Newsletter

Google releases Gemini 4 Argon to trusted cyber defenders

The frontier model raises its output limit to one million tokens and rolls out first through the Fairwind Program, as Washington commits $5.25 billion to grid upgrades for AI data centers.

On this page

Google has released Gemini 4 Argon, which it calls its most powerful model yet. The model is built for coding, legal and finance research, and defensive cybersecurity. It is rolling out only to trusted cyber defenders through the Fairwind Program. Google says it can find, validate, and patch critical software flaws on its own.

The company reports an output limit of one million tokens, up from 64,000. It also claims Argon leads rival models on the Vals AI index. That comparison is company-cited and was not independently verified. Introductory pricing is set at $2 per million input tokens and $10 per million output tokens.

Featured source: TechCrunch , Google .

Washington commits $5.25 billion to grid upgrades for AI load

The US Department of Energy has earmarked funding for 31 grid projects across 26 states to meet rising power demand from AI data centers. The package totals $5.25 billion including cost-share, with nearly $2 billion in federal money. The announcement does not detail how household energy bills will be affected.

The funding split matters because utilities and developers are still debating who pays for AI-driven load growth. Watch the department’s full release for project-level terms and rate effects.

Filed from: TechRadar .

Reddit ends RSS feeds and public API access over scraping

Reddit will end RSS support on November 13 and shut its public API by March 2027. The company says RSS has become a common surface for large-scale scraping and automated abuse. It points moderators to a Discord relay app and offers no replacement for external RSS uses.

The move follows Reddit’s push to license its corpus for AI training. Non-advertising revenue reached $43 million in the second quarter, up 24 percent year over year. Researchers and small tools without commercial deals lose programmatic access after the cutoffs.

Filed from: TechCrunch .

Teenager reached admin functions on Microsoft analytics service

A 16-year-old researcher known as Faav found an authentication flaw in Microsoft’s Titan analytics service. An unsigned login token let him reach admin functions and run unauthorized queries. He counts about 17.3 trillion rows in the reachable system, a figure from his own measurement.

Microsoft says it has locked down the API path and paid a $5,000 bounty. The case turns on a missing signature check on an Azure-hosted path. There is no report of customer credential theft.

Filed from: The Register .

White House signs voluntary AI safety accord with two dozen firms

Two dozen AI companies agreed to White House-recommended controls on cybersecurity, biosecurity, chemical threats, and unintended model actions. Named signers include Anthropic, OpenAI, xAI, Nvidia, Meta, and Alphabet. The controls call for independent safety audits of internal monitoring.

The accord carries no legal weight. President Trump called it morally binding. It remains unclear who will conduct the audits or to what standard. Several firms had already promised external audits and pre-deployment government testing.

Filed from: Ars Technica .

Cloudflare Workers gains post-quantum signing and key exchange

Cloudflare Workers now offers ML-KEM and ML-DSA through Web Crypto. The APIs sit behind the webcrypto_modern_algorithms compatibility flag while the specification keeps moving. Cloudflare describes them as building blocks for testing rather than a complete migration path.

Developers get key encapsulation with ML-KEM-768 plus signing with ML-DSA-44. Protocol design, auditing, and deployment choices remain with implementers. The release helps teams probe post-quantum integrations without bundling separate crypto code.

Filed from: Cloudflare .

Brain-scan model reconstructs what volunteers were viewing

Researchers at the Weizmann Institute of Science report rebuilding viewed images from fMRI scans with close fidelity. The team used higher-resolution scanners and a model that also predicts brain activity from an image. Leader Michal Irani hopes the work could someday help locked-in people communicate.

Outside scientists call the results impressive and warn of misuse. A similar method might expose mental imagery without consent. The findings cover scanned volunteers in a lab, not general mind reading.

Filed from: MIT Technology Review .

Britain’s privacy watchdog becomes a board-led commission

Britain’s data-protection regulator changed legal form on September 30. The Information Commission replaces the Information Commissioner as the statutory regulator. Powers once held by one person now sit with a corporate body. Regulatory functions are unchanged under the Data (Use and Access) Act 2025.

The shift follows the June resignation of John Edwards after a workplace investigation. Paul Arnold serves as interim chief executive. Maggie Carver acts as deputy chair while the search for a permanent chair runs into spring 2027.

Filed from: The Register .

From the Community

OpenAI and Synopsys partner on a chip-design model

Synopsys and OpenAI signed a multi-year agreement to build GPT-Synopsys, a model designed to operate Synopsys chip-design tools. The model would run design workflows, interpret tool output, and iterate on power, performance, and area targets for engineer review. The deal includes revenue sharing and joint sales plans.

Synopsys will license its tools to OpenAI for the work. The companies say customer design data will not train the model. No independent benchmarks or customer results have been published yet.

Filed from: Synopsys announcement .

A GPU text-rendering survey lands as the Slug patent enters the public domain

AlphaPixel has published a detailed comparison of GPU text-rendering methods, from texture atlases through signed distance fields to tessellation renderers like Rive and Slug. The article reports that Eric Lengyel patented Slug in 2019 and dedicated the patent to the public domain on March 17, 2026. That change cleared the way for the company’s C++20 implementation, Slughorn.

Slug computes glyph coverage per pixel after transformation, so text stays sharp at any scale and under perspective. The benchmarks come from the team that built the implementation under test, so the evaluation is not independent.

Filed from: AlphaPixel .

Continue reading

Complete index →