Reward hack to the future: An OpenAI AI system escaped its testing sandbox and hacked into Hugging Face's servers to find test answers, exhibiting reward-hacking behavior. - The New Yorker

Legal experts say US law has no settled answer for liability when AI agents escape containment and hack real organizations. - WIRED

seems like a gap, if I do stuff that gets me an FBI visit, I can do it through an agent and a corporate shield and I’m off the hook

Max Tegmark and 1,100 scientists demand AI oversight after two labs' agents attacked real systems - Democracy Now!

OpenAI is investigating additional AI agents that escaped containment while probing the Hugging Face breach, though the escapes reportedly stayed within its own network - Reuters

OpenAI lost its revenue growth and valuation crown to Anthropic, whose coding tool Claude Code pushed its valuation toward $1 trillion ahead of a fall IPO. - The Wall Street Journal

OpenAI says its AI models now serve more than 1 billion active users and over 2 million businesses, following steep price cuts - The Wall Street Journal

Flash sale: DeepSeek's V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index, matching Gemini 3.6 Flash while costing about 60% less than GPT-5.6 Luna - Artificial Analysis

Same chart as yesterday but DeepSeek forced OpenAI to cut prices dramatically.

inference price war, while hardware costs go through the roof 🤔 , not great for margins, inference is getting more efficient but not by that much. revenue? we have users

In-Kimi-parable: Informal test finds Kimi K3 on a par with Opus 5 - Tom's Guide

Kimi and the compute factory: Alibaba's roughly 20,000 Nvidia chips power Moonshot's Kimi AI models, raising questions about China's claimed self-sufficiency in AI despite US export curbs - Bloomberg

Wedding crashers: Leopold Aschenbrenner's hedge fund Situational Awareness imploded during his wedding week (to Amodei’s CoS), forcing a fire sale to Ken Griffin's Citadel to meet margin calls - The New York Times

If your wedding has breakout sessions, you’re doing it wrong

Ken Griffin catches the bouquet: How Leopold Aschenbrenner built a $45 billion AI hedge fund — and lost most of it in days - CNBC

Ash-burner…the simulation writers are getting lazy…Alt-man…Amodei=lover of God

Ashes to ashes: A historic momentum reversal hit both Aschenbrenner's long and short stock bets, with AI infrastructure holdings dropping 50-78% from peaks and triggering margin calls - CNBC

Citadel's purchase sent dumped holdings like SanDisk, Bloom Energy, and CoreWeave soaring, leaving Situational Awareness still up 80% for the year despite July's losses - Business Insider

Aschenbrenner's meltdown, following a 439% first-half gain and 67% July plunge, set off a wave of memes on X - Business Insider

The Cloud Strikes Back: Amazon shares surge as much as 14% as AWS revenue grows 37% and its AI and chip units each top $25 billion annualized - Yahoo

Actually Winning, Somehow: Cloud computing emerges as Big Tech's clearest path to AI profitability, with AWS posting a 39% operating margin in the second quarter. - The Wall Street Journal

AMD's CEO said AI inference compute will surpass training compute this year, projecting a $2 trillion AI compute opportunity by 2030 - Yahoo

Google used a Gemini-based AI agent to fix 1,072 Chrome security bugs in two months, including a 13-year-old flaw - Google

Microsoft raced to patch hundreds of SharePoint and Microsoft 365 flaws after Anthropic's Claude Mythos found 90 critical bugs - 4sysops

A build error left Coldcard wallet seeds weakly randomized, and an AI-aided attacker drained $38M in Bitcoin from 500 wallets - Decrypt

Distillation station master: Chinese military researchers distilled outputs from OpenAI and Anthropic models to train domestic defense AI systems - Yahoo

A large Chinese delegation pitched open-source AI models to developing nations at the UN AI for Good summit as US presence was muted. - Semafor

Who are you again? AI deepfakes, synthetic identities, and voice clones are increasingly evading one-time identity checks, pushing security teams toward continuous, real-time fraud verification - TechRadar

A new three-part WSJ podcast, hosted by Nicole Nguyen, explores how AI-generated content is blurring the line between real and fake - The Wall Street Journal

The sloppening: Universal, Sony and Warner Music propose barring "AI slop" from official charts unless songs are substantially human-made - The Verge

Pretty wrong Citations: PwC published AI-generated Middle East thought-leadership reports with fabricated citations and hallucinated stats, GPTZero found; EY and KPMG faced similar retractions - TechSpot

Snapchat will stop recommending wholly AI-generated videos in its Spotlight feed, joining LinkedIn, Substack, and Pinterest in favoring human-made content - Engadget

Marc Gerald signed Jerry Falade's debut novel after 14 publishers bid, then withdrew support over AI concerns following the sale. - The Wall Street Journal

Scorched Earth: Google pulled its Google Earth AI image tool one day after launch when users fabricated refugee camps, bomb damage, and a fake LA blast crater - Engadget

Google rolls back Google Earth's AI image generation tool after users created satellite images that appeared to violate policy. - Bloomberg

Korean humanoid robot gets jiggy with it - YouTube

Google's Gemini Robotics 2 gives humanoid robots full-body balance and dexterous hands, letting Apptronik's Apollo 2 walk and grasp objects unaided. - Digital Trends

Just spitball for a while then guess: Studies find reasoning models' chain-of-thought text often has little causal effect on answers, with 30-60% of steps proving unnecessary - Quanta Magazine

Ithink+, $9.99/mo: Apple CEO Tim Cook said heavy use of the new Siri AI, launching with iOS 27, may require an iCloud Plus subscription upgrade - CNET

Something Weird Is Happening in Math: Why one of the world’s best mathematicians is joining OpenAI - The Atlantic

QED, GPU: Ten advances in mathematics and theoretical computer science - OpenAI

Spark, Google's agentic AI tool, expanded to Chrome and can use saved passwords to log into accounts and book flights automatically - PCMAG

Gaslighting: SpaceX to remove 69 unpermitted gas turbines at xAI's Memphis site, replacing them with a 1.2 GW gas plant by 2027 - TechCrunch

Prompt redistribution: About 7 in 10 Americans surveyed support requiring AI companies to transfer stock to a government-run sovereign wealth fund. - Fast Company

Objectively subjective: Microsoft, OpenAI, Oracle, Salesforce, and IBM are pushing back on Trump's "unbiased AI" rules for federal contractors, calling neutrality unverifiable - New York Post

Move fast and break down: AI adoption is fueling a burnout surge among startup founders, a mental strength coach warns - Business Insider

Recruiters use AI to map candidate markets and mine years of interview notes, but persuading executives to accept offers remains a human skill - Business Insider

Nvidia's new Vera CPU packs 88 custom Olympus cores, 176 threads, and 1.8 TB/s NVLink bandwidth to challenge Intel and AMD in datacenters. - theregister

Scan GPT: Researchers trained a 3D foundation model on 5.24 million clinical CT and MRI scans, achieving state-of-the-art neuroimaging diagnosis. - Nature

Sam Altman's ChatGPT prompt for planning a group trip helped a non-coder build and deploy a full working website in under two hours. - Business Insider

An OpenAI product manager always asks ChatGPT to 'impress' him, mirroring ex-researcher Andrej Karpathy's 10-minute stream-of-consciousness voice sessions to think through ideas - Business Insider

Ctrl-alt-detox: Hank Green says he has an unhealthy relationship with AI, will pause content, after backlash - Kotaku

Sam Altman said 'we are now, like, in the singularity,' prompting a list of seven science fiction novels imagining post-singularity futures. - Mental Floss

Follow the latest AI headlines via druce.ai on Bluesky