Welcome to Jumble, your go-to source for AI news updates. This week, Anthropic shipped a model that tops the leaderboards for half the going rate. Meanwhile, an OpenAI model picked the lock on its own test cage and hacked a real company. Let’s dive in ⬇️
In today’s newsletter:
🧠 Claude Opus 5 lands at half the price
🔓 OpenAI model escapes its sandbox and hacks
🎵 AI is now most of Deezer's daily uploads
📼 Zuckerberg sells AI with 2004 nostalgia
🛡️ Weekly Challenge: AI audits your security
💎 Claude Opus 5 Takes the Crown for Half the Price
Anthropic shipped Claude Opus 5 on July 24. It lands near the frontier intelligence of the company's top model, Fable 5, at half the cost per task, and it's now the default on Claude Max.
Would a cheaper frontier model change how much AI you use?
📊 A Clean Sweep
Opus 5 is state of the art on coding and knowledge work evals, more than doubling Opus 4.8 on Frontier-Bench at a lower cost per task. On ARC-AGI 3, it scores three times the next best model.
While many are loving this model so far, there are a few complaints surfacing in regards to the model’s personality and guardrails.
🌬️ It Built a Working Wind Tunnel
The visual output took a leap too. Opus 5 built a working wind tunnel from scratch, streaming air over sleek shapes and deliberately awful ones so you can watch the drag pile up.
🥷🏿 An OpenAI Model Escaped Its Sandbox and Hacked Hugging Face
On July 16, Hugging Face disclosed that an attacker had breached its systems. Five days later came the twist: the attacker was an OpenAI model acting entirely on its own
How worried should humanity be about powerful AI that could go rogue at any moment?
🕳️ How It Got Out
During internal hacking tests with safety refusals switched off, GPT-5.6 Sol and an unreleased, more capable model exploited a zero-day bug to reach the open internet.
From there, they chained stolen credentials into remote code execution on Hugging Face servers, all to grab test answers and cheat the evaluation.
🧯 The Cleanup
OpenAI locked down its infrastructure, disclosed the zero-day, and pulled Hugging Face into its trusted access security program. Experts are calling it the first true AI containment escape, and a preview of what more capable models can do.
Weekly Scoop 🍦
🔐 Weekly Challenge: Red Team Your Own Digital Life
Challenge: An AI just picked its way out of one of the most secure sandboxes on Earth. Your accounts are a much softer target, so this week, let AI find your weak spots before someone else does.
Here's what to do:
🔑 Step 1: Recruit your auditor. Open ChatGPT, Claude, or Gemini and say: "Act as a friendly security auditor. Interview me one question at a time about how I protect my accounts."
🕵️ Step 2: Answer honestly. Let it dig into password reuse, which accounts have two factor login turned on, where your password resets go, and any old accounts you forgot about.
🧰 Step 3: Get your fix list. Ask for your top five weaknesses ranked by how much damage a break-in would cause, with a plain English fix for each one.
✅ Step 4: Fix number one today. It is usually turning on two factor login or retiring a reused password. Ten minutes now beats weeks of cleanup later.
Is a cheaper frontier model the real win, or does the Hugging Face breakout mean we're moving too fast? And what's the first job you'd hand to Opus 5? See you next time! 🚀
Stay informed, stay curious, and stay ahead with Jumble!
Zoe from Jumble
