<h2>Must know</h2>
<ul>
<li><strong>AI agents are still a security nightmare</strong>: The security risks of AI agents continue to escalate, with new vulnerabilities emerging. Google Gemini was found susceptible to prompt injection via calendar invites, allowing private data exfiltration, as reported by Security Boulevard. ServiceNow also patched a critical flaw (CVE-2025-12420) in its AI agents that allowed unauthenticated user impersonation and admin takeover, highlighting the dangers of rushed AI deployments, according to InfoQ. These incidents underscore that AI agents, with their elevated privileges and access to sensitive systems, are becoming prime targets for attackers, demanding a zero-trust approach and rigorous security reviews. <a href="https://app.daily.dev/feed-by-ids?id=dHjNBMkFO&amp;id=wzAqd15ET&amp;id=M77gEo1eY&amp;id=j6bBmLvCW&amp;id=mZZUloj80" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic’s AI agents have a persistent file exfiltration problem</strong>: Anthropic’s Claude Code and its new desktop automation tool, Cowork, are plagued by an unpatched prompt injection vulnerability that allows attackers to exfiltrate sensitive files. Researchers at PromptArmor found that malicious prompts embedded in files can bypass user approval and send data to an attacker’s Anthropic account via the Files API. Despite repeated disclosures, Anthropic places the responsibility on users to monitor for suspicious activity, a stance criticized as inadequate, especially for non-technical users targeted by Cowork, as detailed by InfoQ and other security outlets. <a href="https://app.daily.dev/feed-by-ids?id=s11rN52dz&amp;id=jaKSkIh92&amp;id=rAHVvRPEj" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Small language models are eating the enterprise AI world</strong>: Enterprises are increasingly favoring Small Language Models (SLMs) over Large Language Models (LLMs) due to their lower cost, reduced latency, and better fit for specific business tasks like invoice parsing or contract classification. Benchmarks show SLMs like Phi-3 and Mistral 7B matching or exceeding LLM performance on constrained tasks, often being 10,000 times smaller. This shift is driving a ‘digital factory’ model where specialized SLMs handle routine processes, escalating only complex queries to LLMs, making AI deployment more scalable and affordable, according to The New Stack. <a href="https://app.daily.dev/feed-by-ids?id=HX43N3MkE&amp;id=J1BKrq2u6" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>GitHub Copilot gets agentic memory and new models</strong>: GitHub Copilot is rolling out a ‘memories’ feature that learns and documents coding preferences and project guidelines, aiming to improve code consistency and speed up onboarding. It also introduced GPT-5.2-Codex for all users and new models (GPT-5 mini, GPT-4.1) for its CLI, while deprecating older models like Claude Opus 4.1. <a href="https://app.daily.dev/feed-by-ids?id=jz4TUoMzD&amp;id=nCmOnsQS3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI coding agents are making developers into product engineers</strong>: AI coding agents are drastically reducing implementation time, shifting the bottleneck in software development from execution to problem definition. This is creating a new ‘Product Engineer’ role, where technical skills are combined with a deep understanding of customer problems, strategic thinking, and cross-functional collaboration, as highlighted by Atlassian. <a href="https://app.daily.dev/feed-by-ids?id=ruSCTD0Fh&amp;id=sifgyKGyd" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google previews AI-based SQL generation in BigQuery</strong>: Google is previewing ‘Comments to SQL’ in BigQuery, a feature that uses AI to convert natural language instructions within SQL comments into executable SQL queries, aiming to accelerate data analysis by minimizing manual coding for complex queries. <a href="https://app.daily.dev/posts/5AcBUpc2i" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Android Studio Otter 3 integrates flexible AI and enhanced agent workflows</strong>: Android Studio Otter 3 now supports integration with multiple LLMs (Gemini, GPT, Claude, local models) and features an enhanced agent mode for direct device interaction, natural language UI testing, and remote Model Context Protocol (MCP) server support for external tools like Figma. <a href="https://app.daily.dev/posts/q3jJpSTPR" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ClickHouse raises $400M, acquires Langfuse, targets unified AI analytics</strong>: ClickHouse raised $400 million and acquired Langfuse, an open-source LLM observability platform, aiming to create a unified analytics and observability platform for AI agents. This move positions ClickHouse to compete with Snowflake and Databricks in the ‘operational AI platforms’ market. <a href="https://app.daily.dev/posts/9QbO7XqRY" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Intel expands LLM-Scaler support for generative AI</strong>: Intel released updates to LLM-Scaler-Omni and LLM-Scaler-vLLM, improving support for ComfyUI, SGLang, PaddleOCR, and new AI models like Qwen3 and GLM-4.6v-Flash on Arc Graphics, expanding its generative AI hardware-acceleration ecosystem. <a href="https://app.daily.dev/posts/XBX3pzO5M" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft open sources XAML Studio for .NET UI prototyping</strong>: Microsoft has open-sourced XAML Studio under the .NET Foundation, allowing rapid prototyping and fast feedback loops for XAML-based UI development, with features like a Monaco-based code editor for live XAML rendering and enhanced integration with WinUI 3. <a href="https://app.daily.dev/posts/hUJrXK08w" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI coding tools: functionally correct, architecturally unsound</strong>: Apparently, AI-generated code is ‘functionally correct’ but still a hot mess architecturally, especially for Android and SwiftUI. Reviewers found issues like improper coroutine handling, state management errors, and excessive view re-renders. So, while your AI might churn out code that ‘works,’ it seems you still need a human to prevent your codebase from becoming a future maintenance nightmare. Who knew AI couldn’t intuit good design? <a href="https://app.daily.dev/feed-by-ids?id=O3swuQWEt&amp;id=Oj1MubZPP&amp;id=QjH25jlip" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Razer CEO defends AI investments, because of course he does</strong>: Razer’s CEO, Min-Liang Tan, is out here defending his company’s $600 million AI investment, claiming gamers reject ‘generative AI slop’ but value AI tools that enable ‘better game development.’ He’s trying to thread the needle between gamer skepticism and investor hype, all while acknowledging ‘mental health concerns’ and ‘unresolved trust and safety issues’ with Grok, the underlying model for their ‘Project Ava’ AI companion. It’s almost as if the guy whose company profits from AI wants us to believe AI is good, despite all evidence to the contrary. <a href="https://app.daily.dev/posts/bokhtiGjg" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cursor’s ‘success’ in building a browser: a triumph of marketing over reality</strong>: Cursor proudly announced its AI agents built a web browser from scratch in a week, generating over a million lines of code. Impressive, right? Except the resulting codebase doesn’t compile, has 34 errors and 94 warnings, and no commit actually works. It seems ‘meaningful progress’ and ‘screenshots’ are the new ‘working software.’ Perhaps they should focus less on generating code and more on generating code that actually runs. <a href="https://app.daily.dev/feed-by-ids?id=V3KHuMgUU&amp;id=GPLZOgOgP" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
