Welcome back. AI is changing the way exploits are built. Apple has reportedly spent billions of dollars hardening the M5 chip with Memory Integrity Enforcement, a hardware defense built to make kernel exploits nearly impossible. Now, a small team just bypassed it in five days using an unreleased Anthropic model.

Also: Understand the hidden file structure that runs your AI agent, the 10 papers to read before an AI Engineer interview, and why HashiCorp's co-founder warns of "AI psychosis."

Anthropic's unreleased model reportedly defeats Apple's M5 defenses: The AI lab's restricted-access Mythos model just helped a small security team in Vietnam develop a working macOS kernel exploit for Apple's new M5 silicon in less than a week. Even with humans in the loop, it managed to bypass Memory Integrity Enforcement, Apple’s primary line of defense. For security and engineering teams, the window for potential attacks has narrowed significantly.

OpenAI ships a list of useful improvements for Codex: The ChatGPT maker just rolled out a wave of Codex updates: Git operations are now 10 to 50 times faster, and thread switching is way smoother after cutting out about 75% of unnecessary re-rendering. Startup feels much snappier, too. You can also customize keyboard shortcuts and handle commits, pushes, and PRs directly within the review flow.

Nvidia ships a world model that runs on a single GPU: The chip giant just open-sourced SANA-WM, a 2.6B model that generates a full minute of 720p video from just one image and a camera path. While similar world models usually require eight-GPU clusters, this one runs on a single card and actually beats them on camera accuracy. Even better, a distilled version can render an entire clip in just 34 seconds on an RTX 5090.

AI agents are making "digital twinning at scale" a reality. Test ideas against AI profiles of your customers before spending a dollar on development. Uncover the mechanics in Google for Startups' Future of AI Report. 

Download the report to learn more today!

The decade-long lock-in is officially over. For years, picking a programming language meant committing to it for a decade or more. That changed last Thursday when Anthropic's Bun runtime merged a million lines of Zig to Rust in just a few days. While a heated "Rust vs. Zig" debate followed, most people missed the more interesting takeaway.

Hashimoto saw the bigger picture. HashiCorp founder Mitchell Hashimoto pointed out that the real story isn't about which language won. Bun proved that a project can now migrate to almost any language in a week or two. Fungibility is the real breakthrough here; Rust was just the tool used to prove it.

This isn't an outlier. Django co-creator Simon Willison highlighted a similar shift: one mid-sized company used AI agents to rewrite its mobile apps in React Native, with a clear plan to revert if needed. Cloudflare reproduced the Next.js API in a week, and Ladybird's JS engine moved from C++ to Rust in just fourteen days.

The new stack math. A choice that used to bind you for a decade is now just a quarterly project. If your engineering team has been weighed down by legacy commitments, you can finally move past them. Use this guide from OpenAI to get started.

Tweet Hunter helps 5,600+ founders stay consistent on X.

In one session, you get:

All built on the official X API, so your account stays safe.

Start your free trial for 7 days

Coding on a remote server or cloud dev environment usually means juggling SSH, terminal sessions, and file syncing. OpenAI's Codex simplifies the process by letting you run the tool locally while editing code directly over SSH.

Now, every command Codex runs and every file it edits happens directly on the remote server instead of your laptop.

P.S. You can find 50+ AI coding hacks here.

How to use /grill-with-docs skill for smarter AI coding: This tutorial shows developers how to level up their AI coding using the Grill with Docs prompt. By using Domain-Driven Design to build a shared vocabulary in a context.MD file, the AI picks up on your project-specific terms. This cuts out the repetitive explanations and leads to much cleaner, more accurate code.

Asteroid: Automate complex browser workflows with Asteroid agents. Their agents handle complex, multi-step processes with built-in reliability, observability, and enterprise-grade scaling. From data extraction to full operational workflows, it helps you automate faster and ship with total confidence.

CodexBar (12.7K ⭐): This is a tiny macOS menu bar app that keeps your AI coding provider limits front and center, showing you exactly when each window resets. It has no Dock icon, a minimal UI, and dynamic icons for the menu bar.

Build iterative repair loops with Codex (OpenAI): Technical docs and code examples tend to break or go stale after a while. By using an AI loop to constantly check, fix, and test these files, you can make sure they actually work based on real execution results.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 280K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team