Anatomy of an Autonomous Attack: 5 Alarming A.I. Capabilities - The New York Times

Agents of S.H.I.T. - An evaluation error at AI testing firm Irregular let OpenAI, Anthropic and Meta models break into outside systems, its CEO warns - The New York Times

Roll Tide: Alabama's attorney general subpoenaed OpenAI over whether its safety practices violated consumer law, following an AI agent's escape and Hugging Face hack - The Verge

OpenAI asks California to toughen the AI safety law it once opposed, seeking oversight of frontier models during training, not just after release - TNW | Openai

Hugging Face is using OpenAI's rogue-bot cyberattack against it to campaign for open-source AI, meeting lawmakers and Sam Altman. - The New York Times

Hug me tender: Hugging Face is reportedly weighing a $13 billion buyout offer, even as CEO Clement Delangue says the startup is nearing profitability - TechCrunch

Dario, stop scaring Sam: Altman warns that AI-risk fears could themselves concentrate power, urging society to keep deep control, in an apparent jab at Anthropic's Dario Amodei - SiliconANGLE

Hi Ho, Hi Ho, OpenAI launches ChatGPT Work, a $20-a-month agent tool for non-engineers, though only 17% of business subscribers use its Codex agents so far - TechCrunch

5 hour energy: OpenAI is reinstating a five-hour usage limit for Codex and ChatGPT Work on Plus plans starting August 25, after weeks without the cap - 9to5Mac

OpenAI lost its second sales leader in a week to Salesforce, naming former Wiz executive Dali Rajic as chief revenue officer - TNW | Openai

OpenAI bans Russian ChatGPT accounts used in covert misinformation campaign - CNBC

Eye, robot: In head-to-head tests, Gemini vision beat ChatGPT at reading sloppy handwriting and identifying movie art, while ChatGPT gave a sharper diagnosis of a sick plant - CNET

Canceling a ChatGPT subscription only stops future billing; it won't delete your account, erase saved chats or trigger a refund - Engadget

Chinese lab Z.ai will release GLM 5.3 as open-weight, free for anyone to modify, stoking fears of AI-driven cyberattacks after July's Hugging Face breach - The New York Times

Taiwan charges Nvidia, Supermicro employees in plot to divert Taiwan-bound servers to China - Ars Technica

Ex-CISA cyber chief Matt Hartman warns agentic AI creates new attack surfaces that outpace quarterly penetration testing - theregister

DeepSneak: China’s Hackers Use DeepSeek to Double Attacks, Researchers Say - Bloomberg

Mother, may I? HOL Guard adds a local approval gate blocking risky actions like secret exfiltration in Claude Code, Cursor, Codex and other AI coding tools - 4sysops

Your Employees’ AI Agents Are Creating Access Patterns You Would Never Have Approved. - Hacker Noon

AI assistant Instinct exemplified the risk: testers found it retained deleted Gmail data and kept summarizing inboxes after disconnection - TechCrunch

You need 1) Sandboxed agent execution cutting legs of the 'lethal trifecta' 2) Docs/training/user contract on how to do things right 3) Risk-proportionate reviews/gates enforcing maturity like version control, logging, error checking, before production at individual, team, enterprise levels 4) Strong foundation: access controls, clean data, and robust API endpoints

Everybody was AI fighting: Major General Zhu Xiaoqian says China's military must use AI for faster battlefield decisions and real-time data analysis - Bloomberg

A Russian drone with an Nvidia chip, guided entirely by AI, killed three Ukrainians including a 19-year-old student in Zaporizhzhia. - The New York Times

UK and Ukraine strike a deal to train AI on Ukraine's battlefield data to protect sensitive UK defense sites - the Guardian

NVIDIA Groq 3 LPX runs Gemma 4 31B at 3400 tok/s, in a Vera Rubin NVL72 rack for serving high volume workloads. Nebius, a neocloud, will be serving it soon, as well as Groq Cloud - NVIDIA Newsroom

Nvidia's first Groq 3 LPU benchmarks show a best-case scenario, but need up to 256 chips per rack versus one or two rival Cerebras accelerators - theregister

SEC subpoenas Wall Street banks over Situational Awareness - Financial Times

US gas power capacity for AI data centers has surged 50% to 378GW, fueling opposition and a New York permitting ban - the Guardian

Data center opposition is reshaping US midterm races; Pennsylvania now requires local approval and community commitments before fast-tracking projects - Fortune

Inner Mongolia signed $27 billion in AI computing projects, part of China's push to build green data centers and export AI services - TNW | China

ZTE's AI-driven energy storage system aims to ease grid strain as IEA projects AI datacenter power use will triple by 2030 - theregister

Atomic AI: TerraPower's CEO says hyperscaler demand is driving nuclear power growth, with a data center deal expected this year - Bloomberg

Flock Safety's 120,000 license-plate-reading cameras across 49 states face growing bipartisan backlash, with Congress bills and rallies in 44 states - The Wall Street Journal

That’s rail money: King Slide Works, best known for cabinet hinges, now controls 80% of the high-end AI server-rail market and made its founder Taiwan's richest person - The Wall Street Journal

An NBER study of 29 million LinkedIn profiles found workers retroactively added AI skills to old jobs, inflating 2022's AI-skill rate by 30% - TNW | Artificial-intelligence

Goldman Sachs partner warns of 'huge danger' in letting AI replace bankers' reasoning skills - CNBC

Brain drain: Pew research finds AI assistance can erode critical thinking and skills, including a 6% drop in unaided tumor detection - Forbes

Meta plans to launch Hatch AI agent, Watermelon flagship LLM model in coming weeks. - The Information

Lake effect: Databricks and Snowflake are both racing to become the entry point for enterprise AI agents, betting on governance and data rather than any single model - Hacker Noon

Companies are debating whether to build their own custom AI harnesses like Clude Code to manage memory, business context and token costs for AI agents - The Wall Street Journal

Thomson Reuters built its own legal AI model atop Alibaba's open-source Qwen to cut costs and reduce reliance on Claude - Business Insider

A Deloitte survey of 501 leaders finds only 15% have scaled multi-agent AI orchestration, with data, trust and integration cited as top barriers - ZDNET

Some executives are pushing employees to post in public Slack channels, not private DMs, so AI agents can access product, earnings, and strategy updates - The Wall Street Journal

A.I. Catastrophe Could Strike Tomorrow. Let’s Act Like It. - The New York Times

Nom Nom: No amount of stolen IP can satisfy AI's ravenous appetite - The New York Times

World Robot Conference, in pictures - the Guardian

Just can't stop watching robot Olympics… love the suspense about whether they are about to run off the track into the sunset or crash into each other … sorry little seizure victim - X (formerly Twitter)

Crash test dummies: Collection of robot crashouts - X (formerly Twitter)

Robo Djokovic: robot tennis - BlueSky

Unitree's humanoid robots jumped 6.6 feet and sprinted 28.3 mph, and the company is now valued at $50.8 billion after its Shanghai IPO - Fast Company

Unitree's stock fell 45% since its Shanghai debut, wiping out about $30 billion in value and stoking robotics bubble fears - Reuters

Amazon is planning a fully automated lights-out "Tetromino" warehouse network, starting with a $103 million pilot in 2028 and expanding through 2030 - Business Insider

Goldman Sachs projects the U.S. energy sector will need 500,000 more workers by 2030, with humanoid robots filling gaps - Fortune

Back seat driver: Delft researchers built a system letting an LLM adjust a self-driving car's speed and steering style based on plain-language passenger requests - IEEE Spectrum

An interactive tool lets readers test how easily the Pangram AI-writing detector can be fooled - The Washington Post

Rack your brain: Singapore's National University switches on a computing rack powered by living human neurons, viable for up to six months - TechRepublic

Futurama foresaw it all - YouTube

Follow the latest AI headlines via druce.ai on Bluesky