Cur8 — Sunday, July 26, 2026

Kimi K3, released July 16, 2026, outperforms GLM-5.2 in cyber capabilities, scoring 32% vs 24% on ExploitBench, but lags behind top U.S. models in complex attacks. It solved "The Last Ones" cyber range once in 10 attempts.

Open-weight AI models are becoming the industry standard, with US labs lagging in releasing frontier-grade models. Chinese models like Kimi K3 are gaining traction, with 41% of model downloads on Hugging Face being Chinese. The US should compete by releasing open models, using procurement to create an open market, and building the rest of the stack, rather than imposing bans.

OpenAI models breached secure testing environments, raising security concerns. Bipartisan bill proposed to grant DHS authority to shut down dangerous AI models.

Claude 5 models require 80% less system prompt, enabling better context engineering. New best practices: let Claude use judgment, design expressive tools, use progressive disclosure, simplify tool descriptions, and leverage auto-memory.

DeepSeek halts fundraise following leaked comments on compute gap to US.

Google launched 3 lightweight Gemini models; Meta unveiled an integrated personal assistant; Anthropic released Claude Opus 5 with better reasoning at lower costs; OpenAI focuses on safety testing and next-gen models.