<p><strong>TLDR:</strong> Cursor launched a new coding model without disclosing it was built on Kimi k2.5 from Moonshot AI, and the co-founder later admitted that was a mistake. Alibaba confirmed it’s keeping the Qwen and Wan series open source after internal team turbulence. Claude Code’s skills ecosystem is getting serious attention today, with a large-scale security audit finding command execution patterns in 27% of public skills. Stripe’s autonomous “Minions” agents are generating thousands of pull requests weekly, which is either impressive or alarming depending on your tolerance for unsupervised agents.</p>
<hr>
<h2>Cursor’s Kimi base</h2>
<p>Cursor shipped Composer 2 as a frontier-level coding model without mentioning it was built on Kimi k2.5, an open-source model from Chinese AI company Moonshot AI. A user identified the base quickly, and Cursor’s co-founder confirmed it while noting roughly three-quarters of the training compute was Cursor’s own. Moonshot confirmed the use was an authorized commercial partnership via Fireworks AI. The omission matters beyond optics given the ongoing sensitivity around U.S. companies building on Chinese AI models.</p>
<h2>Claude Code skills security audit</h2>
<p>A security audit of 22,511 public AI coding agent skills found 140,963 security findings across four registries. One in six skills contained curl-pipe-bash remote code execution patterns, 27% had command execution patterns, and researchers found API traffic hijacking to third-party servers and hidden Unicode steganography. The structural problem: registries scan at publish time, but installed skills execute with full developer permissions and no runtime verification. There’s no npm-audit equivalent for this ecosystem yet.</p>
<h2>Alibaba doubles down on open source Qwen</h2>
<p>After a turbulent shake-up in its large model team, Alibaba confirmed it’s keeping the Qwen and Wan series open source. This matters because Qwen 3.5 27B is already running on a single 3090 at 50 tokens per second and handling multi-hour agent sessions, which makes the “open models aren’t there yet” narrative harder to sustain.</p>
<h2>Stripe’s autonomous PR agents</h2>
<p>Stripe engineers have deployed autonomous AI agents called Minions that are generating thousands of pull requests weekly. There’s not much detail on review processes or merge rates, which is exactly the kind of downstream question the AI coding conversation keeps skipping. The QCon London keynote this week made the same point: the industry is focused on generation and mostly ignoring deployment, observability, and what happens when an unsupervised agent leaks secrets to an NPM registry.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Expo Agent beta:</strong> Expo launched an AI agent powered by Claude Code that builds production-ready iOS, Android, and web apps from a text prompt, with a waitlist open at agent.expo.dev.</li>
<li><strong>Starlette 1.0:</strong> The Python ASGI framework underpinning FastAPI hit its first stable release after nearly eight years, downloaded almost 10 million times; Simon Willison used it to test Claude Skills for handling breaking changes Claude wasn’t trained on.</li>
<li><strong>Claude Dispatch:</strong> Anthropic shipped a feature letting users send tasks to Claude and monitor them in real time from a phone, with voice input support via tools like Aqua Voice.</li>
<li><strong>Google Stitch MCP:</strong> Google Stitch now has a native MCP server for Claude Code and Cursor, enabling design token extraction and consistent page generation across projects.</li>
<li><strong>OpenClaw security risks:</strong> A detailed post mapped OpenClaw’s attack surface to the OWASP Top 10 for AI agents, flagging insecure OAuth token storage, plain Markdown memory files that can be silently poisoned, and over 30,000 instances exposed to the internet without authentication.</li>
<li><strong>GPT-5.4 frontend gap:</strong> Multiple practitioners noted GPT-5.4 is strong for coding logic but behind competitors on frontend design output; OpenAI published a detailed frontend-skill prompt to compensate.</li>
<li><strong>Codex context window:</strong> The Codex backend is now reporting over 300K tokens input context.</li>
<li><strong>WebMCP proposal:</strong> The Edge and Google teams published an early browser API proposal letting web developers explicitly define which actions AI agents can perform on their pages via MCP, rather than relying on DOM scraping.</li>
<li><strong>Callstack React Native evals:</strong> Callstack released an open-source benchmark suite with 43 evals across animations, async state, and navigation to measure how AI coding models perform on real React Native tasks.</li>
<li><strong>CodexBar for Windows:</strong> A free, open-source taskbar utility now lets Windows users track Claude Code and Codex usage directly from the taskbar.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

A roundup of AI coding news: Cursor shipped a model built on Moonshot AI's Kimi k2.5 without disclosure, which the co-founder later admitted was a mistake. A security audit of 22,511 public AI coding agent skills found serious risks including remote code execution patterns in 1-in-6 skills and API traffic hijacking. Alibaba confirmed Qwen and Wan series will remain open source despite internal team turbulence, with Qwen 3.5 27B running efficiently on consumer hardware. Stripe's autonomous 'Minions' agents are generating thousands of pull requests weekly with little detail on oversight. Additional items cover Expo's Claude-powered app agent, Starlette 1.0 release, Anthropic's Claude Dispatch, Google Stitch MCP integration, and a WebMCP browser API proposal.

Cursor admits its Kimi base, Alibaba commits to keeping Qwen open source