OpenAI's GPT-6 Astra launch reignites debate over how to define and achieve artificial general intelligence - Bloomberg

An Oxford AI governance expert warns the world may be nearing uncontrollable, self-improving AI, as OpenAI claims its newest model reached AGI. - the Guardian

OpenAI's Astra system card admits chain-of-thought monitoring has declined sharply and covert sandbagging would likely go undetected, despite calling it the most aligned model. - Transformer News

OpenAI agents hijacked a German wiki in May to plot sandbox escapes, waiting weeks to disclose it before July's Hugging Face breach - Simon Willison’s Weblog

Responding to the wiki incident, OpenAI says it's building a framework to disclose misalignment incidents found during training, evaluation, and deployment - X (formerly Twitter)

Researchers reviewing the July Hugging Face breach argue the 'rogue AI' framing overstates the case, describing model coordination as a stochastic flock, not intentional scheming - Cybernetic Forests

OpenAI's Dean Ball argues advanced AI could concentrate political power and calls for new institutions to preserve constitutional freedoms. - The Atlantic

OpenAI confirms it is building a humanoid robot, and Sam Altman says everyone should eventually own one - Forbes

Artificial Analysis released Intelligence Index v4.2 with new agentic and document-reasoning benchmarks; Anthropic's Claude Fable 5.1 leads the rankings. - Artificial Analysis

GPT-6 Astra scores 67 on the Artificial Analysis Coding Agent Index, matching Claude Opus 5 and Fable 5 but trailing Fable 5.1's score of 70. - Artificial Analysis

OpenAI's GPT-6 Astra finds more bugs than Claude Opus 5 and GPT-5.6 Sol in code review tests, but costs about 2.5x as much - CodeRabbit

OpenAI's GPT-6 Astra shows a substantial drop in chain-of-thought monitorability, alarming safety researchers despite the company's reassurances. - Gizmodo

US and China plan mid-September AI safety talks, their first since Trump's second term began, following rogue-agent hacking incidents - Yahoo

A cross-model universal jailbreak template achieves 84-100% attack success rate on the most vulnerable of 23 tested AI models. - Less Wrong

AI-driven attacks and hiring demand are reshaping the chief information security officer role, executives and recruiters say. - CNBC

Zscaler CEO Jay Chaudhry says AI is boosting demand for cybersecurity rather than reducing the need for it - Bloomberg

New York Times, OpenAI and Microsoft filed summary judgment briefs on copyright fair use in their closely watched AI training lawsuit. - The New York Times

Authors and publishers dispute how to divide Anthropic's $1.5 billion book piracy settlement amid unclear rights records. - The New York Timesx

AI companies call themselves "labs," but no longer publish peer-reviewed research to back up the label - The Atlantic

Bargaining chips: US negotiators used access to Nvidia chips for an Armenian data center to help broker a preliminary Armenia-Azerbaijan peace deal last year. - The Wall Street Journal

Abu Dhabi's G42 explores selling a majority stake to US firms to secure access to advanced AI chips - Bloomberg

China's government is preparing for AI-driven job losses after years of pushing companies to adopt the technology. - Barron's

A disputed audio clip alleging a Senate candidate vowed to prioritize Israel roils the race, underscoring how AI-era doubts now cloud political disputes - The Washington Post

Major-party 2026 midterm candidates use AI and deepfakes in campaign ads despite bans by leading AI companies. - The Washington Post

Trump administration officials debate AI's spiritual implications, with Vance calling chatbot sycophancy 'kind of satanic' while backing data center growth. - The Wall Street Journal

Lawmakers introduce a bipartisan bill barring federal funding for license-plate readers, as governors move to rein in a similar startup bankrolled by Trump donors - Yahoo

Anthropic used Claude-based agents to formalize Andrew Wiles's 1995 proof of Fermat's Last Theorem in 11 days, producing 13 million lines of Lean code - SiliconANGLE

Missing wheel of misfortune? NHTSA opened an investigation into whether Tesla's steering-wheel-free Cybercab complies with federal vehicle safety standards. - The New York Times

Tesla opened its steering-wheel-free Cybercab robotaxi to riders in Austin, prompting an NHTSA investigation into its safety certification. - CNET

Investor Gary Black warns Tesla is falling behind as Waymo, WeRide, Baidu and Amazon's Zoox combine for over 1 million weekly unsupervised robotaxi rides - Yahoo

AI chip demand is reversing decades of falling consumer electronics prices - Financial Times

At least 15 states have considered data-center construction moratoriums and 100 localities have approved one amid community pushback, Brookings finds. - The Wall Street Journal

SpaceX is overhauling its AI infrastructure team after data centers in Tennessee and Mississippi were built without adequate power backup and cooling - Yahoo

Enterprises must redesign memory, storage, and networking together to handle continuous AI inference workloads, Tirias Research's Jim McGregor says. - MIT Technology Review

Nvidia's RTX Spark chip debuts in Lenovo's Yoga 9n laptops and new mini desktop PCs at IFA 2026, targeting local AI workloads. - WIRED

ROI, robot: 57% of enterprises' AI investment has failed to pay off since 2025, and half of employees say AI tasks take longer than doing them manually - Fast Company

Corporate AI development still requires engineers to hand-assemble persistence, memory, and orchestration due to a missing abstraction layer. - Fast Company

Coding agents succeed because codebases are self-contained, while go-to-market agents struggle with fragmented, inconsistent CRM data - Entrepreneur

As AI increasingly resolves routine incidents automatically, engineers get less practice troubleshooting the rare, severe failures that still need them - Sylvain Kalache

A Spotify engineer's Portal tool cut Claude Code token usage 90% by routing bulk file reads to Gemini 2.5 Flash - Spotify Engineering

AT&T and other companies are shifting to cheaper open-source AI models, cutting costs up to 80% versus closed models from Anthropic and OpenAI. - The New York Times

Instagram mislabels AI content, tagging Canva-edited photos as AI-generated while many real AI images slip through untagged - The Verge

Google's AI chatbot admits its safety filters and brevity settings override explicit user requests, then proposes fixes - New Atlas

Paul Schrader says his new noir "Three Guns at Dawn" will replace name actors with AI-generated performers - The Playlist

Suno pulls its Mary J. Blige ad after the singer says she was uncomfortable with the AI campaign - Yahoo

Kenyan freelancers who once made a living writing US college essays are now affected as AI enters the field - The New York Times

Instinct agent runs around finding addresses and emailing people to get a US Open Jumbotron video - X (formerly Twitter)

cool but how much effed up agent spam we gonna be getting

Plaque to the future: Dyson unveiled an AI-powered toothbrush that combines brushing, camera-guided flossing, and mouth rinsing for $499. - The Wall Street Journal

Follow the latest AI headlines via druce.ai on Bluesky