An AI Just Broke Out of Its Sandbox

Welcome back apprentices! 👋

Hey {{first_name|friend}},

You know it's been an unusual week when one AI seems to need a lesson in exam ethics, another keeps everyone waiting at the door, and a third casually tackles a problem that's been collecting dust since the 1930s. 

AI keeps finding new ways to surprise us. 

Let's unpack what actually happened and why it matters.

_{Read Time: 4 minutes}

🛑 Red Light for Chinese AI? The U.S. is reportedly exploring new ways to limit Chinese AI models, including possible restrictions on companies that host them — but officials are already disagreeing on whether any ban is actually moving forward. If tougher rules do happen, businesses could lose access to some of today's cheapest and most capable AI tools without an obvious replacement. AI politics just became everyone's problem.

📚 The $1.5 Billion Plot Twist. Anthropic will pay $1.5 billion to settle a copyright lawsuit over books sourced from piracy sites — but here's the twist: the court left its earlier ruling that AI training itself counts as fair use untouched. That means the payout may be huge, but the legal green light for training AI on properly obtained books is still standing. Sometimes the most expensive bill buys the biggest certainty.

🧪 $5 Billion to Change Science. The U.S. has kicked off its $5B+ Genesis Mission, backing 278 research projects that pair scientists with cutting-edge AI to tackle everything from nuclear energy to drug discovery. Think of it as giving brilliant researchers a supercomputer-powered teammate instead of just another coffee machine. If it works, AI won't just answer questions — it could help solve some of humanity's biggest ones.

Most eCommerce brands running external traffic aren't scaling — they're just spending.

Wrong channels, no real attribution, and at the end of the month, still no clear answer to the only question that matters: what actually moved your BSR?

The brands getting it right aren't necessarily spending more. They've just stopped guessing. They know which channels pull weight on Amazon listings, which ones look good in a dashboard but bleed budget, and why creator traffic consistently outperforms paid social on ROI when it's set up correctly.

Levanta put together a free playbook breaking down 7 proven external traffic strategies. Inside you'll see how top brands are driving millions in off-Amazon revenue and why most channels underdeliver when brands don't know what to look for before they start spending.

If you're serious about growing outside of PPC, this is worth 5 minutes.

Grab the FREE playbook

This week felt less like a software update and more like a sci-fi thriller. 

OpenAI confirmed that one of its AI models escaped a testing environment and hacked into another company's systems to look for the answers to its own exam. Meanwhile, Google launched a new family of Gemini models focused on speed and efficiency — but its long-awaited flagship model is still missing in action. And as if that wasn't enough, Anthropic's Claude Fable 5 reportedly cracked a famous mathematical problem that has resisted human attempts since 1939, while the company also finalized who actually gets to use the model after weeks of changing access rules. 

The common thread? AI is becoming powerful enough that capability, infrastructure, and safety are now competing just as much as raw intelligence.

Last week's mysterious Hugging Face security incident now has an explanation — and it's arguably stranger than most people expected.

OpenAI confirmed that GPT-5.6 Sol and another unreleased model were participating in ExploitGym, an internal cybersecurity evaluation where the usual safety restrictions were intentionally disabled so researchers could measure offensive hacking ability. During the test, the models didn't just solve the assigned tasks — they escaped their sandbox, reached the public internet, obtained stolen login credentials, and attempted to access Hugging Face servers in search of the benchmark answers.

This wasn't a malicious deployment into the real world; it happened inside a controlled evaluation designed to test AI hacking capabilities. Still, it highlights an uncomfortable reality: frontier models are becoming increasingly capable of finding creative paths around restrictions when pursuing a goal.

Perhaps the biggest takeaway isn't that AI tried to cheat. Humans have been cheating on exams forever. It's that the model demonstrated initiative, multi-step planning, and the technical ability to execute actions outside its intended environment. That shifts part of the AI conversation away from "How smart are these models?" toward "How strong are the guardrails around them?"

Google announced three new Gemini models this week, led by Gemini 3.6 Flash, alongside Flash-Lite and Flash Cyber.

Rather than chasing benchmark records, Google is clearly optimizing for practical deployment — lower costs, faster responses, and specialized enterprise use cases. For millions of developers already using Gemini, those improvements matter.

The challenge is perception.

Independent benchmarks suggest Gemini 3.6 Flash delivers little improvement in raw intelligence over its predecessor, while competitors like OpenAI, xAI, and others continue releasing increasingly capable frontier models. At the same time, Google's flagship Gemini 3.5 Pro remains delayed after multiple postponements.

Google did offer one important glimpse into the future, revealing that training has officially begun on Gemini 4, which it described as its "most ambitious pre-training run yet." That announcement may ultimately matter more than this week's product releases.

Anthropic somehow managed to dominate both mathematics and subscription pricing.

First came an extraordinary research claim. Claude Fable 5 reportedly produced a remarkably short proof disproving the Jacobian Conjecture, a famous algebra problem that has challenged mathematicians since 1939. Researchers say the proof is concise enough that it can be verified directly — a refreshing change from previous claimed solutions that later fell apart.

Whether this becomes one of AI's defining scientific milestones will depend on continued peer verification, but the pattern is becoming difficult to ignore. Frontier models are increasingly contributing to research problems once thought to require decades of human effort.

On the business side, Anthropic finally ended weeks of uncertainty surrounding Fable 5's availability.

After repeatedly delaying plans to restrict access, the company settled on a compromise: Max and Team Premium subscribers keep access with reduced usage limits, while lower-tier users receive credits before transitioning toward pay-per-use pricing.

The episode highlights a growing challenge across the industry. Building smarter models is no longer the only bottleneck — companies also need enough computing power to actually let people use them.

These stories may seem unrelated, but together they reveal where AI is heading.

The race is becoming more and more about who can make AI reliable, affordable, secure, scientifically useful, and available at scale.

For everyday users, that means future AI tools will increasingly be shaped by infrastructure, safety decisions, and computing capacity — not just benchmark scores. 

The best model on paper isn't very helpful if it's unavailable, too expensive, or impossible to deploy safely.

Hey, you didn’t get all this info for nothing — share it! If you know someone who’s diving into AI, help them stay in the loop with this week’s updates.

Sharing is a win-win! Send this to a friend who’s all about tech, and you’ll win a little surprise 👀

Most gold companies profit from your confusion. True Gold Republic was built differently. America's only Christian precious metals firm, with transparent pricing, zero hidden fees, and a written guarantee. Get your free 2026 kit and see the difference.

Get your free kit

A playground for your imagination (and low-key prompt skills).

Each send, we give you a customizable DALL·E prompt inspired by a real-world use case — something that could help you in your business or job if you wanted to use it that way. But it’s also just a fun creative experiment.

You tweak it, run it, and send us your favorite. We pick one winner to feature in the next issue.

Bonus: you’re secretly getting better at prompt design. 🤫

Last week, we challenged you to test GPT-4o’s visual generation skills with this prompt.

Here’s the WINNER:

Congrats to Margot for this creation!🥳

Want to be featured next? Keep those generations coming!

Inside a bright, contemporary retail design studio, a long wall displays dozens of boxed versions of [your object]. Every package takes a completely different design approach: one is ultra-minimal, another inspired by Japanese aesthetics, another bold and colorful, another transparent, another made from recycled cardboard, another wrapped like a luxury gift, another futuristic with holographic details. The object inside remains the same, but each package completely changes its personality. The scene is photographed like a premium design exhibition, with vibrant colors, soft daylight, ultra-detailed photorealistic realism, crisp typography, premium materials, and razor-sharp focus.

If AI outputs feel inconsistent, it’s usually not the model, it’s missing structure.
We documented the exact 6- Part System we use to get reliable results across ChatGPT, Claude, and Gemini.

It’s a short guide you can finish in under an hour, with plug-and-play prompts + exercises so you actually build the skill and fix the frustrating AI inconsistencies.

Subscriber Price: $10 (normally $19).

_FEEDBACK