<h2>Must know</h2>
<ul>
<li><strong>AI agents are still pretty dumb at real-world tasks</strong>: A new benchmark, Apex-Agents, from Mercor, tested leading AI models like Gemini 3 Flash and GPT-5.2 on complex white-collar tasks. Turns out, they only hit a maximum accuracy of 24%, nowhere near human-level performance. This means AI agents can’t reliably replace knowledge workers in tasks requiring multi-domain reasoning and tool use, despite year-over-year improvements. So, maybe hold off on firing your entire consulting team for now. <a href="https://app.daily.dev/posts/OsmrZ9UpM" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub releases SDK for Copilot agent integration</strong>: GitHub just dropped a multi-platform SDK for integrating Copilot’s agentic workflows into your applications, supporting Python, TypeScript, Go, and .NET. This SDK lets you embed the same engine used by the Copilot CLI, enabling orchestration, planning, tool invocation, and programmatic file edits. It’s currently in technical preview, so expect some rough edges, but it signals GitHub’s push to expand Copilot’s ecosystem, as reported by GitHub Changelog. <a href="https://app.daily.dev/feed-by-ids?id=GXXnGjqMp&amp;id=GuG7K2ryZ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Ireland wants to use spyware for surveillance</strong>: Ireland is pushing a new bill, the Communications (Interception and Lawful Access) Bill, to update its surveillance laws from 1993. This would let law enforcement deploy commercial spyware to intercept encrypted digital communications, promising judicial oversight and privacy safeguards. This comes as spyware use is on the rise across Europe, with documented abuses in countries like Greece and Hungary, according to various reports. <a href="https://app.daily.dev/posts/wO3SoInHG" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AI agents for IT operations: proceed with extreme caution</strong>: AI agents introduce non-deterministic actions into IT operations, which is basically a fancy way of saying they can go rogue, so it’s best to limit them to design-phase tasks and keep them away from direct production modifications. <a href="https://app.daily.dev/posts/U8PYVjc73" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New Linux kernel vulnerability for GPUs</strong>: A security flaw in the Linux kernel’s DRM graphics driver allows unprivileged users to hog kernel memory, potentially crashing the system, but a fix is apparently on the way. <a href="https://app.daily.dev/posts/9SIWTOgdA" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Feast joins PyTorch ecosystem for better feature management</strong>: Feast, an open-source feature store, has integrated with the PyTorch Ecosystem to help manage features consistently across AI model training and deployment, aiming to reduce training-serving skew. <a href="https://app.daily.dev/posts/dPG1hvjH4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Rust 1.93.0 released with musl update</strong>: Rust 1.93.0 is out, updating the bundled musl libc to 1.2.5, which improves DNS resolution for static Linux binaries and stabilizes 23 APIs, according to the release notes. <a href="https://app.daily.dev/feed-by-ids?id=wF6DW7Hp4&amp;id=O9g10PE7Q" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel Sandboxes now support filesystem snapshots</strong>: Vercel Sandboxes introduced filesystem snapshots, letting you save and reuse the complete state of a running sandbox, which should make parallel testing and experimentation much smoother. <a href="https://app.daily.dev/posts/Uv3Zx9otG" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ClickHouse launches managed Postgres service</strong>: ClickHouse just rolled out a new managed Postgres service, promising up to 10x faster performance for disk-bound workloads and native integration with ClickHouse for a unified transactional and analytical data stack. <a href="https://app.daily.dev/posts/WHTgaLabX" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Mozilla expands product portfolio and builder program</strong>: Mozilla’s New Products division is launching Solo (web presence), Tabstack (agentic experiences), 0DIN (AI vulnerability research), and an enterprise Firefox, alongside a new ‘Pioneers’ program for external collaborators. <a href="https://app.daily.dev/feed-by-ids?id=qHiMQ30IZ&amp;id=qHiMQ30IZ" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>OpenAI Codex in JetBrains IDEs: because who needs to learn to code anymore?</strong>: OpenAI Codex is now natively integrated into JetBrains IDEs, powered by GPT-5.2-Codex, offering adjustable autonomy levels and ‘reasoning budgets’ for refactoring, code migrations, and security vulnerability detection. JetBrains says it’s free for a limited time, because nothing says ‘we believe in our product’ like a free trial for something that might just write all your code for you. What could possibly go wrong when developers delegate large-scale tasks to an AI that’s ‘accessible through three authentication methods’? <a href="https://app.daily.dev/posts/uAvdGTErT" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google AI Mode: your personal data, now with extra AI creepiness</strong>: Google’s new AI Mode for AI Pro and AI Ultra subscribers lets the AI snoop through your Gmail and Google Photos to suggest vacation itineraries and shopping items. Google assures us it ‘does not train on full Gmail inboxes or photo libraries,’ only ‘specific prompts and responses.’ So, just a little bit of your personal life, for your convenience. What a bargain for a subscription service! <a href="https://app.daily.dev/posts/Kw6SK0Mqr" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agents for security: because 90% of vulnerabilities are just ‘noise’</strong>: ZEST Security launched AI Sweeper Agents that claim to identify which cybersecurity vulnerabilities are ‘truly exploitable,’ dismissing 11 million vulnerabilities in six months for early adopters. Apparently, over 90% of high and critical vulnerabilities are just ‘noise’ that traditional scanners can’t handle. So, instead of fixing the actual problems, we’re just going to let AI tell us they don’t exist. Problem solved, right? <a href="https://app.daily.dev/feed-by-ids?id=LDTl5G8Jm&amp;id=LDTl5G8Jm" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
