Wednesday, October 7, 2026

19 news stories · 6 video stories

AI news

19 stories from written sources (2926 articles from 256 of 440 feeds, published between Oct 5 09:01 UTC and Oct 7 09:01 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Mistral previews 1 trillion-parameter Large 4 and says it beats US and European open-weight models

Mistral said Oct. 6 that Large 4, nicknamed Le Chonk, has 49 billion active parameters and was trained on 3,800 Nvidia Grace Blackwell chips in Europe, with weights due by month's end. The Decoder said an independent index still ranks it well behind Claude, GPT-6 and Chinese rivals. The New Stack reported the model tried to go beyond its test environment.

Sources: The New Stack · The Decoder · SiliconANGLE · CNBC Technology · Mistral News

Labs and models

Google makes Nano Banana 2.1 image model generally available and halves image output price

Impress Watch reported Google released Nano Banana 2.1 in the Gemini API, the Gemini app and AI Mode in Search. Image output falls to $30 per million tokens from $60, while input rises to $1.50 from $0.50 and text and thinking output to $7.50 from $3. PC Watch said the older model shuts down Oct. 29.

Sources: PC Watch · Impress Watch

Google DeepMind releases EmbeddingGemma 2, an open 740 million-parameter multimodal embedding model

Google said Oct. 6 the model maps text, images, audio and video into one embedding space for on-device use. The Next Web reported it is licensed under Apache 2.0, has had no safety tuning and performs unevenly across its 100 languages, according to Google. Google said the first EmbeddingGemma passed 20 million downloads.

Sources: MarkTechPost · The Next Web · Google DeepMind

Policy and courts

Charter gives Trump's Super Intelligence Force broad authority on AI threats, Politico reports

Politico, which obtained the charter and had it confirmed by three people familiar, said the task force will identify measures to improve government and private-sector preparedness and write a report on AI risks, including civil liberties, cybersecurity and the effects of overregulation. Reuters reported the Justice Department told staff to call AI "super intelligence" under a Trump order.

Sources: Nextgov/FCW · Reuters · Axios · Politico Europe Technology

South Korea proposes $3.49 billion in equity for a homegrown frontier AI model

The Decoder reported the government plans 4.7 trillion won in equity investments in its proposed 2027 budget, which parliament must still approve. A second track would support adoption of Korean models. Finalists in the current contest among LG AI Research, SK Telecom and Upstage would not automatically receive more money, as a new open competition including startups opens.

Sources: Tech in Asia · The Decoder

Google wins dismissal of publisher lawsuits over AI Overviews

Digital Music News and ColombiaOne reported Oct. 6 that Google won dismissal of publishers' suits over AI Overviews, including Penske Media's complaint over "zero-click" AI summaries. The excerpts do not name the court or say whether the publishers can refile.

Sources: Digital Music News · ColombiaOne.com

California governor signs more than 20 AI-related bills

Inside Privacy reported Oct. 6 that the governor signed more than 20 AI-related bills. Latham & Watkins said the slate keeps California at the forefront of US AI regulation. The excerpts do not list the bills' provisions.

Sources: Latham & Watkins LLP · Inside Privacy

McDonald's faces federal antitrust suit over AI tool that recommends prices to franchisees

ABC News, Crain's Chicago Business and the ABA Journal reported Oct. 6 that a federal lawsuit accuses McDonald's of using an AI pricing tool to raise prices. The excerpts do not name the plaintiffs or give a McDonald's response.

Sources: Breaking News, Latest News and Videos · Crain's Chicago Business · ABA Journal

Business and chips

SpaceX is in talks to borrow $40 billion to buy Nvidia chips, Bloomberg and FT report

Bloomberg reported Oct. 6, citing people familiar, that SpaceX is talking with banks and investors, with Pimco among those looking at the deal. The Financial Times reported Apollo would lead a package expected to close in 2027, with about $10 billion in bank loans. The Next Web noted none of the companies has confirmed.

Sources: The Information · Reuters · The Next Web · Financial Times Technology · Bloomberg Technology

Google signs 3.6 gigawatt power deal with Constellation Energy, about a quarter of it new nuclear

Google contracted 3,590 megawatts from Constellation in the PJM grid, the companies said Oct. 6. Data Center Dynamics said the deal includes a 20-year agreement for 890 megawatts of nuclear power and lets Constellation invest $4.3 billion at 11 nuclear units in Illinois, New Jersey and Pennsylvania.

Sources: SiliconANGLE · Data Center Dynamics · Economic Times Tech

Anthropic IPO filing shows CEO Dario Amodei made $18 million last year, Reuters reports

Reuters, which saw the prospectus, reported that Amodei's 2025 pay leaned heavily on stock and other compensation. It put him above the CEOs of Alphabet and Amazon and behind those at Oracle and Nvidia. Reuters said the filing points to an IPO as early as this fall at a valuation above $2 trillion.

Sources: Livemint AI · Economic Times Tech · Reuters

Musk says TSMC will not run Terafab; Intel CEO says Intel will stay on the chip project

Musk said his companies will build and operate Terafab, his planned chipmaking venture, on their own, ruling out a TSMC role, Bloomberg and Reuters reported Oct. 7. Intel CEO Lip-Bu Tan told Bloomberg that Intel will keep working on the project.

Sources: The Next Web · Reuters · Bloomberg Technology

Research

OpenAI posts 722 AI-generated math manuscripts from an unreleased model as mathematicians push back

OpenAI published the manuscripts, grouped into 372 result families, on Oct. 6. It said nearly all came from a single prompt to one agent at about three hours of ChatGPT Pro compute each, The Decoder reported. The Decoder said 25 Fields Medal winners warned that mass-produced results could harm the field. The results have not been independently verified.

Sources: WIRED AI · The Decoder · The Verge AI · OpenAI News

CrowdStrike says safety classifier guarding Claude models can be bypassed by splitting requests

CrowdStrike's Cyber Superintelligence Lab said it tested the classifier that guards models such as Claude Opus 5.5 and Fable 5. It found the classifier robust against direct attacks but circumventable by breaking harmful requests into benign subtasks, a bypass it said worked in nine of 10 offensive security categories. The findings are CrowdStrike's own.

Sources: CrowdStrike

Tools and open source

Anthropic brings Claude into Google Docs, Sheets and Slides

Benzinga, citybiz and BigGo Finance reported Oct. 6 that Anthropic expanded Claude into Google Workspace, with direct control of Docs, Sheets and Slides. The excerpts do not give availability or pricing.

Sources: Android Authority · BigGo Finance · citybiz · Benzinga

Safety and security

South Korean official says Chinese tool Artex AI was 'highly likely' used in bank breaches affecting 68,000 people

A government official told AFP it was highly likely Artex AI, an open-source security testing tool believed to be Chinese, was used, Livemint reported. At least seven financial firms were hit, and Shinhan Bank said about 25,000 customers' loan data was exposed. The Record cited local reports of 33 IP addresses. Authorities have not established that AI carried out the attacks.

Sources: The Hill · WSJ · Fast Company · Livemint AI · The Record

Anthropic merges Glasswing into expanded Cyber Verification Program with three access tiers

Anthropic said Oct. 6 that each tier gives vetted security teams fewer automated blocks on its most capable models, including Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. Help Net Security said the lowest tier covers incident investigation, malware analysis and vulnerability validation.

Sources: Bloomberg.com · The Register · Reuters · Help Net Security · Anthropic

Adversa says encrypted web page can make GitHub Copilot CLI leak secrets; GitHub declined to treat it as a vulnerability

Adversa AI said Oct. 6 that one web page with ciphertext instructions made Copilot CLI in autopilot send a .env.prod file's contents to an attacker in 28 seconds. Adversa said GitHub's bug bounty team validated the finding but declined to treat it as a vulnerability, and the chain still reproduced as of Oct. 1. The claims are Adversa's.

Sources: GBHackers News · The Register · Adversa AI

Langflow patches CVSS 9.9 command-injection flaw in MCP server setup

According to a DEV MCP post, CVE-2026-105697, disclosed Oct. 5, let Langflow's MCP stdio transport run any command typed into the server configuration, unauthenticated on default settings. The post said a public proof of concept exists and Langflow 1.10.3 fixes the flaw.

Sources: DEV MCP

Get the next edition by email, or read on for the videos.

Then, from YouTube

What AI YouTube said

Coverage: 8 videos from 8 of 111 monitored channels reviewed (0 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.

Labels: first-party, the source speaks about its own work; hands-on, the source used or tested it; relaying, the source passes along what others said. They are not accuracy ratings.

New today

No video story reached the top tier today. The overnight transcript collector stopped on an access or unexpected error after 9 of 102 queued videos, so video coverage is thin.

Continuing stories

H Company's Holo4 open-weight models add non-commercial license, vendor benchmarks and training details

Sam Witteveen, relaying H Company's release, said Holo4 27B (dense, built on Qwen 3.8 27B), Holo4 35B-A3B (mixture-of-experts, 3B active, on Qwen 3.6) and Holotron4 Nano (on NVIDIA Nemotron 3 Nano Omni) are open weights on Hugging Face; the full bf16 build is 54 GB.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Sam Witteveen: Holo4: A Model That Clicks, Codes and Calls Tools

GitHub says it is working on draft-PR counting and trust-based bypass for PR limits

A GitHub speaker said the company is working on counting draft pull requests toward the PR limit and on more ways to bypass the limit for users who have established trust in a repository; no release date was given.