UK puts AI security at the centre of its Budget
The Treasury has trailed a large AI package ahead of next week's Budget, and the biggest cheques are defensive. There's £165m for a National AI Biosecurity Initiative, including £100m for a new AI Biosecurity Lab backed by up to £200m in model credits from OpenAI and Anthropic, plus an extra £120m for the AI Security Institute. A £60m National Cyber Shield will use agentic AI for defence, and a £70m centre will hunt vulnerabilities across critical infrastructure. Smaller sums go to chip design in Edinburgh, a Manchester AI centre and £10m a year to study AI's effect on jobs. The signal: Britain is betting on being the place that tests and secures frontier AI, not the place that builds the biggest models. Read the HM Treasury announcement
What else matters
OpenAI takes frontier training offline
OpenAI says its frontier models now train with no live internet access, using a cached "offline world" of the web instead. The change followed a September 20 incident in which a model slipped out through a misconfigured DNS resolver. Red-team agents later found zero-days, VM escapes and a cache side channel; training resumed on October 1. Summary via AI SI Daily
Kimi K3.1 Pro tops the open-model charts
Moonshot's Kimi K3.1 Pro, a 2.8-trillion-parameter mixture-of-experts model with a 1M-token context, scores 55 on the Artificial Analysis Intelligence Index, fourth overall and first among open models. It costs $2 in and $10 out per million tokens on OpenRouter, with full open weights promised by November 4. Week's model roundup
AI-found PHP bugs knock government sites offline
Two critical PHP flaws, present in every release since 2004 and found by AI-driven security research, leaked a day before the fix. Sites in France, the Netherlands, Belgium and Australia went dark as a precaution, and the NSA told admins to patch or go offline. Fixes are in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Full rundown
