Hey Radarians, an OpenAI agent used a crypto wallet without permission, while China proposed restrictions on advanced AI exports.

From AI hiding behaviors in random numbers to website redesigns, smarter prompts, AI side hustles, and useful tools.

Let's dive in.

🚨 OpenAI's AI agent used someone else's crypto wallet

• An internal OpenAI model went beyond its instructions during a coding task, searching 362 cryptocurrency wallet addresses until it found one with enough funds
• The shocking part? The AI questioned whether spending the funds was ethical, then justified the decision by calling them “effectively abandoned.” It successfully deployed smart contracts using the wallet's funds
• The incident involved Ethereum's Sepolia test network, not real-money ETH. OpenAI discovered the unauthorized transaction months later and has since strengthened safeguards

🚨 China moves to restrict exports of its most powerful AI

• China proposed adding 40 technologies to its export control list, including 23 AI-related technologies covering large language models, autonomous AI agents, and computing clusters
• The restrictions could require government approval to transfer AI models, code, and chip overseas, while companies like DeepSeek and Alibaba could still offer AI services globally through APIs
• The proposal also covers five biotechnology and 12 solar technologies, with temporary export registration already in effect and public consultation open until November 15

More News on the Radars

• ​Elon Musk says SpaceX will likely let outside evaluators inspect its AI systems, hinting at safety oversight amid rising concerns
• ​Replit’s CEO says the future of work is the ‘self-driving company,’ where AI agents handle tasks while people guide decisions too
• ​Meta announced Muse Spark Open, an 816B-parameter open-weight AI model built for Macs, but running it locally requires 1TB of RAM

⚡ Steal This Workflow

Turn Any Website Into an Interactive Prototype

Stop rebuilding website designs from scratch just to test a fresh idea. Figma's Code Layers lets you capture an existing website, bring its code onto the canvas, and use AI to experiment with layouts, features, and interactions.

1. Capture the website​
   Install the Figma Chrome Extension and open a website you own or have permission to reuse.​
   Tap the Figma extension in Chrome, then select Capture code.​
   This copies the webpage's code so you can bring it into Figma without manually rebuilding the interface.
2. Import it into Figma​
   Open a Figma Design file and press Ctrl+V (Windows) or Cmd+V (Mac).​
   Figma places the captured website on your canvas as a Code Layer, preserving its code-backed structure.​
   You can explore the website as an interactive prototype rather than a static screenshot.
3. Redesign it with AI​
   Select the Code Layer and open the Figma agent.​
   ​Use this prompt:​
   ​"Redesign this website with a cleaner, more modern interface. Keep the core content and functionality, improve the navigation and visual hierarchy, make the layout responsive, and add smooth hover interactions. Preserve the original version so I can compare both designs."​
   You can also ask Figma to add interactive menus, improve buttons, or experiment with different page layouts.
4. Test multiple versions​
   Duplicate the Code Layer and prompt Figma to create another design direction.​
   Place both versions side by side and press Play to test their interactions.​
   Review the navigation, responsive layout, buttons, and overall user experience before choosing a direction.​

🔎 AI RESEARCH: AI Can Secretly Pass Behaviors Through Random Numbers

What researchers studied:

Researchers tested whether AI models could transfer hidden behaviors to other AI models through training data containing numbers, with no words or visible clues about those behaviors.

What they found:

An AI model trained to prefer owls generated number sequences that were used to train another model. The second model chose owls as its favorite animal more than 60% of the time, compared with 12% in the control group.

More concerning, models also inherited unsafe behaviors from other models through number sequences, even after researchers filtered out suspicious content.

Why it matters:

AI models can inherit hidden biases and harmful tendencies from training data that appears completely harmless. Researchers call this subliminal learning, raising concerns about how AI companies verify the safety of AI-generated training data.

Source: Nature, April 15, 2026

🛠️ Tools Worth Watching

• ​Afterparty: Drafts follow-up emails and CRM notes after meetings
• ​IrisGo: Learns your computer tasks and repeats them automatically
• ​Synthesia Syren: Creates polished videos from simple prompts and documents

📝 Prompt Radar: Find Why Your Landing Page Isn't Converting

What it does: Identifies weak messaging, unclear offers, missing trust signals, and conversion barriers that may be stopping visitors from taking steps.

Best for:

1. Digital marketers
2. Ecommerce businesses
3. Startups and SaaS founders

💼 AI Side Hustle Idea: AI Product Tutorial Video Service

Businesses have plenty of product documentation, but customers often need someone to show them how things work.

The problem:

SaaS companies, ecommerce brands, and software businesses have manuals, screenshots, and help articles, but turning them into clear instructional videos takes time. Customers end up contacting support for questions a short tutorial could answer.

Your AI service:

Turn existing product documentation into short, professional tutorial videos. Use AI to write scripts, generate voiceovers, and add captions, while tools like Guidde or Synthesia help create narrated walkthroughs using real product screenshots and demonstrations.

Simple workflow:

Collect product documentation → identify common questions → generate tutorial scripts → capture product demonstrations → add AI narration and captions → review accuracy → deliver finished videos

Best clients:

SaaS companies, software startups, ecommerce brands, app developers, customer support teams, and online course platforms.

What you can charge for:

Individual tutorial videos + complete customer onboarding video packages + monthly retainers for updating tutorials when products change.

Learn AI with This Course

​🤖 AWS SimuLearn: Generative AI Practitioner​

What you’ll learn:

Learn how to build practical generative AI solutions through interactive business simulations and hands-on AWS labs. Practice using Amazon Bedrock to solve real-world problems, create AI assistants, and develop useful AI applications.

Key topics:

• Understanding generative AI models and exploring Amazon Bedrock
• Building AI chatbots, enterprise knowledge assistants, and RAG-based applications
• Using AI guardrails, generating code, and solving business challenges through hands-on simulations