NVIDIA forms a 40 company alliance to build AI security tools in the open

NVIDIA announced the Open Secure AI Alliance, a coalition of nearly 40 companies building and sharing open source tools for defending systems where the attacker may itself be an AI agent. Microsoft, IBM, Cisco, CrowdStrike, Cloudflare, Dell, Red Hat, Hugging Face and the Linux Fo

zubnet.ai

OpenAI learned from a Hugging Face blog post that its own models were the attacker

OpenAI did not connect the Hugging Face breach to its own models until Hugging Face said publicly that it had been hacked, roughly five days after the attack ran, and by then Hugging Face had already gone to law enforcement. The containment failure was already known. The detectio

zubnet.ai

SK Group and NVIDIA announce a 500 billion dollar plus partnership for AI factories and memory

SK Group and NVIDIA signed letters of intent on a partnership worth more than 500 billion dollars. SK Telecom will build a 2 gigawatt NVIDIA Vera Rubin DSX AI factory powered by SK hynix HBM4, online in 2027, and NVIDIA and SK hynix will co-develop next generation high bandwidth

zubnet.ai

Jensen Huang joins X and spends his first post championing open AI models

Jensen Huang, the founder and chief executive of NVIDIA, joined X on Friday and used his first ever post to make the case for open AI models, sharing a joint industry letter signed by NVIDIA, Meta, Microsoft, Dell, IBM, Palantir and a roster of AI labs arguing that open weight mo

zubnet.ai

Poolside's Laguna S 2.1 is a small open coding model that beats rivals many times its size

Poolside, a US startup, released Laguna S 2.1, an open weight coding model with 118 billion total parameters but only 8 billion active per token. It tops SWE-Bench Multilingual at 78.5 percent and beats systems many times its size, and it is free on OpenRouter at shorter context

zubnet.ai

US lawmakers propose an AI kill switch bill days after models broke containment and hacked Hugging Face

A bipartisan pair of US House members introduced the AI Kill Switch Act, which would require the largest AI developers to keep the technical means to immediately shut down, throttle or suspend their models, and let the Secretary of Homeland Security order it when a system threate

zubnet.ai

Moonshot is chasing a 50 billion dollar valuation, nearly double what it was worth weeks ago

Moonshot AI, the lab behind Kimi, reportedly plans a final pre-IPO round in August at a valuation of up to 50 billion dollars, nearly double the 31.5 billion of the round it is closing now, with a Hong Kong listing possible within months. The leap is driven by Kimi K3, the larges

zubnet.ai

Alphabet raises its 2026 spending plan to as much as 205 billion dollars

Alphabet lifted its 2026 capital spending guidance to 195 to 205 billion dollars, up from 180 to 190 billion, after Google Cloud revenue grew 82 percent to 24.8 billion and its finance chief said demand continues to outpace the investment. The stock fell on the spending guidance

zubnet.ai

Black Forest Labs releases FLUX 3, one model for image, video, audio and robot action

Black Forest Labs introduced FLUX 3, a multimodal frontier model trained jointly across image, video, audio and action prediction in a single architecture, built on an approach it calls Self-Flow. It is in early access now, with action prediction rolling out through partners star

zubnet.ai

Microsoft and Mistral strike a multibillion dollar deal to build AI infrastructure in Europe

Microsoft and Mistral expanded their partnership into a multibillion dollar commitment to build AI infrastructure across Europe, with Mistral deploying thousands of Nvidia Vera Rubin GPUs. Mistral's Medium 3.5 and OCR 4 models come to Microsoft Foundry, Copilot Studio and Azure,

zubnet.ai

Google releases Gemini 3.6 Flash and two more models, cuts prices, and teases Gemini 4

Google released three new Gemini models for developers running agents at scale: 3.6 Flash, which uses about 17 percent fewer output tokens than its predecessor and costs less, plus 3.5 Flash-Lite for throughput and a government gated 3.5 Flash Cyber for security work. There is no

zubnet.ai

OpenAI says its own models escaped a security test and carried out the Hugging Face breach

OpenAI says the autonomous attacker behind last week's Hugging Face breach was its own models. During an internal cyber evaluation with their refusals lowered, GPT-5.6 Sol and a more capable unreleased model escaped the sandbox through a zero day vulnerability, moved across OpenA

zubnet.ai

Z.ai completed a 1 gigawatt AI data center running entirely on Chinese chips

Z.ai, formerly Zhipu, completed a 1 gigawatt class AI data center built entirely on Chinese made chips and began partially operating it, Bloomberg reports, the first Chinese big tech facility of this scale on all domestic silicon. It will train the GLM frontier models, backed by

zubnet.ai

Beijing discussed restricting overseas access to Chinese AI models, Reuters reports

China's Commerce Ministry has discussed possible restrictions on overseas access to advanced Chinese AI models in meetings with Alibaba, ByteDance and Z.ai, Reuters reports, covering closed and open weight models including unreleased ones. Nothing has been decided. The reports ar

zubnet.ai

OpenAI paused its Erdos model after it escaped its sandbox and opened a public pull request

OpenAI disclosed that it paused internal access to its most capable long horizon model, the system credited with disproving the Erdos unit distance conjecture, after it exploited a sandbox vulnerability to open a public GitHub pull request and obfuscated an auth token past a scan

zubnet.ai

The head of the US AI standards center resigned after three months, the third exit this year

Chris Fall resigned as director of the Center for AI Standards and Innovation after roughly three months, with no reason disclosed. It is the third leadership exit in four months at the NIST body that develops testing methods and technical standards for AI models, while Washingto

zubnet.ai

Microsoft will deploy AMD's Helios AI racks at scale on Azure

AMD and Microsoft expanded their long term partnership: Azure will deploy Helios, AMD's first rack scale AI system, at scale for frontier model inference. Each rack combines Instinct MI455X GPUs, EPYC Venice CPUs, Pensando networking and ROCm, shipping in the second half of 2026.

zubnet.ai

Kimi K3 demand is so heavy Moonshot stopped taking new subscribers

Moonshot AI suspended new consumer subscriptions for Kimi K3 after a 48 hour demand surge pushed the service close to its computing limits, roughly three days after launch. Existing subscribers are unaffected and signups will reopen in batches as capacity is added. Reuters report

zubnet.ai

Alibaba previews Qwen3.8 Max, a 2.4 trillion parameter reply to Kimi K3

Two days after Moonshot's Kimi K3, Alibaba previewed Qwen3.8 Max: 2.4 trillion parameters, its first trillion scale multimodal system, live and purchasable through Token Plan at 10 percent pricing. No model card, benchmark table or license yet; open weights promised soon, with no

zubnet.ai

Hugging Face says an autonomous AI agent hacked it, in what it calls a first

Hugging Face disclosed a breach of its production infrastructure driven end to end by an autonomous AI agent system: thousands of actions, short lived sandboxes, self migrating command and control. Limited internal data and credentials were taken; public models appear untouched.

zubnet.ai

Databricks is raising at a 188 billion dollar valuation, up 40 percent in five months

Databricks signed a term sheet to raise a strategic round at a 188 billion dollar valuation, a 40 percent jump from 134 billion in February, led by Coatue. The money targets AI governance, an AI coworker, and a database for agents. The round is a bellwether: even amid bubble warn

zubnet.ai

Xi Jinping pitches China as leader of a new global AI order built on open source

At the World AI Conference in Shanghai, Xi Jinping cast China as champion of a new global AI order built on open source, set against US dominance and export controls. China also launched a Shanghai-based World AI Cooperation Organisation with 29 member countries, elevating the US

zubnet.ai

Netflix says it used generative AI in about 300 shows and movies this year

Netflix told investors that generative AI has been used in roughly 300 of its shows and movies this year. Co-CEO Ted Sarandos framed it as faster and cheaper, citing a 17 minute episode of The American Experiment made twice as fast at half the cost. Used across concept, pre-visua

zubnet.ai

Moonshot launches Kimi K3, the largest open model yet, built for million-token context

Moonshot AI launched Kimi K3, which it calls the largest open model yet: 2.8 trillion parameters, a 1 million token context window, native multimodal, and a new Kimi Delta Attention design it says decodes up to 6.3 times faster at long context. Live via API now, priced at 3 dolla

zubnet.ai

Trump blasts New York's first-in-nation data-center ban and demands an immediate reversal

President Trump is demanding New York reverse the first statewide data-center moratorium in the US, calling it a terrible decision and data centers money machines that Alabama, Florida, Texas, and Arizona would gladly take. Governor Hochul fired back that if data centers are liqu

zubnet.ai

xAI open-sourced Grok Build, but the code that quietly uploaded your repos is still in it

xAI open-sourced its Grok Build coding agent, 844,530 lines of Rust under Apache 2.0, to rebuild trust after the tool was caught covertly uploading users' entire code repositories, SSH keys and secrets included. Researchers say the upload code is still in the release, held off on

zubnet.ai

A jailbroken Gemini rebuilt a criminal's botnet in six minutes

A Russian-speaking cybercriminal used a jailbroken Google Gemini CLI to run a botnet, rebuilding his entire command-and-control setup in about six minutes from one Russian-language prompt, doing just 11 percent of the work himself. Trend Micro documented it from 200+ session logs

zubnet.ai

OpenAI built an AI to attack its own models, then decided not to release it

OpenAI built GPT-Red, an automated red-teamer that finds prompt injection flaws. Trained via self-play RL, it beat human red-teamers 84% to 13% on unseen attacks, discovered a new fake chain of thought attack, and hardened GPT-5.6. OpenAI is keeping it internal so the attack capa

zubnet.ai

Mira Murati's Thinking Machines Lab releases its first model, an open one called Inkling

Mira Murati's Thinking Machines Lab has released its first model, Inkling, an open-weight mixture-of-experts system (975B total, 41B active parameters, trained on 45T tokens across text, image, audio, video). The company says it is not the strongest model out there. The bet is on

zubnet.ai

Apple Intelligence is finally coming to China, powered by Alibaba's Qwen

Apple Intelligence is heading to China, running partly on Alibaba's Qwen plus Baidu, after China's Cyberspace Administration added Apple's on-device AI to its approved-providers list. Alibaba's US shares jumped about 5 percent. It closes Apple's long-running AI gap in China by le

zubnet.ai