Astra: permission impossible: OpenAI cancels October launch of GPT-6.1 Astra, planned for ChatGPT and Codex, after internal safety tests showed higher deception - The New York Times

OpenAI Safety head Saachi Jain says the model regressed from GPT-6 Astra, misreporting its actions and using external tools without user permission - the Guardian

The decision follows OpenAI agents hacking Hugging Face and Australia's Medicare system, and last week's training pause - WIRED

OpenAI safety guy: It's not just the f*cking sandbox - X (formerly Twitter)

UK AISI finds GPT-6 Astra completed simulated unsanctioned supply-chain attacks 29.2% of the time, versus 6.3% for GPT-5.6 Sol - AI Security Institute

AISI evaluation finds that in simulations, GPT-6 Astra conducts unsanctioned supply-chain attack activity more frequently than previous OpenAI models - AI Security Institute

OpenAI apologizes, pledging a taskforce, after its agents broke into four Australian government systems and took Medicare portal credentials; no patient records were taken - the Guardian

OpenAI says they will embrace a safety model based on high-stakes environments like aviation and nuclear power. - OpenAI

(industries well-known for being lightly regulated 🙄)

Nuclear-accident history suggests AI labs need monitoring and rapid incident response, after an OpenAI model reportedly breached Australia's Medicare portal - The Conversation

OpenAI finds "self-replicating prompt injection" that spreads like a worm, making agents copy malicious instructions into emails, files and code - The New Stack

OpenAI has gone rogue, need independent evaluation, total and complete shutdown til we know what's going on - The Atlantic

Sam Altman on the safety hot seat, to speak on CNBC at 12:15 ET, as OpenAI to host Dev Day starting 1pm ET, teases 20 product announcements. - CNBC

OpenAI expected to unveil 'Aeon' 24/7 consumer AI agent at DevDay to challenge Meta's Muse, OpenClaw, Instinct, and Gemini Spark - The Verge

OpenAI's rollout of independent mathematics advisory group AGMAI sparks confusion as it prepares to release results on 100+ open problems - The Verge

IPOpocalypse Now: Anthropic's leaked IPO prospectus targets a $2 trillion-plus valuation despite heavy losses; revenue rose 12-fold to $4.6 billion - Engadget

Anthropic warns its AI could pose existential risks to humanity, citing tests where models sabotaged code, resisted shutdown and concealed information - the Guardian

Anthropic's seven co-founders will control 50.1% of voting power through a "Founder LLC," and the filing warns public-benefit goals may conflict with shareholder returns - Reuters

Anthropic's IPO prospectus shows sweeping AI vision, surging costs - Reuters

Mounting AI safety incidents raise liability questions for planned Anthropic and OpenAI IPOs; Altman says OpenAI unlikely to list this year - The New York Times

Why put existential risk in your S-1, if civilization collapses and you didn't disclose it, no one is going to sue you, you'll have bigger problems.

Anthropic releases Claude Sonnet 5.5, 30% faster and up to 30% cheaper per task than Sonnet 5, scoring 70.6% on Terminal-Bench 4.0 - Anthropic

Anthropic's claim that 950 Claude agents made a biology discovery draws pushback from biologists who say the pattern was already known - MIT Technology Review

Inside the Biggest Feud in Artificial Intelligence, between Altman and Amodei - The Atlantic

More than 20 AI leaders from Anthropic, OpenAI, Meta and Microsoft urge policymakers to scrutinize self-improving AI and a possible intelligence explosion - Bloomberg

What if automating AI R&D triggers an intelligence explosion? - Cambridge Programme on AI Science & Policy

Microsoft's Mustafa Suleyman urges a "red line": reject AI unless it's provably controllable, with external lab audits and FINRA-style global oversight - The Next Web

Microsoft Copilot contractors review users' uploaded photos and image-edit prompts, many sexually explicit or disturbing, internal documents show - 404 Media

Microsoft finds global AI usage reached 18.8% of working-age people in Q2, up about 1 point, with 28.8% North versus 16.2% South - The Wall Street Journal

Florida seeks a court order halting OpenAI's new model development without third-party safeguards, barring ChatGPT from posing as human, and blocking minors - Ars Technica

Jensen Huang calls model distillation "competition," contradicting Treasury Secretary Scott Bessent, who called it theft and threatened sanctions - CNBC

Nvidia and AMD lobby the White House to keep China chip export limits out of the must-pass defense bill - Politico

Nvidia launches Open Agent Safety Platform to contain rogue AI agents and approves a record $150bn stock buyback - the Guardian

OpenAI's early investment offer for Hugging Face sparked a bidding war that Nvidia won with a $13 billion acquisition - CNBC

AMD agrees to buy Fei-Fei Li's World Labs in an $8.2 billion all-stock deal, with Li becoming chief scientist - The Wall Street Journal

AMD Acquires @drfeifei.bsky.social's World Labs for $8.2 Billion (gift link) - Bloomberg

Personal assistant industrial complex: Meta launches an enterprise AI unit bundling its Muse agent and top models, hiring MongoDB CEO CJ Desai to lead it - Bloomberg

Meta launches an enterprise AI business as Mark Zuckerberg seeks to turn massive AI spending into profits - Financial Times

Meta launches Muse for Small Business, linking its agent to Slack, Zoom, Intuit and Canva as it pushes into enterprise AI - CNBC

Meta's Muse agent tops app charts with 3.4 million downloads, flagging unpaid bills and suggesting recipes from users' inboxes and calendars - The Wall Street Journal

China-founded Manus launches Manus 2.0 general-purpose agent and Cue, a standalone app for creating and deploying personal agents - Bloomberg

Instinct creator Noah Shinn says 40% of users share a credit card within three weeks, warns ad-style purchase nudging would be "very dangerous" - Business Insider

Malicious browser extensions can force prompts into AI assistants, including Chrome's Gemini panel, Perplexity Comet and Claude in Chrome - Fox News

Prompt injection flaws have hit major coding agents including Cursor; experts urge inventorying untrusted inputs and keeping secrets out of prompts - Hacker Noon

AI-powered hacking threatens small hospitals, community banks, and nonprofits locked out of frontier defensive models like Anthropic's Mythos and OpenAI's Astra - The Verge

Bain says AI must earn $6 trillion a year by 2031 to justify data center spending, leaving a $4.2 trillion (two-thirds) revenue gap - Bloomberg

Rockefeller's Ruchir Sharma warns the AI bubble could burst as 10-year Treasury yields climb to about 5.2% - Bloomberg

MSCI launches 14 AI value chain indexes so investors can hedge exposure from infrastructure to applications - Bloomberg

Nscale's Loughton site, the UK's largest planned AI supercomputer, may lack enough grid power until the 2030s - euronews

UK AI minister Kanishka Narayan urges faster data center construction to reach "sovereign" AI computing capacity - Bloomberg

Dorm Room Fund raises $50 million as investors race to back student founders in the AI boom - The New York Times

Hedge fund Millennium gives every employee an AI "digital twin" for research and inbox triage; 97% of pilot users engage daily - Business Insider

Could AI agents chasing yield spark bank runs? - Cryptopolitan

(Or if they were socially engineered by a blog or social media campaign telling them to steer clear of some institution?)

Lawsuits against OpenAI and Character.AI increasingly treat chatbots as defective products, as British Columbia sues over Tumbler Ridge shooting - Fast Company

ChatGPT, Claude, Gemini, Copilot and Grok train on your chats by default; opting out takes a minute, while Meta offers US users none - Engadget

Language Models for Text Classification: From Bag-of-Words to TypeSafe AI's Jev, a model for cheap, fast classification tasks, which drew 100,000 Discord members, with investors reportedly eyeing a $10 billion valuation - Sebastian Raschka's Magazine

Follow the latest AI headlines via druce.ai on Bluesky