AI News09 Aug 2026
OpenAI paused its own model for getting too good at hacking
Every AI lab publishes a safety framework, and until Friday not one of them had ever been tripped. OpenAI now says its upcoming model, Astra, is the first it treats as critical for cybersecurity: able on its own to find security holes nobody knew existed and run the attack. Parts of the work are paused, the rest moved behind tighter controls. Watch the wording though. OpenAI says only that it cannot rule the capability out, and it is the sole judge of its own voluntary standard, with no outsider having seen the numbers. If your security rests on attacks being slow and expensive, that is the assumption going away.
AI News09 Aug 2026
The US found the hole in its own chip ban
Washington has spent three years keeping its best AI chips out of China. American officials are now mapping how Chinese labs get the hardware anyway, and the answer turns out to be dull: they rent it, in data centres in Thailand and Malaysia, entirely legally. Moonshot, the Chinese lab behind the Kimi K3 model, rented around 20,000 Nvidia chips from Alibaba and trained on those. One estimate puts the extra computing power China gets this way at 60% above what the rules intended. The ban policed the boxes, and nobody was watching the rental market. If you were counting on export controls to hold Chinese AI back, count again.
AI News09 Aug 2026
Alibaba starts charging for the free Qwen model
Plenty of companies have quietly built products on free Chinese AI models, on the reasonable assumption that free means free. Alibaba's Qwen is the most downloaded of them all, ahead of Meta's. Reuters reports that the next version arrives with a condition: if you make money reselling the model as a service, Alibaba wants a share. Run it inside your own business and nothing changes. Put it in something you invoice for and there is a new line in the budget. Moonshot attached the same clause to Kimi K3, the same Moonshot renting chips in the story above. The free years bought market share, and the billing starts now.
AI News09 Aug 2026
AMD buys the startup that burns AI models into the chip
Every AI model you use today runs on general-purpose hardware, hauling the model in and out of memory millions of times a second. Taalas, a three-year-old Toronto startup, casts a finished model straight into the transistors instead, so the chip runs that one model and runs it very fast. AMD bought the company this week and wants the technology inside its big rack systems next year. Worth a note of caution: the speed claims are Taalas's own, measured on a small model. And a chip that runs only one model becomes scrap the day you change models.
AI News09 Aug 2026
Meta approved ads carrying AI-generated child abuse imagery
Wired and the Tech Transparency Project, a US research group that audits the big platforms, found more than 50 paid adverts on Facebook, Instagram, Messenger and Threads carrying AI-generated child sexual abuse imagery. Several pointed to apps that digitally undress people in photos. They ran from November until early August, reached thousands of accounts across the US, UK and Europe, and Meta approved every one of them and took the money. Meta has removed them and blames older systems. The part worth sitting with is that automated review is now the only thing standing between a paying advertiser and your feed.