AI News Daily

New Google and RadixArk models, NVIDIA’s scientific AI resources, Microsoft’s security agents, an EU attack report and Palo Alto Networks’ proposed acquisition.

Show Notes

New Google and RadixArk models, NVIDIA’s scientific AI resources, Microsoft’s security agents, an EU attack report and Palo Alto Networks’ proposed acquisition.

Stories and sources

Google begins Gemini 4 Neon rollout

Hosted access expands, with separate restricted cyber access planned.

RadixArk releases DaisyRose-27B

Downloadable weights accompany training details and a benchmark-version qualification.

NVIDIA introduces Nemotron for Science

Chemistry data and adaptation recipes accompany a reported simulation-repair experiment.

Microsoft expands Perception security agents

General availability adds investigation inputs and reviewed detection workflows.

EU confirms AI-assisted attack

AP reports official confirmation.

Palo Alto Networks agrees to acquire Applied Compute

The $4.5 billion all-stock deal remains conditional.

This episode uses an AI-generated narrator.

What is AI News Daily?

A concise English-language briefing on artificial intelligence: new models and research, AI security, documented threat-actor activity, and the business of AI. Approximately five minutes daily, with an approximately twenty-minute Sunday review of the week's most important substantive developments. Produced by ResearchNode.dev using AI-generated research and narration, with source checks before release and links to original reporting and research in the episode notes.

Google has begun rolling out Gemini 4 Neon, while RadixArk has released downloadable DaisyRose model weights. Microsoft has expanded its security agents, and Palo Alto Networks has announced a deal for an AI training platform. This is AI News Daily for November twentieth, twenty twenty-six.

Google announced Gemini 4 Neon on November eighteenth. Access is rolling out through the paid Gemini API, Gemini for business, and Pro and Ultra subscriptions in the Gemini app and Antigravity. Google lists a million-token context window and support for combining documents, charts, audio and video. Its announced price is three dollars per million input tokens and fifteen dollars per million output tokens. A separate cyber version has a narrower route: Google said eligible Fairwind defenders would begin receiving access over the following days. That announcement does not establish that every eligible defender already has access.

RadixArk released DaisyRose-27B on November eighteenth, with weights under Apache 2.0. Built on Qwen 3.8-27B, the model retains image understanding and a context window of about two hundred and fifty-six thousand tokens. RadixArk says training used about twelve hundred curated tasks over eighteen days on one hundred and forty-four H200 GPUs. Sessions could span a million tokens by summarizing their history across multiple trajectories; that does not mean the model has a million-token context window.

There is also a qualification in the benchmark evidence. RadixArk reports twelve percent on Terminal-Bench 4.1, averaged over eight repetitions using its internal agent setup with summarization. Its comparison table lists the base model at five point six percent, but the model card identifies that figure as a Terminal-Bench 4.0 result. Those different benchmark versions limit the comparison. The reported DaisyRose scores also select the higher result from two numerical formats.

NVIDIA introduced Nemotron for Science on November sixteenth. The package combines its Nemotron 3.5 Super VL model with scientific evaluations, chemistry data and recipes for adapting the model to specialist work. NVIDIA says the model and LitmusBench version zero point two are available on Hugging Face. The dataset contains nearly fifty thousand questions for training and evaluating chemistry reasoning, drawing on molecular properties and experimental drug-discovery data.

The company also describes an agent from KTH Royal Institute of Technology that repairs broken GROMACS molecular-simulation setups. NVIDIA reports that, in one overnight run on a single DGX Spark, the agent fixed eighty-one percent of test cases with reasoning enabled. GROMACS itself checked each fix. This is a reported result from that test run, rather than a success rate established for every simulation workflow.

Microsoft announced general availability of Perception on November seventeenth. Its agents coordinate security investigation and remediation, with people approving critical actions. Microsoft says its Triage Agent covers about eighty percent of the alert volume in an analysis of roughly eight million Microsoft Defender alerts. That measures coverage, rather than a guaranteed triage rate for each customer.

The update also lets the Attack Investigation Agent use Okta and Amazon Web Services signals. Microsoft's Detection Authoring Agent can turn external threat intelligence into detection recommendations, test them against historical customer data, and leave analysts to review and adjust them before deployment.

An EU official has confirmed an AI-assisted cyberattack. The Associated Press reported on November nineteenth that the October attack affected the European Commission. The spokesperson did not identify an attacker. He said most applications taken offline had been restored, and no evidence had been detected of a breach affecting commission staff as a whole.

Palo Alto Networks has agreed to acquire Applied Compute. Its November eighteenth announcement values the all-stock transaction at four point five billion dollars, subject to adjustments. Applied Compute provides training and inference for open-weight AI models. Palo Alto Networks says the acquisition would support enterprise model customization and deployment in public-cloud or sovereign environments. Those are plans for the combined business. The transaction still requires regulatory clearance and other customary conditions, with closing expected in the company's fiscal third quarter of twenty twenty-seven.

Today's releases bring new hosted and downloadable model options, alongside scientific training resources. Microsoft's security update retains human review, while the Applied Compute acquisition remains pending. You've been listening to AI News Daily.