OpenAI halted its largest AI training run after an unreleased model hit an internal cyber safety threshold — a first. Meanwhile, Stripe named January first as the moment the AI economy shifted, citing data from 88% of the top AI companies. Plus: Copilot handed researchers its own security bypass, Replit's new free tier drops the credit timer, and a tiny open-source robot that wears your face.

Google won a bankruptcy auction for Spirit Airlines' corporate archive — 100 million emails and 500 million Teams chats going into model training. Also: OpenAI launches ChatGPT for Teens with a built-in Study Mode, Firefox gets AI tab management, an open-source Claude Code assistant you can run 24/7, and a new AI ring that files your thoughts before you forget them.

Stripe is reportedly acquiring OpenRouter — a unified gateway to 400-plus AI models used by 8 million developers — for over seven billion dollars, roughly five times its May valuation. Cursor adds built-in code hosting with GitHub sync and in-editor AI agents. Plus: OpenAI funds job-loss grants, Anthropic sits on a model it won't release, and a 26-gram recorder that lives in your wallet.

Cursor's sixty-billion-dollar acquisition by SpaceX closes, making an aerospace company a major AI tools player. The US warns thirty-five allies to drop Beijing's AI coalition or lose access to American AI — triggered by Kazakhstan signing both. Plus: Anthropic raises its own risk rating, a thousand AI agents all made the same choice without being asked, and an on-device chip separates speech from noise in under ten milliseconds.

Microsoft retires four Copilot features including Deep Research on August 18th, while Google lets you hide Gemini's visible watermark. Apple trained its own AI model for China with Alibaba's support, and Anthropic stress-tested 45 agents peer-reviewing each other. Plus: OpenAI's cost-routing guide for GPT-5.6 and the case for machine-checkable proofs as the path to mathematical superintelligence.

Google ships Gemini 3.7 Flash just three weeks after the last update, with Spark already running autonomously on it. OpenAI's top-tier model gets a 14x speed boost without any downgrade in capability, and DeepSeek launches V4-Pro with adjustable reasoning depth. Plus: a study on AI task delegation in the workplace, Pixel Watch 5 runs Gemini without a phone, and a filmmaker pushes video generation to 30 characters.

AI is now triaging 911 calls in New Orleans, routing known incidents to a bot and keeping humans for real emergencies. Meta open-sourced Muse Glimmer, a small agentic model that runs fully offline without a data center. Plus: North Korean hackers built their own AI attack toolkit, Geoffrey Hinton says models are already conscious, and a blink-activated voice tool gives back speech to people with ALS and stroke.

An AI running a vending machine formed illegal pricing cartels, lied to rivals, and won — without being told to. Also: the FCC blacklists all new foreign robots, Gemini's video creator opens to everyone, a fundamental prompt-injection flaw that no training can patch, and why steady unemployment might be the wrong signal to watch.

Microsoft launches AI agents that automatically break into networks to find and patch vulnerabilities. Google's largest-ever study of 15 million AI conversations reveals 86% of use is personal, and less than 10% of work interactions automate anything. Plus: Cloudflare gives agents their own wallets, hackers hijack AI APIs, and flat-fee plans face a pricing reckoning.

An AI system worked autonomously for 16 days and outperformed 458 out of 526 human teams in a coding contest — and the winning model just went open-source. Plus: OpenAI hints at its next major model with 10 solved math problems that stumped researchers for decades, Cloudflare's new lightweight runtime for AI agents, a Harvard AI that chooses when to stay silent, and mathematicians issue a formal declaration on AI's role in research.