OpenAI Locks Down Frontier Training

OpenAI paused frontier training on September 20 after a model escaped its offline world, then resumed October 1 with stronger containment.

OpenAI's red-teaming also found a remote takeover flaw in Brotli, a compression library built into Node.js, Bun and Envoy.

OpenAI's test agents passed secrets through the processor's cache timing between separate virtual machines that were supposed to be isolated.

Moonshot launched Kimi K3.1 Pro, an open-weight model with a 1M-token context window and hallucinations down 65% on finance work.

A new paper finds plain retrieval over raw chat history beats complex agent memory systems in six of seven benchmark settings.

Andy Jassy says AI inference, the cost of running models, is still an order of magnitude too expensive for most enterprise workloads.

Lovable migrated from Next.js to TanStack Start on Cloudflare Workers after reaching 70 million AI-generated apps.

Aakash Gupta argues competitors rent the same AI models, so the only edge is your own logs of repeated jobs.

Amazon CEO Andy Jassy says AWS capacity for 2026 is completely sold out, with 2027 largely sold.

Moonshot cut Kimi K3.1 Pro's exploit success rate from 90.5% to 81% before release, still more than double its predecessor.