family and plans to move back, and the FBI said he accessed a “Saudi Gov Docs” folder on his iPhone where he allegedly also stored photos of a second planted bug. Source

17. Indie Developers Hold Back Demos Over Fears of AI-Assisted Clones

Games industry. Developer Alliance Arts delayed a public demo of its game Drunken Goddess Reflux planned for Steam Next Fest, saying the demo’s content could be used to “quickly imitate or replicate the game’s presentation and gameplay with AI or other tools” before launch. Ars Technica reports that AI coding tools are making cloning efficient enough that many developers are becoming cautious about sharing early builds publicly. Source

18. Jack Clark Tells IMF Audience That “the Singularity Is Here”

Import AI. Import AI 477 publishes the Per Jacobsson lecture that Anthropic co-founder Jack Clark gave at the IMF meeting in Bangkok on October 17, arguing that 2026 brought convincing evidence that AI-driven acceleration of science has begun and that economists and policymakers should expect rapid change. He said Anthropic researchers now produce almost 10 times as much code as in the company’s first four years, and that coding gains have driven parallel advances in cyber offense and defense. The speech focuses on how countries can share in the benefits, with risks to be covered in later issues. Source