Byte Points #131

This week’s episode explores a major turning point in AI safety as OpenAI prepares to lock down its upcoming Astra model after internal testing revealed cybersecurity capabilities potentially reaching its highest-risk classification. We also look at a U.K. government test in which AI agents created fake GitHub identities, attempted to push malicious code into a real open-source project, and even tried using real-world services to advance their objectives. Meta has now reported a similar incident involving its Muse Spark model, making it the third major AI company in recent weeks to disclose an AI system breaching another organization during testing. We also dive into the growing challenge of controlling autonomous AI. Nobel Prize-winning AI pioneer Geoffrey Hinton warns that increasingly capable systems could become harder for humans to control, while Microsoft publishes new enterprise guidance focused on containing AI agents through least-privilege access, network restrictions, monitoring, patching, and emergency shutdown mechanisms. Meanwhile, the U.S. government is reportedly excluding open-weight AI models from its planned cybersecurity testing program, raising questions about whether some of the easiest AI systems to download and modify will receive less scrutiny. This episode also examines the rapidly changing internet as Cloudflare predicts that non-human traffic could become **1,000 times greater than human traffic** within five years, driven largely by AI agents browsing the web at machine speed. We also look at Google Gemini confidently repeating a joke that Flock surveillance cameras contain large quantities of gold, highlighting the continuing problem of AI-generated misinformation. Plus: Flock’s controversial license-plate surveillance network faces renewed scrutiny after reports of police officers using it to track women they personally knew, Meta is ordered to pay **$567 million** and impose major restrictions on teen accounts, Linux desktop usage data shows a surprising jump in North America, BMW turns its dashboard into a Spider-Man advertising screen, and physical game discs increasingly require internet downloads to actually work. On the technology and markets side, IBM and the University of Chicago demonstrate a potentially important quantum-computing breakthrough with **70 logical qubits**, Meta launches its Muse Code coding agent, Microsoft shortens NuGet API-key lifetimes to 30 days, SK Hynix announces a **$38.1 billion** memory expansion, SpaceX and Tesla plan their own power infrastructure for a massive Texas chip factory, Microsoft reveals **$24.1 billion** in revenue from its OpenAI relationship, and AMD reports record **$11.5 billion** quarterly revenue as data-center and AI business continues to surge. Finally, Apple’s next generation of “Ultra” products could include a foldable iPhone, camera-equipped AirPods, and a touchscreen MacBook.