M-Pesa commissions low, BGP routes faked, Starlink V3

The UK AI Security Institute and US Center for AI Standards and Innovation jointly assessed Moonshot AI's Kimi K3, the Chinese open-weight model that claimed near-frontier performance at a fraction of the cost. On cyber exploit development tasks, Kimi K3 scored 32%, far behind frontier models at 76%. The assessment suggests Kimi K3's impressive general benchmarks do not translate to practical offensive capability, though the model still assisted with basic exploit scaffolding. The findings are relevant as Kimi K3's weights are scheduled for public release on July 27.

The ARC-AGI-2 leaderboard update for July 2026 shows GPT-5.6 Sol at 92.5%, leading Claude Opus 5 at 90.4% and GPT-5.5 at 85%. ARC-AGI measures abstract visual reasoning designed to resist rote memorisation, making it a useful gauge of genuine model improvement. A separate ARC-AGI-3 leaderboard tracks agentic task completion under budget constraints, where the gap between models is narrower. The scores suggest frontier models are still improving on reasoning tasks, but the gap between the top two has shrunk to 2.1 percentage points.

Dan Williams, a long-time Linux kernel contributor and maintainer of the libnvdimm subsystem and other core infrastructure, passed away on July 21. Colleagues Dave Hansen and Thomas Gleixner published an obituary remembering both his technical contributions and his character. Williams was known for his collaborative approach, deep knowledge of memory management and persistence, and his willingness to mentor newer contributors. His death is a significant loss for a kernel community that has been reckoning with the sustainability of its maintainer base.

A workshop concept called 'Avoiding AI' is spreading virally through US public libraries, drawing attendees frustrated with AI being pushed into every corner of digital life. The workshops teach practical strategies for opting out of AI features, avoiding AI-generated content, and maintaining analogue workflows. Organisers say demand is unprecedented, with sessions filling within hours. The trend signals a growing cultural countercurrent to Big Tech's relentless AI push, though sceptics argue it is a niche reaction that will not slow enterprise or consumer adoption.

OpenAI's first piece of hardware, the Codex Micro, is a $230 keypad built in collaboration with Work Louder. It features a joystick, programmable buttons, and RGB feedback that maps to Codex agent states, letting developers control prompts, accept code, switch reasoning modes, and manage multiple AI agents without touching a mouse. The device is genuinely useful for heavy Codex users but will mystify anyone outside that workflow. It is the latest signal that AI tool makers are moving beyond software into physical interfaces for power users.