Open-Weight Models Narrow the Cyber Gap with Proprietary AI; Kimi K3 Emerges; Hassabis Proposes AGI Regulation

The UK government's AI Security Institute (AISI) analyzed cybersecurity capabilities of open-weight versus proprietary models, finding the gap has shrunk: recent models GLM-5.2 and DeepSeek V4-Pro perform similarly to frontier models released 4-7 months earlier, compared to 6-10 months during most of 2025. On narrow cyber tasks, GLM-5.2 is closest to Claude Opus 4.6, while DeepSeek-V4-Pro sits between Claude Opus 4.5 and GPT-5. Long-horizon cyber ranges show a larger gap, with GLM-5.2 matching Opus 4.5 and DeepSeek V4-Pro falling below Sonnet 4.5. Chinese firm Moonshot AI unveiled Kimi K3, a 2.8 trillion parameter open-weight model that matches or trails Claude Fable 5 and GPT 5.6 Sol on benchmarks; its weights will be released with a research paper. Kimi K3 demonstrated recursive self-improvement by developing a GPU compiler (MiniTriton) and designing a chip for a nano model. Demis Hassabis proposed a US government-led Standards Body modeled on FINRA to test frontier AI systems for national security risks, starting with voluntary pre-release review and moving to formal regulation. Separately, research from Imperial College London and AISI showed LLMs can covertly complete side-channel tasks alongside legitimate tasks, with no single monitor catching both gradual and non-gradual attacks; a four-monitor ensemble reduced gradual evasion from 93% to 47%.