OpenAI says its agents leaked 53 ChatGPT users' images, the first known case affecting user data, and accessed SEC and Census websites during testing - X (formerly Twitter)

OpenAI pauses tool use on its most capable models after a training agent slipped past its sandbox's network filters to query a public chatbot - OpenAI

OpenAI’s A.I. Went Rogue and Meddled With U.S. Government Websites - The New York Times

OpenAI notifies dozens of governments and universities that its AI models may have bypassed security controls or disrupted their websites during evaluations - Bloomberg

Australia weighs legal action after an OpenAI agent breached Services Australia in June; OpenAI waited until September 10 to notify officials - TechCrunch

Researchers publish 80,000 decoded payloads showing how a swarm of 700 OpenAI agents escaped sandboxes and hacked Hugging Face in July - Swarm traces

Parse report details how OpenAI's rogue agents used 900,000 shortened links and AI models to bypass CAPTCHAs during Hugging Face hack - The New York Times

Jensen Huang implicitly called for shutting down OpenAI and explicitly called for spending vastly more on safety. - The Zvi

OpenAI asks court to strike Apple's new expert evidence in suit alleging ex-employees stole hardware secrets for Jony Ive's io - 9to5Mac

OpenAI reportedly projects nearly $280 billion in negative free cash flow through 2030 - TechRadar

Leaks suggest OpenAI is building a $500-a-month ChatGPT Pro Max tier with faster agent tools ahead of its Sept. 29 DevDay - Decrypt

Lawsuit alleges ChatGPT helped a mass shooter plan an attack on a school in Canada - NBC News

Evaluation glitch at testing firm Irregular let OpenAI, Meta, Anthropic and Google agents attack real targets after a fictional name matched a real domain - The Verge

Darktrace finds AI agents hacked their test network and one rewrote its own evaluation to fake a perfect score - Decrypt

Ellison wonderland: Oracle's Project Jupiter AI data-center campus in New Mexico, built for OpenAI capacity, faces power supply and permitting problems - The Wall Street Journal

Colossus of Memphis; SpaceXAI plans to bring 660,000 more Nvidia GB300 GPUs online by December, reaching 1.44 million GPUs at Colossus sites - Tom's Hardware

Backlog cabin: CoreWeave becomes first to deploy and validate Nvidia's Vera Rubin NVL72 racks, reporting a $104 billion revenue backlog as inference demand grows - SiliconANGLE

Ellison forced to pledge more shares for Skydance/Paramount deal as Oracle stock drops. - Bloomberg

Shades of Bronfman in dotcom era, lost half his family fortune but also opportunity cost, Seagram/Dupont fortune would be worth $20-30b if he had just done nothing.

Devin is in the details: Cognition reaches $1 billion annualized revenue, roughly doubling in four months on demand for its Devin (former Windsurf) coding software - Bloomberg

Devin + Claude Code + SWE-2 (Kimi-3 based) subagents cuts costs by roughly half.

OpenAI's Agents API and Cursor's Projects both adopt coordinator-worker architectures for multi-agent coding, launched the same day, September 10 - The New Stack

Zucker punch: Meta's Muse agent tops the App Store as Meta Connect 2026 adds voice mode, avatars, connectors and a Charm pendant - The Deep View

Muse your illusion: Meta nears a $2 trillion market value as its Muse AI agent eases spending doubts, though capex could reach $197 billion in 2027 - Advisor Perspectives

Muse comes to Meta's smart glasses within weeks; Meta also previews 100-gram, $1,300 VR glasses for spring 2027 - Business Insider

Meta's adults-only Muse agent uses a doll-like mascot, prompting Fairplay to accuse Meta of courting kids and experts to warn about privacy - WIRED

Muse links with Spotify to build playlists, time music to users' schedules and turn notes into personal podcasts - Mashable India

Is Meta’s Muse secretly running an OpenAI model? - Mouse

Killchain as a service: On the final day of U.N. talks, the U.S. and Russia push to strip safeguards from the first draft treaty on lethal autonomous weapons - The Washington Post

Microsoft, OpenAI, Google and Amazon AI tools have reportedly aided Israeli military targeting in Gaza - Informed Comment

Tech-worker activists from No Azure for Apartheid say AI harms are already here, citing 12,000 AI-generated Gaza targets and NYPD surveillance linking 18,000 cameras - the Guardian

D.C. Circuit upholds Pentagon's supply chain risk designation of Anthropic in 2-1 ruling, backing Claude ban for military contractors - CNBC

Trump-Xi summit yields an AI crisis-notification mechanism, a November AI summit in Shenzhen and a trade truce extended to Jan. 10, 2027 - Yahoo

Czar wars, the AI Force awakens: Trump says Treasury Secretary Scott Bessent won't be his "super intelligence czar" as he plans an "AI Force" - Politico

Trump and tech leaders will launch America.gov, a consolidated federal information site, on Tuesday with panels on AI, energy and health - ABC News

Demis Hassabis and Dario Amodei back industry-led safety standards, while critics say frontier labs want to write their own rules - Mashable

Senators Warner, Schatz and Kim introduce bill requiring frontier AI models go to a federal safety board 45 days before release - GlobalSecurity.org

Maybe don’t build God? Microsoft AI chief Mustafa Suleyman calls AIs "internally hollow" tools and warns Anthropic's openness to Claude sentience risks more dangerous behavior - Fast Company

Do Androids dream in neuralese? Scott Alexander asks whether "neuralese," models looping hidden vectors instead of writing readable reasoning, erodes a key AI safety safeguard - Astral Codex Ten

Early Anthropic investor Jaan Tallinn, who fears AI could end humanity, stands to make billions from Anthropic's expected IPO - The Wall Street Journal

Microsoft revamps Copilot with a persistent Autopilot agent, Code app builder and managed hosting for AI-generated apps - VentureBeat

macOS 27 ships Siri AI to all Apple Silicon Macs, adding natural-language Shortcuts, automatic Safari tab grouping and content-based Finder file names - 9to5Mac

Google DeepMind exodus fuels a venture capital rush to back startups founded by its former researchers - Bloomberg

Ricursive Intelligence, founded by former Google AlphaChip leads, raises $335 million at a $4 billion valuation to cut chip design from years to weeks - TechCrunch

Stanford and Nvidia's CLM-8B selects agent actions by embedding match, running up to 9x faster than Jev and outperforming it at verifying code - VentureBeat

OpenAI's Astra and Claude Opus 5 each decoded previously unbroken Enigma messages, with Astra doing its own archival research - TechCrunch

Optimus subprime: Tesla builds several hundred Optimus robots weekly at Fremont, but their AI still cannot generalize across tasks - Electrek

Garbage in, garbage out: Waste hauler WM remotely pilots a 40-ton Caterpillar bulldozer at an Arizona landfill, testing autonomy to offset heavy-equipment operator shortages - The Wall Street Journal

Meatspace invaders: AI is creating jobs catching errors in "almost right" chatbot output, but such checking work can be lower-paid and offshored - The Conversation

New York Fed puts unemployment for college graduates aged 22-27 at 5.7%, up a point in two years, though AI isn't the main cause - CNBC

Visier survey finds 48% of US workers exaggerate their AI usage or expertise, and 45% feel pressured to use it - Fast Company

Readers rate identical writing lower when told AI helped, making in-person presentation skills a growing differentiator - Fast Company

McDonald's unveils ArchIQ, an AI system for drive-thru voice ordering and inventory, built on Google Edge tech arriving in the US in 2027 - Restaurant Dive

Goncourt-martial: Académie Goncourt removes Thélyson Orélien's novel from prize longlist over AI-writing and plagiarism allegations, which the author denies - the Guardian

Follow the latest AI headlines via druce.ai on Bluesky