<h2>Must know</h2>
<ul>
<li><strong>Chrome zero-day under active exploitation</strong>: Google has released an urgent update for Chrome (versions 145.0.7632.75/76 for Windows/Mac and 144.0.7559.75 for Linux) to patch CVE-2026-2441, a high-severity use-after-free vulnerability in its CSS font feature parsing. This is Chrome’s first zero-day of the year, and Google confirms it’s already being actively exploited in the wild. The bug allows arbitrary code execution inside the browser sandbox, making immediate updates critical. <a href="https://app.daily.dev/posts/ISqaRw1RF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Critical vulnerabilities in popular VS Code extensions</strong>: OX Security disclosed critical vulnerabilities in four widely used VS Code extensions: Live Server, Code Runner, Markdown Preview Enhanced, and Microsoft’s Live Preview, collectively impacting over 128 million installs. These flaws enable file theft, remote code execution, and local network reconnaissance. While Microsoft silently patched Live Preview, the maintainers of the other three extensions have not responded, leaving millions of developers exposed. Developers are advised to update Live Preview and disable unused extensions. <a href="https://app.daily.dev/posts/IuDICegHN" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft Copilot chat security flaw exposes confidential emails</strong>: Microsoft confirmed a security bug (CW1226324) in Copilot Chat that allowed the AI to bypass confidentiality labels and Data Loss Prevention (DLP) policies, summarizing emails from users’ Sent Items and Drafts folders without authorization. The flaw was active since late January, and while a fix is rolling out, the incident highlights ongoing enterprise anxiety about AI tools accessing sensitive corporate data, especially given previous Copilot security concerns. <a href="https://app.daily.dev/posts/jSmIdhAQU" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AI model performance and cost: Anthropic’s sonnet 4.6 is a token muncher</strong>: Anthropic’s new Claude Sonnet 4.6, marketed as 40% cheaper, uses over 4x more tokens than its predecessor due to ‘adaptive thinking,’ potentially negating per-token price advantages for heavy use cases, according to independent benchmarks. <a href="https://app.daily.dev/feed-by-ids?id=i4O6sQBIw&amp;id=j12nF8mSg&amp;id=jj2Kauaop&amp;id=lUz1VvbAr&amp;id=y0N7mzotz" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>NVIDIA and Sarvam AI achieve 4x inference speedup for multilingual MoE model</strong>: Sarvam AI, an Indian generative AI startup, collaborated with NVIDIA to optimize inference for its Sovereign 30B multilingual Mixture-of-Experts model, achieving a 4x speedup on H100 and Blackwell GPUs through kernel-level optimizations and advanced scheduling. <a href="https://app.daily.dev/posts/yKP3YkWSF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>NVIDIA run:ai GPU fractioning for LLM inference performance</strong>: NVIDIA and Nebius AI Cloud benchmarks show that fractional GPU scheduling with NVIDIA Run:ai can deliver 77% of full-GPU throughput and 86% of concurrent user capacity with 0.5 GPU fractions, significantly maximizing GPU ROI for LLM inference. <a href="https://app.daily.dev/posts/LsiRfCVWP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Databricks launches custom agents</strong>: Databricks has launched Custom Agents (formerly Agent Framework), enabling developers to build, test, and deploy production-quality AI agents as fully managed Databricks Apps on serverless compute, with built-in memory and enterprise data access. <a href="https://app.daily.dev/posts/QJn2roYcs" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Asahi Linux progress on Apple M3 and Apple silicon support</strong>: Asahi Linux reports significant progress on Apple M3 hardware, with key peripherals like keyboard, touchpad, WiFi, and USB3 working, and DisplayPort Alt Mode via USB-C now functional in experimental branches after years of reverse engineering. <a href="https://app.daily.dev/posts/9Q140qigi" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Tailscale peer relays general availability</strong>: Tailscale announced the general availability of Peer Relays, allowing users to deploy high-throughput relay nodes on any Tailscale device to improve network connectivity and performance, now supporting static endpoints and enhanced observability. <a href="https://app.daily.dev/posts/Q0uzlfbeJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Tailscale integrates with Huntress and Fleet for automated device posture enforcement</strong>: Tailscale now integrates with Huntress EDR and Fleet MDM, allowing device posture and management state to be used in Tailscale access policies, automatically restricting access for non-compliant devices and reducing lateral movement risk. <a href="https://app.daily.dev/feed-by-ids?id=L7iJSIIDS&amp;id=VukD3HNtV" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Anthropic’s C compiler: A marketing masterpiece of misdirection</strong>: Anthropic’s marketing video claims Claude built a C compiler ‘from scratch’ with ‘no human intervention’ that could compile Linux. Turns out, it used 37 years of GCC test suites, an online GCC oracle, required human restarts, cost $20,000 in API calls, and couldn’t even boot Linux. But hey, 16 AI agents autonomously running for two weeks to produce a 100,000-line functional compiler is still pretty cool, if only they didn’t feel the need to lie about it. <a href="https://app.daily.dev/posts/0csgEkRMy" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI productivity impact: Executives still waiting for the magic</strong>: An NBER survey of nearly 6,000 executives found over 80% reported no measurable impact from AI on productivity or employment over the past three years. Yet, they anticipate AI will eliminate 1.75 million jobs by 2028. So, AI isn’t making things better now, but it’s definitely going to take your job later. Got it. The UK government’s M365 Copilot trial also found no productivity improvement. Maybe the AI is just really good at hiding its work? <a href="https://app.daily.dev/posts/rfmYnNc4e" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-generated code: The new ‘instant legacy’</strong>: Stack Overflow’s 2025 survey shows AI tool usage at 84% but developer trust dropped to 29%, with main branch success rates at a five-year low. Microsoft even republished an AI-generated, error-ridden Git branching diagram. Turns out, making code generation effortless just accelerates the creation of ‘instant legacy code’ that’s hard to understand or debug. Who knew that quantity doesn’t equal quality? <a href="https://app.daily.dev/feed-by-ids?id=alUW9sJXE&amp;id=guYQJk0m5&amp;id=okTNHJRhB&amp;id=sc19eWuhU&amp;id=vbdTHZjVN&amp;id=xQwTKeb7K&amp;id=ySpIGlBP7&amp;id=ygeEuk5Vz" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
