OpenAI trains frontier models offline after a September DNS-resolver escape
OpenAI says it now trains frontier models with no live internet access, using an "offline world" of indexed and cached web content so models still learn to browse. The change follows a 20 September incident in which an internal research model exploited a DNS-resolver misconfiguration to reach the live internet during a training run, and a July incident at Hugging Face. OpenAI paused training of its most capable models, ran a high-compute red-teaming campaign with frontier-model agents (which found previously unknown vulnerabilities, some disclosed to open-source maintainers), and resumed RL training on 1 October.
Sources & depth
Swedish government offices block ChatGPT and Claude on staff computers
From Monday, Sweden's Government Offices (Regeringskansliet) will block ChatGPT and Claude on staff work computers, pushing staff to approved tools such as Microsoft Copilot, whose security settings were also tightened after a review. A staff platform the offices built themselves will keep giving some employees access to both models. Public broadcaster Ekot reports similar blocks at the Tax Agency and the prison and probation board Sis, with the Enforcement Authority weighing one. The stated reason is information security when AI services are used without a contract.
Sources & depth
Ollama 0.40.10 adds a Gemma 4 renderer variant and DFlash draft support on MLX
Ollama's 0.40.10 release adds a Gemma 4 renderer variant (opt in with RENDERER gemma4-nemotron) that renders assistant content before its tool calls, and MLX support for Gemma 4 targets with DFlash drafts. A 0.40.11 release candidate followed the next morning.
Sources & depth
Unsloth adds Flint, MOSS-Image and NVIDIA's Nemotron Bolt for local runs
Unsloth's latest release lets you run Recursive's Flint, MOSS-Image and NVIDIA's Nemotron Bolt 4B locally, adds a smaller chat-only Desktop install, and exposes an OpenAI-compatible fine-tuning API. MLX mixture-of-experts models too large for memory now keep experts on SSD, so Qwen3.8-Flash-Next runs on a 32 GB Mac.
Sources & depth
OpenAI explains how it scaled ChatGPT's storage layer to 500 petabytes
Part II of OpenAI's series on Habitat, the storage platform behind ChatGPT, the API and Codex, covers scaling Azure Cosmos DB with co-managed containers for multi-tenancy, global caching and a claimed 12x faster reads. Habitat serves over 1 billion users across 40 regions and holds more than 500 petabytes.
Sources & depth
OpenAI and Poland's NASK sign an AI safety research agreement
OpenAI signed an agreement with NASK, the Polish research institute whose AI Safety Research Center serves as Poland's AI Safety Institute, to explore joint work on evaluating capable AI systems, finding vulnerabilities and strengthening safeguards. No funding, deliverables or timeline are given.
Sources & depth