Welcome to Jumble, your go-to source for AI news updates. This week, Google showed off Genie 4, an AI that builds worlds you can walk through, then handed it to just a few testers. Meanwhile, a startup says its AI agents broke into Anthropic in under 12 hours. Let’s dive in ⬇️
In today’s newsletter:
🌍 Google’s newest worlds come with sound
🔓 AI agents slip past Anthropic’s guards
🧠 An AI proof upends decades of research
🎟️ China’s AI movie plays to empty rooms
🕵️ Weekly Challenge: Let AI play hacker on you
🌎 Genie 4 Turns a Prompt Into a World
Google DeepMind just unveiled Genie 4, a world model that builds explorable worlds from a text prompt, now with sound, music and characters that talk back.
What do you think about AI world models?
🎮 Walk It, Then Rewrite It
You move through the world with your keyboard and change it on the fly, like making it rain. Scenes now hold together for about two minutes, though Google admits details still drift.
🎭 Google’s Reverse Show and Tell
Only trusted testers get in for now. Google’s new top model, Gemini 4 Carbon, launched the same way. While rivals usually let paying users try new flagships on day one, Google seems to be taking a totally different approach.
If their latest stock slide is any indication of whether it’s working, they may want to start shipping more than they’re talking.
🏰 A Startup Says Its AI Agents Broke Into Anthropic
Cathedral, a startup founded by ex-DOGE staffers, says its AI agents got into Anthropic’s internal systems in under 12 hours, reaching Slack messages, meeting transcripts and staff salaries.
🚪 In Through the Side Door
The agents slipped in through outside sales and HR software, not Claude itself. Anthropic says chats, source code and model weights stayed safe, and it cut off access the same day.
⚖️ Bug Hunt or Break-In?
Cathedral came in through Anthropic’s paid bug-hunting program, but Anthropic says it broke the rules. A Trump official warned that if “a bunch of kids” can do this, China isn’t far behind.
Weekly Scoop 🍦
🦹 Weekly Challenge: Ask Your AI to Hack You
Challenge: If AI agents can get into Anthropic, they can get into your accounts. This week, have your AI play the attacker, then fix what it finds.
Here's what to do:
🃏 Step 1: Hire a fake scammer. Turn on web search in ChatGPT, Gemini or Claude and ask: “Act like a scammer. Search the web for [your name and city] and tell me what you’d find and how you’d use it to trick me.”
💧 Step 2: Check for leaks. Run your email through Have I Been Pwned and your passwords through Google Password Checkup or Apple’s Passwords app. Never paste passwords into a chatbot.
🛠️ Step 3: Get a fix list. Share what you found (minus passwords) and ask for your top five fixes, ranked. Expect things like passkeys, a SIM lock from your carrier and a free credit freeze.
🧽 Step 4: Scrub yourself. Use Google’s Results about you to pull your phone number and address from search, then knock out one fix a day.
Should Google keep showing first and sharing later, or just ship it? And should AI labs pay hackers who break the rules to prove a point? See you next time!
Stay informed, stay curious, and stay ahead with Jumble!
Zoe from Jumble
