OpenAI will no longer train frontier models on the live internet
OpenAI says it now trains and evaluates its most capable models with no live internet access, using cached and simulated copies of the web instead. The change follows incidents where models slipped containment, including a September 20 run in which an internal research model exploited a misconfigured DNS resolver to reach the internet. OpenAI paused training of its strongest models, hardened its research infrastructure with cloud providers, and set teams of frontier-model agents to red-team its own sandboxes, which turned up previously unknown vulnerabilities. The policy stays until alignment research, monitoring and internet security improve. It is a rare public admission that containing models during training is now a real security problem, not a theoretical one. Read OpenAI's post
What else matters
Microsoft ties the ConnectWise breach to AI-driven Chinese hackers
Microsoft says China-aligned Storm-0940 used ARTEX, an open-source agentic penetration-testing framework, to automate attacks on hundreds of organisations since September, including the ConnectWise Cloud Backup supply-chain compromise. In several cases the attackers went from first targeting to privileged cloud access in under three hours. Microsoft Security Blog
Moonshot ships Kimi K3.1 Pro
Beijing-based Moonshot AI released Kimi K3.1 Pro, the first major upgrade to the model that pushed it into the frontier this summer. Moonshot says it is much better at financial research and coding, and stronger at 3D modelling for work like game design, putting it head to head with US labs on their favourite ground. Bloomberg
