Good Morning! OpenAI is quietly finalizing a powerful new model with cybersecurity capabilities and is now doing a restricted rollout to certain companies only, same playbook as Anthropic is used with Mythos. Plus, you’ll learn how to run Google's Gemma 4 locally with Claude Code for free.
Plus, in today’s AI newsletter:
OpenAI Prepares Restricted Cybersecurity Product
Google DeepMind CEO Wishes AI Stayed in the Lab
Perplexity Partners With Plaid for AI Wealth Management
How to Run Google's Gemma 4 Locally with Claude Code
4 new AI tools worth trying
AI MODELS
OpenAI is finalizing a new product with advanced cybersecurity capabilities that will be released exclusively to a small, handpicked group of partners.
The move closely mirrors Anthropic’s recent decision to limit access to its new "Mythos" model over fears of its advanced autonomous hacking capabilities.
Security experts warn we've reached a tipping point: models can now autonomously find software flaws and potentially write new exploits, raising alarms about attacks on critical infrastructure.
OpenAI previously launched an invite-only "Trusted Access for Cyber" pilot in February alongside GPT-5.3-Codex to accelerate legitimate defensive work.
Top security leaders warn that restricting releases only buys a little time, predicting that open-source models with similar capabilities will likely hit the wild within weeks or months.
AI capabilities have officially crossed the threshold from merely assisting human hackers to autonomously finding and exploiting software vulnerabilities. By gating these tools behind closed doors, frontier labs like OpenAI and Anthropic are openly acknowledging that their latest technology is now too dangerous for a public release.
AI NEWS
Demis Hassabis, the head of Google DeepMind and recent Nobel Prize winner for AlphaFold, just gave one of the most revealing interviews from any top AI leader, admitting that the current commercial AI race was premature.
Hassabis stated that if he had his way, he would have left AI in the lab longer to focus on fundamental science: "Done more things like AlphaFold. Maybe cured cancer or something like that."
He noted that the launch of ChatGPT locked the industry into a "ferocious commercial pressure race," redirecting focus from foundational breakthroughs (like energy and new materials) toward shipping chatbot products and quarterly earnings.
Beyond the current commercial sprint, Hassabis warned that his biggest fear isn't just bad actors, but AI itself going rogue in the next two to four years as we enter the "agentic era."
He emphasized that keeping these increasingly capable, autonomous systems aligned with strict guardrails is going to be an "incredibly hard technical challenge."
He called for international cooperation between labs, governments, and academia, stressing that the window to get alignment right is rapidly closing before the next generation of agents arrives.
When the man who built the system that might cure cancer tells you he wishes the industry had focused on that before launching chatbots, and that the next two to four years pose an enormous safety and alignment challenge, it's a massive wake-up call for the entire industry.
AI INTEGRATIONS
Following their recent brokerage integration, Perplexity now allows users to securely link bank accounts, credit cards, and loans via the Plaid network. This creates a unified financial picture where you can query your own real-time data using Perplexity’s advanced reasoning models.
Consolidates data from over 12,000 institutions, including Chase, Fidelity, and Vanguard, into a single, secure, read-only dashboard.
Enables "Perplexity Computer" to perform sophisticated tasks like building custom Excel models, debt payoff planners, and daily net worth trackers through simple natural language.
Pro and Max subscribers can build interactive dashboards and full financial apps that monitor spending and cash flow in the background.
Combines your permissioned personal data with institutional-grade sources like FactSet and Nasdaq for professional-level analysis.
Managing personal finance has historically meant jumping between three or four different apps to get a "holistic" view. By integrating Plaid's massive financial network with a frontier AI model, Perplexity is moving beyond simple web search into high-stakes utility. It’s no longer just telling you how to save money; it’s looking at your actual bills and building the spreadsheets to do it for you.
HOW TO AI
🗂️ How to Run Google's Gemma 4 Locally with Claude Code
In this tutorial, you will learn how to integrate Google's new state-of-the-art Gemma 4 open model with Claude Code using Ollama. This powerful combination gives you a highly accurate, privacy-first local coding assistant that can run entirely on your own hardware without API costs.
🧰 Who is This For
Developers who want to run AI locally
People concerned about privacy (no cloud APIs)
Hackers and tinkerers experimenting with LLMs
Startup founders building AI products
Engineers working with offline / edge setups
STEP 1: Install Ollama
First, you need to get the engine that will run your local model. Head over to the official Ollama website and download the app for your Mac or Windows machine. Once downloaded, open it up and create an account or sign in so you can access the model library.
STEP 2: Pull the Gemma 4 Model
Once done, open your terminal. Gemma 4 comes in different sizes depending on your hardware (like the E4B for laptops or 26B for powerful workstations). Once you decide which size your machine can handle, use your terminal to download it.
For example, you can run a command like ollama pull gemma4:e4b. Ollama will pull the manifest and download the open-source model directly to your computer.
STEP 3: Authenticate Claude Code
To use the Claude Code framework with your newly downloaded Gemma model, you still need to bypass Anthropic's authentication. If you haven't already, log out of your current Claude session in the terminal and log back in to connect your API key.
Note: You will need to deposit a minimum balance (around $5 to $10) in your Anthropic Console to activate the API service.
However, because you are routing the actual processing to your local Gemma model via Ollama, it will not actually consume those funds!
STEP 4: Launch and Code Offline
With your API authenticated and Ollama running in the background, you can launch Claude Code and instruct it to use your local Gemma 4 model. Start giving it prompts.
The model will "think" locally on your machine, write the code, and allow you to accept edits and view your work on a local host, all without ever pinging the internet.
Anthropic makes Claude Cowork, previously available as a “research preview”, generally available to all paid plans, and adds six features for enterprise use.
OpenAI launches a $100/month ChatGPT Pro subscription, which offers 5x more Codex usage than Plus; the $200/month Pro plan offers 20x higher limits than Plus.
Anthropic is weighing the possibility of designing its own chips, but it has yet to commit to a design or put together a dedicated team for the project.
Alibaba and China Telecom launch a data center in southern China that is powered by 10,000 of Alibaba's Zhenwu chips designed for AI training and inferencing.
🔥 Muse Spark: Meta’s new AI for personal, context-aware agents
💎 Gemma 4: Google’s powerful small AI model
🎥 Avatar V: HeyGen’s AI that creates high-quality avatar videos
🧠 PikaStream 1.0: turns AI agents into talking, face-to-face video bots
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