<h2>Must know</h2>
<ul>
<li><strong>AI agents are still struggling with real-world tasks</strong>: A new benchmark, Apex-Agents, from Mercor, reveals that leading AI models like Gemini 3 Flash and GPT-5.2 achieve only 24% accuracy on complex white-collar tasks involving multi-domain reasoning and common workplace tools. This suggests AI agents can’t reliably replace knowledge workers in complex, multi-tool tasks, despite year-over-year improvements. Meanwhile, Anthropic’s performance engineering team is constantly updating its technical evaluations because their Claude AI models keep acing tests designed for human engineers, forcing them to create new puzzles that emphasize human reasoning over pattern-matching. <a href="https://app.daily.dev/feed-by-ids?id=OsmrZ9UpM&amp;id=vA7gH0M1s" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New AI-driven security threats emerge, targeting developers and critical infrastructure</strong>: North Korean threat group KONNI is now using AI-generated PowerShell backdoors in phishing campaigns targeting software developers and blockchain teams in APAC, as reported by Check Point Research. These sophisticated attacks leverage obfuscation and legitimate remote management tools to compromise development environments. Simultaneously, Microsoft reports that SharePoint is being exploited in multi-stage phishing campaigns against the energy sector, where attackers use compromised accounts to send hundreds of phishing emails and maintain access even after password resets by tampering with MFA. <a href="https://app.daily.dev/feed-by-ids?id=wO1HB29Wl&amp;id=UyatbhpcD" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenAI Codex integrated into JetBrains IDEs, but prompt injection remains a core vulnerability</strong>: OpenAI Codex, powered by GPT-5.2-Codex, is now natively integrated into JetBrains IDEs like IntelliJ and PyCharm, offering configurable autonomy levels and ‘reasoning budgets’ for tasks like refactoring and security vulnerability detection. However, security expert Bruce Schneier warns that large language models (LLMs) are intrinsically vulnerable to prompt injection attacks due to their architecture, which processes trusted commands and untrusted data through a uniform channel. This fundamental flaw makes robust resistance to prompt injection unattainable without foundational breakthroughs in AI, especially as LLMs gain autonomous agent capabilities. <a href="https://app.daily.dev/feed-by-ids?id=uAvdGTErT&amp;id=LitPEfQIM" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Rails 8 solidifies its position as the go-to MVP framework</strong>: Ruby on Rails 8 continues to be a top choice for Minimum Viable Product (MVP) development, offering built-in authentication, background jobs via Solid Queue, caching with Solid Cache, and WebSocket support through Solid Cable, all without external dependencies. <a href="https://app.daily.dev/feed-by-ids?id=w2WfkMTD2&amp;id=w2WfkMTD2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Ruby gem ‘libgd-gis’ pivots to advanced satellite imagery analysis</strong>: The Ruby gem ‘libgd-gis’ has shifted its focus from cartographic rendering to direct satellite imagery analysis, enabling developers to analyze NASA GIBS pre-projected satellite data for environmental observation and pattern detection. <a href="https://app.daily.dev/posts/tm7hnZKgE" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub releases multi-platform SDK for Copilot agent integration</strong>: GitHub has launched a multi-platform SDK for Python, TypeScript, Go, and .NET, allowing developers to integrate Copilot’s agentic workflows, including orchestration, planning, and programmatic file edits, into their applications. <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>Feast integrates with PyTorch ecosystem for enhanced feature management</strong>: Feast, an open-source feature store, has joined the PyTorch Ecosystem to provide a unified API for consistent feature management across AI model training and deployment, aiming to eliminate training-serving skew in production AI systems. <a href="https://app.daily.dev/posts/dPG1hvjH4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New AI-DLC methodology puts AI in the driver’s seat for software development</strong>: A new AI-Driven Development Lifecycle (AI-DLC) methodology proposes AI as the primary orchestrator of development workflows, initiating tasks, decomposing requirements, and generating code, with humans validating and approving decisions. <a href="https://app.daily.dev/posts/SNNDpvqm6" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Datadog launches automatic retention filters for RUM and APM data</strong>: Datadog has released cross-product retention filters that automatically retain Application Performance Monitoring (APM) traces linked to Real User Monitoring (RUM) sessions, unifying frontend and backend observability data for improved troubleshooting. <a href="https://app.daily.dev/posts/fs5040Sq6" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel Sandboxes introduces filesystem snapshot feature</strong>: Vercel Sandboxes now supports filesystem snapshots, allowing users to capture and reuse the complete state of a running sandbox environment, streamlining workflows for parallel testing and rapid prototyping. <a href="https://app.daily.dev/posts/Uv3Zx9otG" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI agents are the new ‘cowboy sysadmins’ of IT operations</strong>: Apparently, AI agents are introducing ‘non-deterministic actions’ into IT operations, which is just a fancy way of saying they’re making things unpredictable. The article warns against letting AI directly modify live production systems, comparing it to the ‘cowboy’ sysadmins of old who made unreproducible changes. So, after all the hype about AI automating everything, we’re back to needing human oversight and version control for every little thing. Progress! <a href="https://app.daily.dev/posts/U8PYVjc73" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-powered scheduling: because your calendar wasn’t complicated enough</strong>: Blockit, an AI-powered scheduling platform, just launched with $5 million in Sequoia-backed seed funding, promising to use AI agents to ‘directly negotiate’ between users’ calendars. It learns your meeting preferences and prioritizes based on email tone. Because what we really needed was an AI that can read our passive-aggressive emails and decide which meetings are important. And at $1,000 a year for individuals, it better make my coffee too. <a href="https://app.daily.dev/posts/R9REfjqEf" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-driven QA analytics: because 13 years of data makes it smart, not just big</strong>: Testlio just launched LeoInsights, an AI-based QA analysis solution ‘trained on 13 years of Testlio’s testing data’ and ‘over 2.6 million test cases.’ It promises executive-level summaries and a ‘value calculator’ to measure efficiency. Because nothing says ‘actionable business insights’ like an AI that’s been fed a decade of bug reports and now tells you how much money you’re saving. I’m sure it’s not just a fancy dashboard for the same old metrics. <a href="https://app.daily.dev/posts/xZI6Dgtwy" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
