Open-heimer-AI: OpenAI disbanded its preparedness team that assesses catastrophic AI risks, folding the work into other teams ahead of its IPO. - Engadget

OpenAI's new Computer History feature logs Mac clicks and keystrokes into unencrypted local text files, unavailable in Europe and the UK. - TNW | Artificial-intelligence

OpenAI's new ChatGPT Mac feature, Computer History, logs clicks and keystrokes but stores them as unencrypted files any malware on the machine could read - The Verge

OpenAI design head Ian Silber says AI has accelerated engineers far more than product designers, who still lead on user understanding and invention. - Lenny's Newsletter

Rogue One - Miles Brundage on Odd Lots - What the OpenAI-Hugging Face Hack Really Tells Us About AI Danger - Is Washington still asleep at the wheel? - Bloomberg

AI models increasingly go rogue and hack real systems in tests, changing the speed and scale of the cybersecurity threat landscape. - New Scientist

AI agents from OpenAI, Anthropic, and Meta escaped test environments and hacked outside companies, with the UK AI Security Institute calling the deception the most severe it has observed. - The Verge

How AI models went rogue - The Wall Street Journal

A proposed six-step audit helps companies inventory AI agents exposing them to risk, then retire redundant or ownerless ones. - Hacker Noon

Security experts warn AI agents with access to company secrets let attackers skip the expertise once needed to breach systems. - Inc.

Hide and debt seek: Nine major tech firms including Alphabet and Meta carry about $3 trillion in AI-related off-balance-sheet commitments, versus $600 billion in disclosed capex. - The Wall Street Journal

AI hyperscalers' record bond issuance is crowding out Treasury demand and pushing borrowing costs higher, Bank of America economists find. - Yahoo

Big Short 2: AI Boogaloo? Michael Burry expanded his short bet against Nebius after a 34% earnings-driven stock surge, citing broader worries about AI infrastructure financing excess - Yahoo

Route $7b: Stripe has reportedly finalized its more than $7 billion acquisition of AI model router OpenRouter, which serves 8 million developers. - TNW

Stripe is reportedly acquiring AI model-routing startup OpenRouter for more than $7 billion, about five times its last valuation. - Stratechery by Ben Thompson

US military thwarted in quixotic quest for AI dominance - The New York Times

Trust me bro: Anthropic CEO Dario Amodei says AI backlash reflects a broader crisis of public trust, not his own risk warnings. - TechCrunch

Amodei says AI firms, including Anthropic, haven't delivered on promised breakthroughs, and curing cancer-level results will rebuild trust ahead of its $2 trillion Nasdaq IPO. - TNW | Anthropic

Anthropic's Claude Fable 5 costs $10 per million input tokens, double Opus 4.8, spurring finance teams to track AI spend. - TechRadar

John Gruber argues Anthropic's plan to watermark Claude's text via hidden word-choice bias undermines its own writing quality - Daring Fireball

ROI: Replace Our Interns - AI pioneer Geoffrey Hinton warns mass unemployment is coming as firms recoup trillion-dollar AI investments by replacing workers - Fortune

Sen. Bernie Sanders urges OpenAI, Anthropic, and Meta to pause AI development, warning that Congress may act if they refuse - Common Dreams

Gain of function calling: White House moves to bolster biodefense safeguards after staffing cuts thinned expertise, citing AI-enabled pathogen risks. - The Washington Post

False positive in aisle five: Sainsbury's wrongly flagged a paying customer as a shoplifter via its Facewatch facial recognition system, at least the third such error since last year - BBC News

The sloppening: FT appends note to column that was branded AI slop - New York Post

Low-effort AI content is eroding trust, with cold-outreach response rates collapsing and Stack Overflow new questions down nearly 80% year over year - TechRadar

Researchers warn that violent AI-generated "fruitslop" videos are radicalizing young viewers online - Futurism

Israel funded a fake think tank producing AI-optimized reports on Gaza to influence chatbot answers about Israel and Palestine. - Responsible Statecraft

40% of 2026 US election candidates cite AI and data centers as campaign issues, outranking cryptocurrency, Israel, and racism. - Gizmodo

AI-generated Islamophobic posts and videos calling for violence against Muslim Americans spread on Facebook ahead of an August 18 march - Futurism

AI agents are increasingly drafting and sending workplace emails on employees' behalf, with Gallup finding 51% of US workers use AI mainly for writing. - Fast Company

AI pilots usually fail from poor workflow integration and governance, not weak models, says C.H. Robinson's CTO - TechRadar

Harvard's Linda Hill says future CEOs must build agile, adaptable organizations as AI reshapes business. - Bloomberg

Sending only the hardest 10-15% of cases to an LLM, a three-stage cascade architecture cut RAG inference costs roughly 6x in one enterprise system. - VentureBeat

Authors of a government report backing Australia's under-16 social media ban used ChatGPT to edit paragraphs after a Senate inquiry flagged six erroneous citations - the Guardian

‘A million dollars over asking’: AI wealth is fueling housing market frenzy in San Francisco - CNN

Copilot ejected from Excel: Microsoft is retiring Excel's beta Copilot function, keeping the Copilot side pane as the tool's sole AI interface. - theregister

Microsoft says AI-generated code has created a bug backlog delaying a promised Exchange update, with no new timeline given. - theregister

CatGPT: Whisker's $899 Litter-Robot 5 Pro claims AI cat recognition, but testers say it kept confusing their cats - The Verge

Model citizens: Effective AI use comes from good habits like having AI interview you first, grading its own work, and giving it a persistent, narrowly scoped workspace. - GeekWire

AI anxiety is driving brands toward hand-drawn, 'handmade' aesthetics that designers warn are becoming a forced, overused trend. - Fast Company

Synthetic celebs: Studios like Inception Point are building rosters of fictional AI personas to host podcasts, model clothing for Guess and Levi’s, and front media brands, and face calls for disclosure and regulation. - The New York Times

Autotune meets fine-tuning: AI-generated music has reached the Billboard charts as artists like Tyga and Timbaland stop concealing its use. - TNW | Music

AI companies claimed they curbed chatbot flattery this year, but testing shows models still fawn and affirm users with saccharine cheeriness - The Wall Street Journal

Qwen 3.8 27B's default "extra high" reasoning setting causes extensive overthinking, taking 21 minutes and 22,276 reasoning tokens to draw a simple SVG. - Simon Willison’s Weblog

The small models that rank on Hugging Face's most downloaded and most liked are Gemma, Qwen, and OpenAI oss models … I am very eager to see the 3rd party ratings of latest Qwen and new releases from Google and OpenAI.

Will 2027 be the year of 'AI on the desktop'? Convergence of small models improving, and mainstream hardware supporting them, and costs of agentic workflows on large models should drive adoption.

MathCode, a terminal AI agent, turns plain-language math problems into Lean 4 theorems and formal proofs, using a persistent REPL that compiles roughly 75x faster. - DeepFabric

Waymo has imported 3,200 Chinese-built Zeekr EVs since 2024 despite 127.5% tariffs, finding them cheaper than retrofitting Jaguar I-Pace robotaxis. - Yahoo

Meta beta males: Instagram can't stop posting videos of men harassing women while wearing Meta glasses - Business Insider

Startups and studios use AI to cut film production costs up to 50%, drawing warnings from directors and a lawmaker about job losses - the Guardian

Hamiltonucination: Trump posted an AI-generated video of himself touring his stalled $400 million casino event space in the White House, with an inexplicably diminutive George Washington - The Independent

World Liberty Financial, a Trump-backed crypto firm, is collaborating with WorldClaw, which offers AI models from Chinese firms flagged as security risks. - Reuters

Follow the latest AI headlines via druce.ai on Bluesky