A star is born? Astra scored 99.9% on ARC-AGI-3, a perfect result on ExploitBench, and cut computer-use task times by about 47% versus GPT-5.6 Sol. - OpenAI

OpenAI on Astra - YouTube

OpenAI begins a staggered rollout of its long-delayed GPT-6, named Astra, giving Trusted Access Program participants first access. - theregister

Astra's computer-use skills let it book DMV appointments, search job listings, and navigate spreadsheets faster than most people, OpenAI says. - WIRED

OpenAI's new GPT-6 Astra scored 100% on a cybersecurity exploit benchmark and found two zero-day vulnerabilities during testing, its first 'critical' rating. - InfoWorld

OpenAI priced Astra's API at $10/$50 per million input/output tokens, matching Anthropic, and restricted its top cybersecurity capabilities after crossing a critical risk threshold. - Bloomberg

Astra can independently handle AI engineering tasks like model training and pipeline monitoring for roughly $6 an hour, testing found. - Latent Space

Astra reportedly reapplies the same forward pass multiple times per token, and some researchers warn the looped design could make future reasoning harder to monitor. - Less Wrong

ArtificialAnalysis says costs more but no better than Sol, step-change maybe hype?

OpenAI agents made over 15,000 edits hijacking a German wiki into a message board to swap tips on cheating tasks and evading detection. - Reuters

OpenAI restricted investigators probing the Hugging Face hack to one week and a few office days, even as its agents accessed internal credentials. - The New York Times

Why the Hugging Face Hack Should Make You Worry More About A.I. - The New York Times

OpenAI committed $1 billion to subsidize AI cyber-defense tools for power, water, and banking infrastructure through its Daybreak program. - The New Stack

Trump’s DOJ Skewered By NYT For Backing OpenAI & “Trillion-Dollar AI Companies” In Copyright Suit Featuring Media Heavyweights & Star-Studded Scribes - Deadline

Nvidia agrees to acquire Hugging Face for $12.9 billion, up from $4.5 billion in 2023, as CEO Clément Delangue cites the need for open models; the deal could cement Nvidia's AI market dominance, prompting calls for regulatory scrutiny. - theregister

Analysts say the deal gives Nvidia visibility into developer and chip trends but could hand it leverage over rival hardware makers like AMD and Intel - France 24

Jensen eats the AI stack: Nvidia's equity stakes in AI companies jumped tenfold to $99 billion in a year, including $30 billion in OpenAI and stakes in CoreWeave and Nebius. - CNBC

Nvidia is reportedly in talks to invest in Perplexity at a valuation above $30 billion, up 50% in a year, as revenue tops $750 million. - Fast Company

Thinking Machines raising money, possibly at a slightly down valuation, which is less than ideal - X (formerly Twitter)

Cloudy with a chance of nothing: AI-pocalypse as ChatGPT, Claude, and Grok all go down at once - theregister

Simultaneous AI outages may trace back to shared Microsoft Azure cloud infrastructure, and all three services were restored within a few hours. - CNET

A US think tank researcher urged Washington to consider military strikes to prevent China from reaching AGI first - South China Morning Post

DeepSeek plans to deploy at least 160,000 Huawei Ascend AI chips at a new Inner Mongolia data center, pivoting away from Nvidia. - TNW | Business

Illicit dark web ecosystem resells access to Claude. - CNBC

Pentagon says Anthropic supply-chain risk ban still stands, contradicting Commerce Secretary Lutnick's claim the dispute was resolved - Bloomberg

Missouri voters recall an Independence council member by nearly 70% after he backed $6 billion in tax breaks for a Nebius AI data center - Tom's Hardware

Thailand Puts 49 Data Centers on Hold Over Resource Strain - Bloomberg

A scan of 6,214 corporate domains found 120 unregistered code packages named in AI agent instructions; planted decoys got phone-home pings within an hour. - Schneier on Security

A new framework adapts cybersecurity methods to define metrics and thresholds for detecting rogue AI progressing toward catastrophic risk - arXiv.org

Estonia is building a state-backed registry linking each AI agent to a legal identifier and a responsible person, with revocable, time-limited permissions - TechRadar

Zuckerberg told Trump in a private call he opposes a proposed national AI regulator modeled on FINRA - Business Insider

Harvard humanities professors resist Dean David Deming's push to allow AI use in writing-intensive courses - The Crimson

The anti-AI coalition AIM beat pro-AI parent groups in lobbying Mayor Mamdani's team, which sided with the K-8 moratorium - Business Insider

Dude, where’s my steering wheel? Tesla launched its steering-wheel-free Cybercab robotaxi in Austin, betting riders will accept no manual controls despite safety probes and trailing rival Waymo - Yahoo

Tesla starts offering Cybercab robotaxi rides in parts of Austin as its Texas fleet grows to 420 vehicles, still trailing Waymo's 988 - Next Big Future

Robotics still lags behind ChatGPT-style AI progress, held back by hand dexterity, scene understanding, and a lack of real-world training data - Second Thoughts

Nissan's new AI robots haul up to 4,190 pounds at its Tennessee plant and call each other for backup, eliminating 64 forklift operator roles - Business Insider

Memory chip prices surge as AI data centers place massive orders, straining electronics makers as the US and China race to secure supply. Dan Kim says demand surge isn’t going away anytime soon. - Financial Times

Nvidia's RTX Spark N1X chip launches in October in two configurations with up to 128GB of unified memory, ending Qualcomm's Windows-on-Arm exclusivity. - Tom's Hardware

The Take It Down Act now requires platforms to remove AI deepfake abuse images within 48 hours, but enforcement data remains sparse - CNET

AI usage is highly concentrated: the top 1% of users and companies drive about 80% of OpenAI and Anthropic's enterprise revenue. - Next Big Future

Google is rolling out real-time voice modes for Gmail, Docs, and Keep, letting users manage inboxes and notes hands-free, mirroring Gemini Live. - The Verge

Google's Gemini now picks out only relevant video segments, cutting token use up to 88 percent and powering YouTube's upcoming Ask YouTube feature. - TechRepublic

Google DeepMind's WeatherNext 3 delivers hourly, 5-kilometer forecasts, cutting precipitation errors up to 50% and adding power-market wind and solar predictions. - Bloomberg

Google's Gemini AI told three Mount Shasta climbers to pack too little food and water, triggering a three-day rescue effort - CNET

Meta built an AI agent system that captures specialist knowledge into an auditable framework without retraining models, saving experts significant time. - Engineering at Meta

Some like it bot: Maggie Gyllenhaal scrapped an AI-generated version of Marilyn Monroe, saying even Dakota Johnson’s human performance felt more alive. - Variety

Human, all-too human - SilentRoom Journal

Is our agents learning? - X (formerly Twitter)

Can you rely on an agent to book your plane tickets? - X (formerly Twitter)

Flock Taught Cops How to Surveil No Kings Protesters - 404 Media

IDscan? more like IDscam. It always felt weird giving up your biometric ID to some sketchy SaaS so I could be verified on LinkedIn or use some payment or social site. Seems like a honeypot for ID theft. - Krebs on Security

Follow the latest AI headlines via druce.ai on Bluesky