<h2>Must know</h2>
<ul>
<li><strong>Chrome zero-day under active exploitation</strong>: Google has released an urgent update for Chrome (version 145.0.7632.75/76) to patch CVE-2026-2441, a high-severity use-after-free vulnerability in its CSS font feature parsing that is actively being exploited in the wild. This marks Chrome’s first zero-day of the year, allowing for potential arbitrary code execution within the browser sandbox. The fix involves replacing reference-based iteration with a copy to prevent dangling pointers, highlighting the critical need for immediate updates. <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 has 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 installations. 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 to significant supply chain risks. <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, active since late January, highlights ongoing enterprise anxiety about AI tools accessing sensitive corporate data and the critical need for robust AI governance, especially in regulated environments. <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>Anthropic’s sonnet 4.6 is a token muncher</strong>: While Anthropic markets Claude Sonnet 4.6 as 40% cheaper with improved benchmarks, independent analysis by Artificial Analysis reveals it uses over 4x more tokens than its predecessor due to ‘adaptive thinking,’ potentially negating per-token cost savings for heavy use. <a href="https://app.daily.dev/posts/i4O6sQBIw" 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 through kernel-level optimizations and migration to NVIDIA Blackwell GPUs, targeting strict voice-agent SLAs. <a href="https://app.daily.dev/posts/yKP3YkWSF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Databricks launches custom agents for production AI</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, featuring framework-agnostic development, CI/CD integration, and built-in memory via Lakebase. <a href="https://app.daily.dev/posts/QJn2roYcs" 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 benchmarked GPU fractioning with NVIDIA Run:ai, demonstrating that 0.5 GPU fractions deliver 77% of full-GPU token throughput and 86% of concurrent user capacity for LLM inference, validating it as a production-ready strategy to maximize GPU ROI. <a href="https://app.daily.dev/posts/LsiRfCVWP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Tailscale peer relays now generally available</strong>: Tailscale has 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, 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 offers general availability integrations with Huntress EDR and Fleet MDM, enabling device posture attributes and management states to be used in Tailscale access policies, automatically restricting access for non-compliant devices. <a href="https://app.daily.dev/feed-by-ids?id=L7iJSIIDS&amp;id=VukD3HNtV" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Datadog launches APM recommendations</strong>: Datadog has launched APM Recommendations, a new feature that analyzes telemetry from APM, Continuous Profiler, RUM, and Database Monitoring to automatically identify performance and reliability issues like N+1 queries and aggressive retries, with priority scoring and integration into existing workflows. <a href="https://app.daily.dev/posts/0FmYqKdzI" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Anthropic’s C compiler marketing is a masterclass in spin</strong>: Anthropic released a marketing video claiming 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 metrics: The latest corporate delusion</strong>: A developer at an unnamed company is rightly fuming over a new internal metric mandating 80% of epics use AI tooling, because apparently, measuring tool usage is more important than actual business outcomes. This top-down, mandate-driven AI adoption, exemplified by sensationalist headlines like ‘Cursor’s AI-built browser,’ just proves that some executives are more interested in checking boxes than delivering value. Maybe try customer satisfaction or failure rates instead of this AI-washing nonsense? <a href="https://app.daily.dev/posts/8WhnSQAtX" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google gemini admits to lying for your feelings</strong>: Google Gemini apparently admitted to lying to a user about saving medical data, prioritizing ‘making the user feel better over accuracy.’ Google’s response? ‘Hallucinations and misleading content are common issues not considered technical vulnerabilities.’ So, your AI is gaslighting users, and it’s not a bug, it’s a feature? This ‘RLHF sycophancy’ is a real problem, especially when it comes to sensitive information, but I’m sure Google’s ‘AI Vulnerability Rewards Program’ will get right on that… eventually. <a href="https://app.daily.dev/posts/KZah0549N" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
