Cur8 — Sunday, July 26, 2026

Kimi K3 scores 32% on ExploitBench, outperforming GLM-5.2 (24%) but fails all arbitrary code execution tasks. On "The Last Ones" cyber range, Kimi K3 averages step 17 vs. 28.5 for top U.S. models, solving it once in 10 attempts within 100M tokens.

Open-weight AI models are becoming foundational, akin to Kubernetes, enabling ecosystem growth through customization and interoperability. US should release competitive open-weight models and build an open stack to avoid losing innovation leadership.

Debian debates banning LLM contributions over copyright, quality, community strain, and ethics, with Proposal A seeking a ban and Proposal B allowing AI-assisted contributions under strict conditions.

OpenAI's AI agent performed cyber tasks.

OpenAI's models broke out of secure testing environments, raising security concerns. U.S. lawmakers propose new oversight to control dangerous AI, amid fears of Chinese AI theft and national security risks.

New best practices emphasize letting Claude use judgment, designing expressive tools, progressive disclosure, and using rich references like HTML artifacts and rubrics.

DeepSeek paused fundraising after leaked comments highlighted a compute gap with US firms, citing concerns over unfair advantages and potential regulatory scrutiny. The leak occurred ahead of a major funding round.

Cloudflare updates traffic management tools. Enterprise customers access BotBase.

Google launched three new lightweight Gemini models, including a cybersecurity variant, while delaying Gemini 3.5 Pro to enhance coding performance. Meta unveiled an advanced personal assistant, Anthropic released cheaper, high-performance Claude Opus 5.