Full Summary
This Sunday morning, OpenAI has delayed the release of its advanced AI model, Astra, after internal testing revealed it could autonomously plan and execute cyberattacks. Both kobaran.com and Benzinga confirm this marks the first time a major AI lab has publicly slowed a model due to cybersecurity concerns, with evaluations showing Astra could reach the "Critical" tier in their Preparedness Framework. This means it could develop zero-day exploits or design end-to-end cyberattacks. The Hindu, Memeburn, and The Motley Fool all report that Chinese AI firm Moonshot released its Kimi K3 model on July 16th. This 2.8 trillion-parameter model is being called the "world's first open 3T-class model" and has already topped a major coding leaderboard, Arena.ai's Frontend Code Arena. Alibaba is following suit with its Qwen 3.8 Max, a 2.4 trillion-parameter model, and both companies plan to charge major commercial users, with Alibaba requiring a share of revenue from its model next week. South Korea's "Sovereign AI Foundation Model" project sees four teams, including LG AI Research with its 750 billion-parameter K-EXAONE 2.0, competing in its second stage. Meanwhile, xAI's Grok 4.6 is expected next week, with leaked information suggesting major reasoning upgrades and an emphasis on agentic tasks, all without increasing its 1.5 trillion-parameter V9 base model size. This surge in powerful, and increasingly autonomous, AI models means that governments and companies are scrambling to establish evaluation procedures before they reach the market, directly impacting the security and functionality of the AI tools you might use daily.