Welcome back! OpenAI rewrote ChatGPT's rules and built watchdogs to catch its AI breaking them. Mistral gave away the weights of a trillion-parameter model. Google and NASA are training robots to help astronauts survive on Mars. And researchers used a Chinese AI model to turn Firefox security fixes into working cyberattacks.
In today's Generative AI Newsletter:
OpenAI: Why does ChatGPT need bodyguards?
Mistral: What can you do with a trillion-parameter AI?
Google: Can AI robots keep astronauts alive on Mars?
Firefox: How did Chinese AI turn security fixes into attacks?
ChatGPT's new rules come with bodyguards
OpenAI just rewrote its Model Spec, the public rulebook for how ChatGPT should behave. It now focuses on five values: following human intent, minimizing harm, honesty, human agency, and care.
The company also revealed how it monitors its most powerful AI agents to catch suspicious behavior, including attempts to escape their testing environments.
OpenAI replayed an earlier security incident and found that its new monitors would have raised an alert more than 30 hours before the actual breach.
In a separate live training run, the monitors caught an AI model accessing the internet without permission within 15 minutes.
Serious alerts go to human security teams, and some critical incidents can automatically pause the affected AI activity.
The 30-hour finding came from replaying an old incident, not catching it live.
OpenAI admits its models don't fully follow the new rulebook yet, and the monitoring system still has weaknesses. The company is writing rules for its AI while building ways to catch it when it breaks them.
Learn Practical AI Skills for Just $19.90/Month
Get unlimited access to short, practical courses designed by AI experts, all for just $19.90 a month.
Each course focuses on real tools and skills you can start using immediately. Plus, learners have rated the courses 4.6 out of 5.
The subscription includes:
How to Stop Hitting Your Claude Limit: Understand why Claude interrupts your workflow and learn how to avoid hitting those limits.
Claude Cowork: Build Your AI Business Team: Create your first three AI agents and let them handle tasks for you.
Get Your Business Out of Your Head with Claude: Turn a five-minute conversation about your workflow into a clear, step-by-step SOP.
Ship a Micro-SaaS in a Weekend: Build a simple app, integrate Stripe payments, and start charging customers.
The curriculum is refreshed every quarter to keep pace with the latest AI models and tools.
Europe gave away a trillion-parameter AI
Mistral has released Le Chonk, its Mistral Large 4 model, making the weights of its most powerful AI model publicly available for developers and businesses to download, customize and deploy.
First previewed earlier this month, the model is designed for complex reasoning, coding, multimodal understanding, and agentic workflows.
1.05 trillion parameters, with only 52 billion activated per token through a Mixture-of-Experts architecture.
1 million-token context window, with support for text, images, and more than 160 languages.
82% on a cybersecurity benchmark, according to Mistral, testing its ability to identify and patch real software vulnerabilities.
3,800 NVIDIA Grace Blackwell GPUs used to train the model in Mistral's European data centers.
Mistral's weights are public, so anyone with enough computing power can take Le Chonk for a spin.
Google is training AI robots to keep astronauts alive
NASA and Google Cloud are testing AI-powered robots that could help astronauts handle everyday tasks and medical emergencies on future Moon and Mars missions.
Communication between Earth and Mars can take up to 44 minutes round-trip, making real-time assistance nearly impossible and increasing the need for autonomous robots.
The robot uses Physical Intelligence's 4-billion-parameter Pi0 model to connect visual understanding with physical movement.
In early laboratory tests, it achieved an 85% success rate at locating and moving a medical bag.
Google fine-tuned the model in just 2.5 hours, then demonstrated that it could run locally on compact, energy-efficient devices.
The robot is designed to handle cargo, organize equipment, and eventually assist with medical tasks such as preparing ultrasound machines.
Astronauts could eventually control it using simple voice instructions, without relying on constant communication with Earth.
Still in testing, but one day these robots could become astronauts' extra pair of hands in space.
Chinese AI turned Firefox security fixes into working attacks
Researchers at Epoch AI tested China's GLM-5.3 model and found it could turn upcoming Firefox security fixes into working cyberattacks before the updates even reached users.
The model exploited what's known as the patch gap, the window between a security fix becoming publicly visible and reaching users.
GLM-5.3 successfully developed attacks against all six Firefox releases tested between September 17 and October 26.
It created malicious webpages capable of stealing files from a user's computer simply by being visited, without requiring additional clicks.
Once the testing setup was ready, each successful attack cost around $1,000 in API usage.
The model worked autonomously, identifying vulnerabilities and combining them into attacks without further human instructions.
The attacks happened in controlled experiments, not against real Firefox users.
And with GLM-5.3's weights publicly available, these capabilities are no longer limited to a handful of AI labs.
Tool of the Day: Anam
Anam lets you create interactive AI avatars that hold real-time video conversations. Choose a realistic or animated character, customize its voice and personality, and connect it to AI models and external tools. Think virtual receptionists, customer support agents, or interactive tutors that feel more like face-to-face conversations.
Try this yourself:
Visit Anam and sign up.
Choose an avatar or create your own.
Customize its voice, personality, and instructions.
Start a conversation and watch your avatar respond in real time.
Who it's for: Developers, businesses, educators, and anyone building interactive AI experiences.
Everything else you shouldn't miss
NVIDIA may help finance Anthropic's AI expansion: it's reportedly discussing credit guarantees for Nscale's data center deal with Anthropic, but nothing has been finalized.
Walmart committed $6 billion to warehouse robots: it's expanding its Symbotic partnership to automate 25 US distribution centers over several years.
DeepSeek and Moonshot reportedly trained on GPT-6 Astra outputs: the Chinese AI companies allegedly used OpenAI's model outputs to improve their own, according to The Information.
Google brought Gemini Robotics 2 to the cloud: its new public preview helps robots plan tasks, correct mistakes, and coordinate with other machines.
Google's AI chips are challenging NVIDIA: demand for Google's custom TPUs is helping cloud growth, but the infrastructure investments are putting pressure on cash flow.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Check out courses made by AI experts on GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
