## 🔍 Key Insights **Anthropic** has publicly accused **DeepSeek**, **Moonshot AI**, and **MiniMax** of carrying out '**industrial-scale distillation attacks**,' igniting widespread discussion on AI model security and the boundaries of intellectual property. Concurrently, the industry is rapidly advancing toward **AI Agent engineering**: practical initiatives—including **OpenAI's Codex App**, **TinyFish's $2M seed fund for AI Agents**, and the **Claude Code + Obsidian personal operating system**—are reshaping software development paradigms. ## 🚀 Top Updates - **Anthropic accuses domestic LLM vendors of industrial-scale distillation attacks**: Alleges that DeepSeek, Moonshot AI, and MiniMax systematically distilled core capabilities of **Claude** using tens of thousands of fraudulent accounts—highlighting critical gaps in model training ethics and defensive safeguards. - **OpenAI accelerates infrastructure investment in agent development**: Recruited **Cursor's** core engineer **Rohan Varma**, and aggressively promoted its **Codex App**, widely hailed by the developer community as the best-performing agent-native IDE to date. - **TinyFish × Dify launch a $2M AI Agent seed fund**: Dedicated to supporting early-stage AI Agent startups with capital, cloud resources, and hands-on engineering mentorship—focused on accelerating the full lifecycle from prototype to production. - **OpenClaw's security model is well-defined but accident-prone**: While emphasizing single-user sandboxing and mandatory human approval, it recently suffered a permissions failure that led to the **accidental deletion of Meta AI's Head of Security's email**, underscoring the urgent need for robust Agent permission governance. - **AI-assisted development shifts to an 'iteration-first' quality paradigm**: Moving away from lengthy documentation, teams now prioritize rapid validation via a '**minimum viable runnable version**,' replacing static specifications with high-frequency feedback loops. - **OpenAI deprecates SWE-bench Verified benchmark**: Citing **training data contamination** and flaws in test design, OpenAI recommends upgrading to the more rigorous **SWE-bench Pro**. - **Notion launches an 'AI-First Playground'**: Enables non-technical roles (e.g., PMs and designers) to safely experiment with code logic in isolated environments—significantly lowering the barrier to entry for AI Agent adoption. - **Building a 24/7 personal operating system with Claude Code + Obsidian**: By injecting structured notes to provide long-term memory and personal context, users create an AI thinking partner capable of sustained, context-aware reasoning.