I've been using GitHub Copilot since day one. So finding out that starting April 24, every line of code I write with it becomes training data β unless I manually opt out β honestly caught me off guard. Here's what you need to know this week.
π₯ The Big One
GitHub is training Copilot on your code starting April 24 β and most developers don't even know.
This is one of those changes that flies under the radar until it's too late. Starting April 24, GitHub will use all Free, Pro, and Pro+ Copilot interaction data to train their AI models. That means your code snippets, file names, navigation patterns, comments, and documentation are all going into the training pipeline.
The kicker? It's opt-in by default. If you don't go into your settings and manually disable it before the deadline, your coding patterns become Microsoft's training data. No notification, no confirmation prompt. Just a policy update buried in a blog post.
Enterprise and Business plans are exempt β their contracts explicitly prohibit training on customer data. But the millions of individual developers on Free, Pro, and Pro+ plans? You're in unless you act.
GitHub CPO Mario Rodriguez cited "meaningful improvements, including increased acceptance rates in multiple languages" from internal Microsoft employee tests as justification. Which is great for the product, but it means the improvement you're seeing in Copilot suggestions is literally built on code from developers who didn't realize they were contributing.
This isn't about whether AI training on code is good or bad. It's about informed consent. A 30-day window buried in a blog post isn't consent β it's a countdown. Go to Settings β Copilot β disable interaction data sharing. Do it today.
"All Free/Pro/Pro+ interaction data β code snippets, file names, comments β used for training. 30-day opt-out window open NOW. Enterprise exempt." β GitHub Blog
π οΈ What I built this week
With all the Claude Code updates dropping, I spent most of this week stress-testing the new features and exploring what's possible with MCP integrations.
-
Auto Mode workflow β Tested Claude Code's new Guardian Classifier across three real projects. It auto-approved safe file edits and terminal commands while blocking anything destructive. Replaced my old --dangerously-skip-permissions flag, and honestly, I should have had this guardrail months ago.
-
Notion MCP integration β Connected Notion's official MCP server to Claude Code. My AI tools can now read and write to my entire Notion workspace in real-time. The workflow for documentation just collapsed from hours to minutes.
-
Claude Code Auto-Fix on a live repo β Set up Auto-Fix on one of our GitHub repos. Claude watched for CI failures and review comments, then pushed fixes autonomously. Three PRs fixed without me touching the keyboard
βThe best week to be a Claude Code user. The worst week to be Anthropic's security team." β Sonny
β‘ What shipped this week
1. Anthropic's wildest week ever β leaks, launches, and chaos
Anthropic had the most chaotic week in AI company history. First, nearly 3,000 unpublished assets leaked from a misconfigured CMS, revealing a secret model codenamed "Capybara" (project name: Claude Mythos) β described internally as a "step change" in capabilities. Days later, ~500,000 lines of Claude Code source code across 1,900 files were exposed via an npm package source map. Two security disasters in one week. And somehow, in the middle of all this, they shipped Auto Mode, Auto-Fix, Computer Use from the CLI, and mobile work tools. Four major product launches while putting out two security fires. The tools keep getting better. The operational security needs to catch up.
"Model codenamed 'Capybara,' described as a 'step change' β far ahead in cyber capabilities." β Fortune
2. Notion shipped an official MCP server
Notion's OFFICIAL MCP server lets AI tools read and write your workspace in real-time. Works with Claude, ChatGPT, Cursor, and any MCP-compatible client. Open-source on GitHub. With MCP crossing 97 million installs this week (fastest-adopted AI infrastructure standard ever), this is Notion positioning themselves as the default knowledge layer for AI agents.
"Notion's OFFICIAL MCP server β AI tools can read and write your workspace in real-time." β Notion
3. Google shipped Gemini 3.1 Pro with Deep Think
Google is quietly stacking wins. Gemini 3.1 Pro launched with three thinking levels (Low, Medium, High), and the Gemini app has hit 750 million users. They also dropped Gemini 3.1 Flash Live, their highest-quality audio and voice model, powering both Gemini Live and Google Search Live across 200+ countries. Lower latency, better tonal understanding, real-time voice and camera input. The pace of shipping from Google right now is relentless.
"Gemini 3.1 Flash Live: Google's highest-quality audio/voice model. 200+ countries." β Google AI
4. Amazon and OpenAI partnered on stateful AI agent runtime
OpenAI models are coming to Amazon Bedrock AgentCore with persistent memory and multi-turn state. The key distinction: Azure gives you stateless, AWS gives you stateful. For anyone building production AI agents that need to remember context across sessions, this partnership makes AWS the obvious choice. The agent infrastructure wars are heating up.
"OpenAI models coming to Amazon Bedrock AgentCore with persistent memory and multi-turn state." β AWS
π§° Worth your time
-
Jupid β File your taxes using Claude Code. Yes, really. AI-powered accounting with persistent vendor memory and ~96% accuracy on IRS Schedule C categorization. Hit #1 on Product Hunt with 506 votes. If you're a freelancer or solo dev, this could save you hours.
-
Agentation β Click any UI element and get structured markdown with CSS selectors and React component hierarchy. Paste it straight into Claude Code or Cursor. If you're doing any frontend work with AI tools, this closes the visual feedback gap.
-
Crossnode β Two 19-year-old founders built a platform to vibe code AI agents and put them behind a paywall. Turn n8n workflows into SaaS products. 310 votes and 62 comments on PH. The fact that teenagers are shipping production agent platforms tells you everything about where this space is going.
-
Pendium β AEO/GEO platform that helps AI agents recommend your product more. This is the new SEO: optimizing for how ChatGPT, Claude, and Gemini discover and recommend your brand. Early movers here will have a massive advantage.
My weekly message to YOU!
This week was a wake-up call about who controls your data in the AI era. GitHub quietly changed their training policy. Anthropic accidentally leaked their most powerful model. The tools we rely on every day are moving fast β sometimes faster than their own security can keep up.
The takeaway isn't to stop trusting AI tools. It's to build your own verification layer. Review what your tools have access to. Check your GitHub Copilot settings before April 24. Understand what data flows where. The developers who thrive in this era won't be the ones who blindly trust every tool. They'll be the ones who use everything available while keeping their eyes open.
Your challenge this week: audit your AI tool permissions. Go through every tool that touches your code or data and verify what it has access to. It takes 15 minutes and it could save you from a surprise down the road.
What's one AI tool permission that surprised you when you checked? Reply and let me know.
I read every single one.
Talk soon PAPAFAM,
Sonny ππΌ
ππ½ Don't forget to follow me across socials!
Responses