Weekly Digest: Apr 26

The week’s loudest signal wasn’t a product launch, it was a witness stand. Musk spent three days testifying against OpenAI while the Pentagon quietly handed classified-network contracts to almost every major AI lab except Anthropic. Meanwhile a wildlife YouTuber shipped the #1 paid iOS app, and WordPress got another fifty-thousand-site reminder that file upload plugins remain a liability. The builder repos keep climbing, which is where I’d actually put my attention.

AI & Automation

The Pentagon signed classified-network deals with OpenAI, Google, Microsoft, AWS, Nvidia, and xAI, and conspicuously skipped Anthropic, which tells you usage-policy fights now have real procurement consequences (The Verge).
The DOD’s Anthropic snub is the through-line worth tracking; vendor diversification is the official story, but terms-of-service disputes are the real driver (TechCrunch).
Musk’s case against OpenAI is going badly in open court, with his own emails and tweets being read back to him, and there’s still a long witness list to come (The Verge).
Meta acquired humanoid startup Assured Robot Intelligence to feed its robotics models; a quiet acquisition that says more about Meta’s roadmap than any keynote did (TechCrunch).
DualShot Recorder hit #1 paid on the App Store in 12 hours, built by a wildlife creator solving his own problem. It is the clearest “audience-first product” case study of the quarter (The Verge).

Trending Code & MCPs

n8n crossed 186k stars and is still the workflow layer most serious automation builders default to before reaching for code (GitHub).
Ollama is at 170k stars with day-one support for Kimi-K2.5, GLM-5, and gpt-oss; local-first inference is no longer a hobbyist position (GitHub).
everything-claude-code is the harness repo to watch if you’re tuning Claude Code, Codex, or Cursor; skills, memory, and security treated as a system rather than vibes (GitHub).
prompts.chat (the rebranded Awesome ChatGPT Prompts) hit 161k stars and is now self-hostable, which matters if you don’t want your prompt library on someone else’s server (GitHub).
The Google Sheets MCP server crossed 45k installs on Smithery. It is the boring integration that quietly powers half the agent demos people show off (Smithery).

Side Hustles

Failory’s roundup of the 27 most successful Shark Tank products across 17 seasons is a useful reality check; survivorship is rarer than the show makes it look (Failory).
The companion piece on Google’s 34 biggest failed products is the same lesson from the other side: distribution doesn’t save a bad product (Failory).
Vinted’s business model breakdown is worth reading if you’re thinking about marketplace economics; buyer-pays-fees is doing real work for them (Failory).
Plurai launched on Product Hunt with vibe-trained evals and guardrails — eval tooling that meets builders where they actually are instead of demanding a PhD (Product Hunt).
Velo hit 666 upvotes for “share anything as video messages” — async video keeps eating meetings and the tooling layer is still wide open (Product Hunt).

WordPress Pulse

Ninja Forms File Upload had an unauthenticated arbitrary file upload vulnerability across 50,000 sites — patch now if you’re running it, and stop letting unauthenticated uploads anywhere near your stack (Wordfence).
Wordfence published a candid post on how AI is reshaping their bug bounty program — researchers using LLMs to find vulns faster, which cuts both ways (Wordfence).
WordCamp Asia 2026 in Mumbai pulled 2,627 attendees with a schedule heavy on AI, enterprise WordPress, and developer workflows (WordPress.org).
The session lineup recap is worth a skim if you want to see where the project’s official narrative on AI is heading (WordPress.org).
The weekly Wordfence vulnerability report is the boring habit that prevents the loud incident — bookmark it if you manage more than one site (Wordfence).

Next week I’m watching whether Anthropic responds to the Pentagon snub with a policy clarification or doubles down, and whether the Musk trial produces any actual evidence that reshapes the OpenAI corporate-structure story. On the build side, I’m running a small test pitting n8n against a Claude Code harness on the same content workflow to see which one is actually faster end to end.