<h2>Must know</h2>
<ul>
<li><strong>AI agents are still struggling with complex tasks</strong>: A new benchmark, Apex-Agents, reveals that leading AI models like Gemini 3 Flash and GPT-5.2 achieve only 24% accuracy on complex white-collar tasks requiring multi-domain reasoning and tool use. This suggests current AI agents are far from reliably replacing knowledge workers, despite rapid year-over-year improvements, according to Mercor. Meanwhile, a developer found that AI-generated SwiftUI code caused memory leaks due to improper closure capture, highlighting the need for human oversight even with AI assistance, as reported by Rebeloper.</li>
<li><strong>OpenAI’s Grok faces child safety criticism and EU investigation</strong>: xAI’s Grok chatbot is under fire from Common Sense Media for child safety failures, including generating inappropriate content and discouraging mental health support. This has prompted a Digital Services Act investigation by the European Commission, which could lead to significant fines for X. In response, X has disabled Grok’s image generation feature for non-paying EU users. Separately, the Tech Transparency Project found 55 apps on Google Play and 47 on Apple’s App Store that create non-consensual AI-generated nude images, with both companies profiting before removing some apps after inquiries.</li>
<li><strong>React Server Components hit with DoS vulnerability</strong>: Multiple high-severity denial-of-service vulnerabilities (CVE-2026-23864) have been found in React Server Components, affecting React 19.0.x through 19.2.x and Next.js 13.x through 16.x. These flaws can lead to server crashes or excessive CPU usage via specially crafted HTTP requests. Vercel has deployed WAF mitigations, and official fixes are available, urging immediate upgrades to patched versions.</li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AI-generated malware emerges with VoidLink</strong>: Check Point Research identified VoidLink, the first fully AI-generated advanced malware, created by a single developer using TRAE IDE and AI agents, producing over 88,000 lines of code in under a week.</li>
<li><strong>AI coding tools impact open-source sustainability</strong>: Open-source projects like Tailwind CSS are experiencing revenue decline and staff layoffs due to developers using AI tools instead of visiting documentation sites, threatening the sustainability of open-source ecosystems.</li>
<li><strong>New Linux kernel GPU driver fixes and performance gains</strong>: Mesa 26.1 brings significant updates for Intel ANV and i915 graphics drivers on Linux, including a DriConf workaround for faster shader compilation in games and patches for DG2/Meteor Lake GPUs, yielding up to 260% performance improvements in some benchmarks.</li>
<li><strong>China’s open-source AI shifts to MoE and hardware focus</strong>: The Chinese open-source AI community is adopting Mixture-of-Experts (MoE) as a standard, focusing on cost-performance optimization, multimodal AI, smaller models, Apache 2.0 licensing, and a ‘hardware-first’ strategy for domestic AI chips.</li>
<li><strong>Microsoft and Google face digital sovereignty challenges</strong>: Satya Nadella redefines AI sovereignty as control over proprietary knowledge in AI models, while France plans to replace US videoconferencing platforms with a state-built alternative by 2027 due to data sovereignty concerns.</li>
<li><strong>New tools for AI-native software development</strong>: Atomic SCM and Circuit Breaker are emerging open-source tools designed for AI-generated code, offering change-centric version control and real-time validation to address the limitations of traditional Git workflows when dealing with machine-speed code changes.</li>
<li><strong>AI-powered UI component library launched</strong>: SpicyUI launched a free AI-powered UI component library for React and Next.js developers, featuring AI-generated components and a community feed for sharing and remixing.</li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI-generated code: the fast fashion of software?</strong>: The tech industry is apparently churning out AI-generated code like fast fashion, prioritizing speed and metrics over quality. <a href="http://Fast.ai" target="_blank" rel="noopener nofollow">Fast.ai</a> criticizes this ‘vibe coding’ as a false productivity wave, warning of hidden bugs, excessive complexity, and a decline in software craftsmanship. So, while your AI assistant might whip up some code in seconds, you might just be buying a digital garment that falls apart after the first wash, leaving you with a mountain of technical debt.</li>
<li><strong>Cursor’s AI-built browser: more spaghetti than code?</strong>: Cursor proudly announced their AI agents built a web browser with over 3 million lines of code and a custom Rust rendering engine. Sounds revolutionary, right? Well, developers who actually looked at the code are calling it ‘spaghetti code’ that fails to compile, performs poorly, and relies heavily on external dependencies, contradicting Cursor’s claims of building from scratch. It seems AI can generate a lot of code, but quality and originality are still very much human endeavors, despite the marketing hype.</li>
<li><strong>AI’s impact on jobs: more construction, less… everything else?</strong>: Tech CEOs like Nvidia’s Jensen Huang and Palantir’s Alex Karp are now telling us that AI’s infrastructure buildout is the ‘largest construction project in history,’ creating ‘six-figure’ jobs for electricians and construction workers. While it’s great news for the trades, it feels a bit like a distraction from the 6-10% of US jobs AI is projected to displace by 2030. So, while AI might be building a shiny new future, it seems many of us will be left wondering if our current skills are still relevant, or if we should all just pick up a hammer.</li>
</ul>
