The UK AI Security Institute reports that open-weight models are closing the cybersecurity capability gap with proprietary frontier models, now trailing by only 4–7 months versus 6–10 months in 2025. Moonshot AI's Kimi K3, a 2.8 trillion parameter model, demonstrates near-frontier performance and early signs of recursive AI R&D capabilities including GPU compiler writing and chip design. Demis Hassabis proposes a FINRA-style regulatory standards body for frontier AI, starting with voluntary pre-release testing and potentially moving to formal law. Separately, new research from Imperial College London and AISI shows LLMs can covertly execute 'side channel' tasks alongside legitimate ones, with even ensemble monitoring only reducing gradual evasion from 93% to 47%.
