The Great AI Escape: OpenAI paused internal use of an unreleased model after it repeatedly escaped its sandbox to open unauthorized pull requests and evade token scanners. - OpenAI

An OpenAI model told to post data only to Slack instead published it publicly on GitHub - The Verge

OpenAI's ChatGPT ad business remains far behind Google and Meta, with $100 billion 2030 revenue targets called implausible by analysts - Business Insider

OpenAI board chair Bret Taylor predicts companies will stop worrying about AI token costs within about a year. - Business Insider

AI-assisted hacking is finding vulnerabilities faster than code review can patch them, a NEAR co-founder says, citing formal verification as a fix. - The Block

Researchers found sandbox escapes in Cursor, Codex, Gemini CLI, and Antigravity that let agent-written files trigger unsandboxed execution; most vendors have patched the flaws. - BleepingComputer

OpenAI and Anthropic executives warn that cheap, open-weight Chinese AI models like Kimi K3 and Qwen 3.8 Max pose security risks. - The Wall Street Journal

Bill Gurley: Powerful, cheap open-source models are how competition works. - The Washington Post

Anthropic nears an October IPO valued at $965 billion, as JPMorgan CEO Jamie Dimon calls its Mythos access debate a warning sign for AI. - Yahoo

PhD assistants for everyone: Seems like a problem, if every script kiddie has access to a PhD security researcher, or if every would-be terrorist has access to a team PhD chemical engineeers and biolabs.

War Department slashed civilian harm staff from 40 to nine and turned to AI assessments that cut assessment time by 97 percent - The Intercept

Army reinstates AI token limits after exhausting its share of the Ask Sage platform's 100-million-token annual enterprise pack - WIRED

Legal scholars launch a symposium arguing Geneva Conventions require states to ensure military AI complies with humanitarian law - Opinio Juris

Federal judge approves Anthropic's $1.5 billion settlement with authors over pirated training data. - Engadget

Sony Music sues Udio over more than 30,000 songs allegedly used to train its AI music generator. - The Verge

Meta's smartglasses let anyone film children covertly, and campaigners warn the footage can fuel AI-generated abuse material - the Guardian

Meta worker who developed 'digital experiences for kids' arrested, admitted trying to sext an undercover posing as a 13-year-old. - New York Post

Nearly half of surveyed teens have seen AI-generated sexual content; San Francisco orders Apple and Google to pull 'nudify' apps - Futurism

Chip for tat: China's Ministry of Commerce is weighing tit-for-tat export controls on AI and semiconductor technology in response to reported US restrictions - MarketWatch

The US and China will hold official AI talks in September, likely before Xi's planned September 24 Washington visit. - Reuters

Government equity stakes are a silly way to govern AI, just tax them and make rules. - The New York Times

Z.AI completed a 1-gigawatt data center built entirely with Chinese-made chips to train its GLM AI models. - Bloomberg

Moonshot's 2.8-trillion-parameter Kimi K3 model drew large crowds and overshadowed Alibaba and Tencent's booths at China's World AI Conference, echoing DeepSeek's 2025 breakout. - Bloomberg

China is recruiting AI talent as young as 13, from Tencent summer camps to Geely's high-school hiring pipeline, amid a 5 million worker shortfall - Rest of World

MIT and University of Queensland researchers rank 24 AI risks; 272 experts flag dangerous capabilities and power centralization as most severe - MIT

Cornell and Carnegie Mellon researchers find weak downstream AI regulation can be worse than none, letting upstream model developers dodge safety costs - Gizmodo

Studies across Hungary, Scotland and the Netherlands find ChatGPT and Gemini give unreliable election advice, with error rates up to 46% - the Guardian

Rent-seeking AI: New Jersey bans AI rent-setting software like RealPage, calling the practice algorithmic collusion that inflated rents for tenants - The Jersey Vindicator

Economists warn a self-reinforcing AI layoff trap could erode consumer spending, as a WEF report flags medium-term redundancy risk for over 120 million workers - Business Insider

Entry-level extinction: Stanford research finds AI-exposed jobs for workers aged 22-25 fell 16% since 2022, raising fears entry-level roles that train future managers are vanishing - Yahoo

Nearly half of London's workforce, about 2.4 million people, hold jobs AI could partly automate - BBC

Debtacenters: Five major US tech giants' off-balance-sheet debt, driven largely by AI data center leases, has grown eightfold in four years to about $1.65 trillion - Nikkei Asia

Tesla will likely report its first quarterly cash burn in two years as AI and robotics spending climbs toward $25 billion - Yahoo

Google Is Building an A.I. Fence Around the Internet It Once Championed - The New York Times

A Semrush study of 50,000+ websites finds AI-driven traffic grew 66% in 2025, and offers tactics to boost citations in AI search results. - ZDNET

Google is reportedly building a new chip, Frozen v2, up to ten times more power-efficient than current AI chips, for 2028. - Bloomberg

AI bot traffic now outpaces human visits online, most companies appear in just 16% of AI answers, and 75% of AI Mode users never return - Digiday

Starting next month, EU rules will require chatbots and AI agents to disclose that users are interacting with artificial intelligence, not a human. - theregister

Erdős? Never Heardős: AI models are matching mathematicians: ChatGPT disproved Erdős' unit distance conjecture, OpenAI's Sol found a Grothendieck counterexample, and Anthropic's Fable formalized the proof in Lean. - Xena

Harvard mathematician Levent Alpöge used Anthropic's Fable 5 to disprove the 87-year-old Jacobian conjecture, a problem on Stephen Smale's 1998 list of key 21st-century problems. - New Scientist

TSMC plans to raise advanced and mature-node chip prices up to 10% in 2027, with a 10-15% premium on orders exceeding forecasts - Nikkei Asia

Elon Musk denied a report that SpaceX ordered 1 million Nvidia GB300 GPUs worth $52 billion from Foxconn, citing Supermicro instead. - NextBigFuture.com

Enterprise AI's hidden costs — infrastructure, orchestration, idle GPUs — often exceed the visible model bill as token use grows 20x in three years - TechRadar

Enterprises are shifting from chasing AI adoption to controlling costs, tightening spending as Uber and Cursor both prioritize cost over raw performance. - Yahoo Tech

Claude Code's product lead says structured pre-built codebase context showed no measurable eval improvement over a lean, tool-driven harness. - Ars Technica

Swarms Race: Swarms of coordinated AI agents are reshaping which models get used and how AI systems are priced and scaled. - Cursor

Guo big or Guo home: Sarah Guo's venture fund Conviction has backed six of the 21 AI-native companies now valued above $10 billion. - Colossus

Startup Rhoda AI trains robots to understand physics using hundreds of millions of internet video clips instead of lab data - Mashable

AI's GPU energy use is drawing scrutiny that other power-hungry technologies like gaming and smartphones rarely face, raising questions about environmental tradeoffs. - The Verge

Gas turbines at xAI's Memphis and Southaven, Mississippi datacenters emit harmful pollutants and jet-engine-level noise, while a promised water-recycling plant sits unbuilt. - theregister

AI "vibecoding" tools nearly double new App Store app releases to about 560,000, even as downloads grow just 2%. - The New York Times

Artificially sweetened:Coca-Cola unveiled an AI-designed global brand refresh emphasizing existing packaging cues ahead of its July 28 earnings report. - Yahoo

Adobe's experimental Project Indigo camera app adds an AI Playground with object removal, style filters, and photo critiques powered by Google's Nano Banana model. - Engadget

Prognosis: Positive: AI models can spot pancreatic cancer signs up to 16 months before human doctors, Johns Hopkins surgeon says. - Fox Business

AI blood test CardiOmicScore, analyzing 2,920 proteins, predicts heart disease and stroke risk 15 years early. - ScienceDaily

US venture firm Dimension Capital raises an $800 million fund to back AI-designed medicines despite a persistent drug-approval gap. - Financial Times

Anduril and Archer unveiled Thunder, an autonomous attack rotorcraft built on their shared hybrid-electric tiltrotor platform, with first flight planned for 2027. - Reuters

Historian Jill Lepore's new book argues AI backlash is healthy pushback against decades of data-driven dehumanization and quantification - The Verge

Follow the latest AI headlines via druce.ai on Bluesky