👋 Hello hello,
Somewhere on Google right now, a stranger might've been reading a Claude chat you never meant to leave your inbox. Meanwhile Microsoft built a model whose entire job is finding the crack in your code before someone else does, and Meta decided your smart glasses needed opinions on Threads too.
🚨🚨 Heads up: Every week the PAI team hops on screen and builds AI workflows in real time. If you've been following along but haven't tried building yet, this is your chance!
Spots are limited on purpose, so grab one now. Join the PAI team live this Friday, 10AM PT / 1PM ET →
Let's get into it.
🔥🔥🔥 Hot off the presses
Over the weekend, Reddit users noticed that typing a simple search command into Google surfaced hundreds of publicly shared Claude conversations, complete with crypto wallet keys, API credentials, resumes, and even health records.
Here's the twist: Anthropic says the links were never guessable. This only happened to conversations people had actively shared and then posted somewhere public, like a forum or social post, which search engines could then crawl. Anthropic has since fixed the indexing gap, and the exposed links have largely vanished from search results. so, basically If you try to search with "site:claude.ai/share" now, you receive a failed search result.
Anyway, if you've ever used Claude's share button, it's worth a quick look through your shared chats and revoking anything you don't remember meaning to make public.
Microsoft just introduced MAI-Cyber-1-Flash, a new security-focused model plugged into MDASH, its multi-agent system for finding and fixing software vulnerabilities. On CyberGym, the benchmark for spotting real exploits in complex code, the combo hit 96%, beating out rival models while costing about 50% less than Microsoft's previous best security setup.
The bigger idea: let a cheaper, faster model handle 90% of routine vulnerability hunting, and save the expensive heavyweight models for the genuinely nasty 10%. Microsoft also launched Perception, an agentic system built to continuously monitor and patch threats using this stack. Learn more here.
Meta's latest update for Ray-Ban Display glasses does two things:
it brings Threads directly into the glasses so you can stay connected without pulling out your phone, and
it upgrades Meta AI with the new Muse Spark model, meaning sharper answers, better recognition of what you're actually looking at, and more proactive suggestions throughout your day.
More on this update here.
🔥🔥 Two AI Tools Worth Trying
1. 📲 Pushary
If you run AI coding agents like Claude Code or Cursor, you know the pain of walking away only to find the agent frozen, waiting for a permission you weren't there to give. Pushary sends that single decision straight to your phone so you can approve or deny with one tap, no matter where you are. Great for anyone running agents unattended overnight or during meetings.
2. 🔍 ychasit
Ever wondered if someone's already built the tool you're dreaming up? ychasit lets you describe your problem in plain English and searches over 4,000 active YC companies to find the startups already solving it, complete with reasoning, pricing, and integrations. It's free, no login required.
🔥 Design meets coding agent
Paper is a design tool built specifically to work alongside AI coding agents instead of sitting off to the side. Here's how to put it to work:
Open Paper and connect it to your coding agent, like Claude Code or Cursor, through MCP.
Point your canvas at your actual codebase or content source so you're designing with real data instead of placeholder text.
Sketch layouts and edit copy directly on the canvas.
Ask your agent to sync those design tokens and styles straight back into your codebase, since Paper exports as real HTML and CSS.
Hand off the repetitive stuff, like responsive variants and consistency checks, to the agent and spend your time on the calls only you can make.
🎤 Join Us for a Live AI Build On Friday!
Every Friday, the PAI team gets together to host a live AI build session where we build workflows that you bring to us.
Because we build in real-time, you get to see how we think through the problem, where we get stuck, and what the final outcome is. It's the difference between reading a recipe and watching someone cook it while talking through their decisions.
🗓️ When: Friday, July 31st @ 10AM PT (1PM ET)
📹️ Where: Virtual
Would you trust an AI agent to work unsupervised while you sleep?
Until next time,
Team @PracticalyAI
