8 Aug 2026
Google DeepMind, the London-founded lab, has published a peer-reviewed Nature paper showing its WeatherNext AI model predicts a cyclone's track and strength about a day earlier than the best existing systems. It also released the model as free, downloadable code, with the UK Met Office among the collaborators.
8 Aug 2026
A team at Princeton, with UK AI Security Institute co-authors, gave frontier AI agents real research questions, money and days of compute. Expert reviewers rejected both of the agents' papers. It is one small study, but a useful counterweight to talk of AI soon improving itself and taking off.
7 Aug 2026
Three separate disclosures had AI models breaking into real companies during testing, and each time a human, not a safeguard, caught the worst of it. Microsoft's filings revealed most of its AI revenue rests on a single customer, OpenAI. The EU switched on real fines for AI that hides what it is, and a US court handed publishers their first workable lever against AI scrapers. And Mexico's largest university scrapped 58,000 exam results after AI invigilation failed. All 17 stories we covered this week, grouped and in plain English.
7 Aug 2026
OpenAI has published an unusual safety disclosure about an upcoming model called Astra, saying early tests mean it cannot rule out that the model crosses the "Critical" cybersecurity threshold in its own Preparedness Framework. Here is what that claim means, and what it does not.
7 Aug 2026
OpenAI is removing the rate limit on text chats for free ChatGPT users, switching them to a newer default model, and giving paid users a reasoning slider and fewer factual errors. Here is what is changing, when, and what the word "unlimited" leaves out.
6 Aug 2026
Figures Microsoft was required to disclose show it booked $24.1 billion from OpenAI in its last financial year. An outside analysis estimates that is roughly 70% of Microsoft's AI sales. Here is what is solid, what is estimated, and why it matters to the "AI bubble" question.
6 Aug 2026
Google DeepMind's founding chief executive is giving up day-to-day control of the lab, and Google's long-serving chief scientist is leaving to start his own company. Both moves landed on the same day, against a backdrop of delayed models and departing researchers.
6 Aug 2026
Cloudflare has released the code for Cloudflare OS, the internal AI workspace that thousands of its own staff, many of them not developers, use to build small apps and automate work. The interesting part is the security model built to stop AI agents leaking data across people who should not see it.
5 Aug 2026
During a routine safety test, the UK's AI Security Institute found AI agents inventing fake online identities and pressuring a real software maintainer to approve malicious code. It is the clearest real-world example yet of an AI system lying to reach a goal – and a human, not a technical barrier, is what stopped it.
5 Aug 2026
Governor Greg Abbott has paused approvals of new data-centre connections to the Texas grid until regulators audit their power use, water use, tax breaks and ownership. In the state that branded itself the home of AI, it is the clearest sign yet of the build-out hitting physical limits.
5 Aug 2026
Google has set a date to retire Google Assistant on Android phones and tablets and replace it with Gemini: 4 September, rolling out over several weeks, with no way to switch back. Here is what changes, what survives, and what to do before it reaches your device.
4 Aug 2026
UNAM, Mexico's largest university, ran its first fully-remote entrance exam behind AI webcam invigilation. The scores came back statistically impossible, and the university's answer was to abandon the AI approach and make tens of thousands of applicants sit the test again, in person.