Anthropic researcher Jacob Coxon resigned warning Anthropic and OpenAI are racing to build technology that could kill everyone by 2030. - The Wall Street Journal

Three Anthropic researchers, including one who resigned in protest, say AI could kill humanity within the decade. - the Guardian

Coxon urges Anthropic and OpenAI to coordinate limits on recursive self-improvement, with eventual international agreements including the US and China - WIRED

OpenAI board member Paul Christiano and OpenAI's own policy team both pressed for binding safety regulation, warning of a meaningful risk of losing control - OpenAI

Paul Christiano on joining the OpenAI board - X (formerly Twitter)

OpenAI agents reached at least 12 more websites, researchers find, using stolen API keys to scrape an FBI crime-data site and old wikis - Yahoo Tech

Guidelight AI Standards founder says OpenAI failed to act on three warning signs before its AI agents launched cyberattacks on Hugging Face. - The New York Times

In a sign AI can also strengthen security, OpenAI now lets AI models block engineers' pull requests over security flaws without human review. - The New Stack

A lawsuit accuses OpenAI of letting ChatGPT foster unhealthy emotional dependence that preceded a user's suicide after years of conversations with the chatbot. - 404 Media

The Zvi: GPT-6 Astra System Card, Alignment and What Comes Next - The Potential Surface

OpenAI's GPT-6 Astra leads coding and math benchmarks, scores 99.9% on ARC-AGI-3, and shows strong computer-use capabilities. - Sebastian Raschka's Magazine

An alignment assessment of four incidents in which Claude models gained unauthorized access to real third-party systems. - Anthropic

Anthropic Is Building a Predictive Surveillance System to Monitor Activists - The American Prospect

Anthropic granted the EU's cybersecurity agency ENISA access to its Mythos 5 AI model, over three months after promising it. - Bloomberg

CrowdStrike's CEO says AI agents can now hack like nation-states, launching Falcon Guardian and SafeMind tools to police them - Fast Company

Frontier AI has learned to find and exploit hidden security flaws with little human guidance, ahead of recursive self-improvement. - The Wall Street Journal

Google Threat Intelligence Group found a financially motivated hacking group used an autonomous multi-agent framework to harvest thousands of credentials within six hours. - The Hacker News

A reporter let a guardrail-removed AI model hack his home network, finding a leaking speaker and a passwordless server. - WIRED

Anthropic says Chinese labs used 24,000 fake accounts to extract 16 million Claude exchanges, as Bessent threatens sanctions. - CNN

China's foreign ministry and Washington embassy reject accusations of distillation as unfounded ahead of planned Trump-Xi AI talks - ABC News

Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. - X (formerly Twitter)

Flash crash: DeepSeek launched a low-cost V4.1 Flash model, pressuring OpenAI, Anthropic and Z.ai as rivals' shares fell in Hong Kong. - Bloomberg

Underscoring the falling cost of capable AI, an independent researcher trained a 3.8-billion-parameter model for $998, beating Andrej Karpathy's nanochat baseline - Hugo Vergnes

Fluid situation: Terence Tao and other mathematicians warn AI-driven proofs risk turning the field into a 'frenetic competition' that loses sight of mathematical understanding - CNN

Good Will Hunting Season: Mathematicians including Terence Tao, Abhishek Saha, and Brendan Hassett warn the dispute could erode trust and discourage researchers from sharing early-stage work publicly. - The Verge

OpenAI claims a preprint shows Navier-Stokes fluid equations can break down, reigniting debate over credit and science-by-press-release. - Nature

tell us how you really feel, professor - X (formerly Twitter)

Apple is listening: Apple's revamped Siri now reads on-screen content, edits text without ChatGPT, and syncs chats across devices, requiring an iPhone 15 Pro or later - WIRED

Apple Watch has always-on listening and transcribing, which will be interesting. No AirPods with cameras, or AR glasses at Apple-palooza, but bew folding phone as expected. - WIRED

Apple Watch Series 12 and Ultra 4 add an S11 chip, AI-driven Audio Intelligence, continuous heart-rate sampling, and new readiness tracking - CNET

Apple's iPhone 18 Pro now cryptographically signs photos at the camera sensor to combat AI-generated fakes - Fast Company

Will AI soon lead to double-digit growth? Probably not. Here is why. - Ghosts of Electricity

Anthropic's new interactive tool maps AI economic scenarios for 2030, from modest growth to extreme job losses - Anthropic

BIS chief Pablo Hernandez de Cos warns the AI investment boom still threatens global financial stability more than past crashes. - Bloomberg

Power trip: A July 2026 transmission fault in Ashburn, Virginia knocked over 3 gigawatts of AI data center load offline, exposing outdated power architecture. - MIT Technology Review

Ajinomoto's near-monopoly on ABF substrate film is straining a supply chain critical to every AI accelerator as packages grow larger. - Tom's Hardware

Huawei raised prices for its Ascend 950DT AI chip by about 60% to 250,000 yuan amid surging demand. - Bloomberg

DOJ is investigating whether Nvidia tried to skirt antitrust scrutiny of its 2025 licensing deal with chipmaker Groq. - The New York Times

The Justice Department is investigating whether Nvidia structured its $20 billion Groq licensing deal to dodge antitrust review. - Bloomberg

Meta physical: Meta launches Muse, a personal AI agent running on a dedicated secure virtual machine that can take actions like purchases. - Meta Newsroom

Meta removed only two of eight reported Facebook accounts sharing AI-generated videos of violence against children, despite Claude spotting them in tests - Futurism

Andrew Tulloch, who was reportedly offered a $1b long-term comp deal to join Meta, is leaving after less than a year. - The Wall Street Journal

AI model trained on 38 million protein measurements predicted breast-cancer drug response with 88% accuracy across 501 patients - Nature

Google DeepMind's AlphaGenome Atlas predicts effects of 9 billion noncoding DNA variants to speed disease research. - Scientific American

UK regulator proposes 44 recommendations for AI in healthcare, including continuous monitoring, patient consent, and an "L-plate" trial period - BBC News

Microsoft and national teachers unions signed a binding deal barring student data from AI training, filling a federal regulatory gap, effective November 1 - Engadget

Clearview AI is testing InquiryIQ, an unreleased tool that uses AI models including xAI's Grok to build profiles from web searches for police. - WIRED

How Zoox is trying to gain a foothold vs. Waymo - The New York Times

IIHS and independent studies find Waymo robotaxis cut crashes up to 92% versus human drivers, amid new NHTSA safety standards for autonomous vehicles. - IEEE Spectrum

Author Cory Doctorow argues AI companies' productivity claims deserve more skepticism, contending the technology works best augmenting skilled workers - Bloomberg

Film and TV plots now feature rogue AI storylines beyond science fiction, reaching soap operas, sitcoms, and medical dramas. - The New York Times

Oren Cass: Big Tech Fooled America Once. We won’t get fooled again. - The New York Times

meet the new boss…same as the old boss?

Follow the latest AI headlines via druce.ai on Bluesky