📛 AI Malware rewrites its own code

Good morning, AI enthusiast. Google just uncovered a chilling new development in cybersecurity, malware that can literally rewrite its own code using AI to dodge detection.

In today’s AI newsletter:

• AI Malware learns to rewrite its own code

• OpenAI warns of Catastrophic AI Risk

• Microsoft forms superintelligence team

• How to build a production-ready AI agents

• AI tools & more

AI ETHICS

⚠️ AI Malware Learns to Rewrite Its Own Code

Google’s Threat Intelligence Group (GTIG) has identified a new experimental malware family called PROMPTFLUX, capable of leveraging large language models to dynamically alter its own code in real time, effectively teaching itself how to hide.

• PROMPTFLUX interacts with Google’s Gemini AI API to modify its code mid-execution

• Can generate malicious scripts and obfuscate itself on demand

• Currently in testing phase, incomplete features and limited API calls suggest early development

• Associated with financially motivated groups and a growing underground market for illicit AI tools

• State-backed actors from China, Iran, and North Korea are reportedly experimenting with similar AI-powered threats

This marks a turning point in cybersecurity, where AI is being used by both attackers and defenders. As hackers weaponize LLMs to evolve malware, companies like Google are deploying AI agents like Big Sleep to fight back. It’s officially AI vs AI in the cybersecurity battlefield.

AI SAFETY

⚠️ OpenAI Warns of Catastrophic AI Risk

OpenAI’s blog post (shared by Sam Altman) says systems can already outperform top humans on hard intellectual tasks and that we’re likely to see small scientific discoveries from AI by 2026, with more significant discoveries possible by 2028. The company urges stronger safety, oversight, and an “AI resilience” ecosystem.

• The cost of intelligence has dropped 40× per year, rapidly accelerating development

• Warns that deploying self-improving AI without proven alignment is extremely dangerous

• Calls for shared safety standards, public oversight, and global reporting

• Proposes an “AI resilience ecosystem,” similar to cybersecurity, to track real-world impacts

OpenAI’s latest message marks a shift from product hype to existential concern. The company’s leadership is signaling that the world is unprepared for the speed and power of frontier AI, urging global cooperation before machines outpace human control.

AI NEWS

🤖 Microsoft Forms Superintelligence Team

Microsoft has launched the MAI Superintelligence Team, led by Mustafa Suleyman, the former DeepMind co-founder and Inflection AI CEO. The group will focus on advanced AI research aimed at solving real-world problems while keeping systems controllable and grounded.

• Focus areas include AI companions, medical diagnostics, and renewable energy optimization

• Suleyman emphasizes building practical AI, not uncontrolled superintelligence

• Microsoft aims for “expert-level” diagnostic capabilities and predictive planning in healthcare

• Move follows Meta’s own “Superintelligence Labs” push earlier this year

• Comes amid growing investor scrutiny over AI spending and unclear profit paths

With Suleyman steering this new team, Microsoft is doubling down on human-centered superintelligence, a vision that contrasts the “AGI arms race” narrative. Instead of chasing omnipotent AI, Microsoft wants systems that can work alongside people, tackling global challenges from education to clean energy.

HOW TO AI

💻 How to Build a Production-Ready AI Agent in Minutes

In this tutorial, you’ll learn how to create a fully functional AI agent using Vellum, a no-code platform that turns plain text instructions into deployable workflows.

🧰 Who is This For

• Developers who want to automate workflows fast

• Product teams experimenting with AI-powered features

• Founders building MVPs without writing code

• Anyone who wants to turn text ideas into working automations

STEP 1: Access the Agent Builder

Head over to Vellum.ai and Sign Up for free. You’ll land on a clean, chat-style dashboard that looks familiar if you’ve used ChatGPT before.

Right in the center, there’s a text box that says: “Describe what you want your agent to do.”

This is your command center, everything starts here.

STEP 2: Describe Your Agent in Detail

Now, simply describe what you want your agent to do, just like you’d explain it to a teammate.

For example, you can type:

“Create an agent that reviews legal contracts against a checklist and generates risk assessments with lawyer-friendly summaries…”

Once done, click enter and that’s it.

Vellum will instantly break down your request into clear steps and build the workflow for you automatically.

STEP 3: Customize and Configure

Once the workflow appears, you can click on any part of it to edit or fine-tune.

Change model types, swap integrations, or adjust prompts, all from one screen.

When ready, click Test Run to see your agent in action.

It’ll perform each step in real time, pulling data, writing text, or creating summaries right in front of you.

If something feels off, just tweak your instructions and test again until it’s perfect.

STEP 4: Deploy

When you’re happy with the results, click Publish.

From now on, your agent runs automatically whenever the conditions you set are met, no extra setup needed.

France government launched a new LLM leaderboard focused on European languages and energy efficiency, similar to LMarena.

IBM CEO Arvind Krishna says the company will hire more Gen Z graduates this year even as it cuts thousands of jobs tied to AI restructuring.

Draft documents show the European Commission plans to relax some privacy laws, including the GDPR, to boost AI growth and cut red tape for businesses in Europe.

Scientists just turned human stomach cells into insulin-making cells inside mice. This is an incredible success: a credible path to replacing injections with living cells that sense sugar and release insulin on demand.

🧠 Kimi K2 Thinking: Moonshot AI’s new open-source advanced reasoning model

🖼️ MAI-Image-1: Microsoft’s first in-house text-to-image model

✍️ Arcitext: Find and understand your brand’s unique writing style using AI

🤖 Excelmatic: Use AI to analyze, visualize, and automate your data

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas