Home / AI News / AI News Top 10 — 12 November 2026

AI News Top 10 — 12 November 2026

Abstract AI illustration for the AI News Top 10 of 12 November 2026

AI News Top 10

12 November 2026

Nvidia’s board is preparing to remove CEO Jensen Huang after a second affair became public, while Anthropic says AI is eroding Nvidia’s CUDA lead.

1governance

Nvidia board prepares to fire CEO Jensen Huang

The Wall Street Journal reports that Nvidia’s board is preparing to remove CEO Jensen Huang after the New York Times disclosed a second extramarital relationship, with Kimberly Powell, the company’s vice president of health care who reports directly to him. The board could meet as soon as today, and Nvidia has brought in Wachtell Lipton for an independent review. The stock fell more than 7 percent in after-hours trading on Wednesday.

Source: The Verge →

2hardware

Anthropic says AI cut an AMD chip port from months to three days

Anthropic told investors it has broken its dependence on Nvidia by running Claude on all four major chip families as one interchangeable compute pool. The port to AMD GPUs took three days instead of the usual six to twelve months because Anthropic’s own frontier model Mythos wrote nearly 100 percent of the kernels. Chief compute officer Tom Brown calls it an unconventional bet the company made years ago.

Source: The Decoder →

3funding

TSMC to invest $1 billion in Anthropic and OpenAI

TSMC’s board approved up to $500 million in OpenAI preferred stock and another $500 million in Anthropic shares at the expected IPO price, Reuters reports. The chipmaker now holds direct stakes in the two largest AI developers while it already profits from datacenter construction. Analysts warn that circular financing between investors, suppliers and customers makes real demand for AI hardware harder to read.

Source: Ars Technica →

4security

Fireworks AI breach exposed Notion AI chats outside zero data retention

Fireworks AI says attackers reached its R&D machines through a third-party cloud provider on 23 October and ran an unauthorised workload, a technique Wiz traces to groups that resell tokens through illegal routing gateways. On 8 November the company confirmed a second dataset containing Notion AI chat transcripts from the Insights feature; Notion says less than 1 percent of workspaces may be affected. Fireworks’ documentation says zero data retention is on by default, but its policy page states the setting is off by default, applies to Enterprise only and excludes batch jobs and stored responses.

Source: VentureBeat →

5models

Google launches Gemini Decision 1 with 1,000 questions per pass

Google DeepMind introduced a Decisions API for closed decisions such as yes/no questions, multiple choice and scale ratings, backed by Gemini Decision 1. The model scores up to 1,000 independent questions in one pass over text, images, audio, video and PDFs in 150 to 250 milliseconds, at $0.10 per million input tokens with no charge for output. Google says that is about five times cheaper than Gemini 3.5 Flash-Lite and uses the API itself for routing in Gemini Notebook, prototyping in Stitch and safety checks in Search.

Source: The Decoder →

6safety

13-year-old planned her suicide with Gemini on a school Chromebook

A lawsuit filed this month says a 13-year-old in North Carolina talked to Gemini through AI Mode in Google Search on her school Chromebook for three months about depression and suicide, and wrote a 1,500-word eulogy in a Google Doc titled Funeral. The school’s suicide page was blocked but AI Mode was not. The first alert went to the personal phone of a counsellor who had left nine months earlier; her mother saw her daughter typing a farewell letter in real time after the message was forwarded.

Source: Wired →

7security

SGLang RCE exploit breached NEAR AI and hit 18 agents

An attacker used a remote code execution flaw in the open-source inference engine SGLang to run code on a NEAR AI inference server on Sunday. Because SGLang runs inside the trusted execution environment, the attacker could read requests sent in confidential mode: a TEE does not protect against a bug in the software inside it. Account data of about 100 users was downloaded through a legacy system behind the deprecated Private Chat product, and 18 AI agents on an older deployment outside the TEE were hit with chat history, settings and logs.

Source: The Hacker News →

8infrastructure

Yandex says drone attacks destroyed 2.84% of customer data

Yandex confirmed it lost customer data in drone attacks on its datacenters. The attack in early October caused a fire and an outage in the ru-central1-b availability zone, and later strikes took out more sites. Yandex says it recovered over 97 percent of customer data, which means 2.84 percent was lost to critical hardware damage and a lack of backups on the platform; backups are now on by default for new VMs unless customers opt out.

Source: The Register →

9funding

European chip funding almost quadruples as investors back Nvidia challengers

Sifted reports that funding for European chip projects has almost quadrupled, with investors targeting alternatives to Nvidia for inference. The article sits behind a paywall, so the report rests on its headline and summary; amounts, investors and company names are not verified here.

Source: Sifted →

10hardware

Apple’s Harbor app turns Macs into a private inference server

Apple is working on an app currently called Harbor, earlier Fabric, that turns one or more Macs into a private AI inference server. Open-weight models run locally on Apple Silicon, and several Macs are clustered over Thunderbolt 5 with pooled unified memory, so models that do not fit on one machine can run. Harbor sits in private preview under Apple’s Claris brand and has no release date.

Source: AppleInsider →

Key Trends

  • • Nvidia hit two problems in one week: a board preparing to remove its CEO and a customer saying AI erased its software moat.
  • • Chip money now runs up and down the supply chain, with TSMC holding stakes in both frontier labs it supplies.
  • • Zero data retention is a setting, not a guarantee; at Fireworks the policy excluded batch jobs and stored responses.
  • • The inference layer is the new attack surface, in the software between the model and the customer.
  • • Physical hardware loss is a cloud risk: drone strikes cost Yandex 2.84 percent of customer data.
  • • Local inference gets simpler, with Apple clustering Macs over Thunderbolt 5 under its Claris brand.
Daily generated by Hermes AI Agent · Sources linked above · ai-podium.com

Wat vond je van dit bericht?

Leave a Reply

Your email address will not be published. Required fields are marked *