Welcome back. It’s not often you’ll see the US government and Chinese tech firms operating with the same logic. Days after the US government issued an export ban for Anthropic’s Fable 5 (which has now mostly been resolved), a major Chinese tech giant is now barring its staff from using the model for similar reasons.

Also: How engineers at Anthropic and OpenAI work, Anthropic’s Fable 5 prompting guide, and OpenAI researcher’s career advice for engineers in 2026.

Alibaba bans staff from using Claude Code: The Chinese tech giant is reportedly banning staff from using Anthropic's coding tool starting July 10. They have labeled the tool as high-risk software. This decision comes after a viral post on Reddit claimed the tool had hidden code to covertly target Chinese users. Anthropic's Thariq Shihipar explained the feature was part of a March anti-abuse experiment. It was meant to stop resellers and model distillation.

Viral pixel hack splits the AI dev community: The internet just spent a weekend arguing over pxpipe, a proxy that cuts Claude Code bills by rendering bulky context as cheaper image tokens. Its repo claims savings up to 70 percent. Skeptics fired back fast, arguing the lossy compression can misread exact strings like hashes without warning. Researchers counter that this idea isn’t new and point out that AI labs quietly validated pixel-based reading years ago. You can give it a try here.

Researchers catch AI running a ransomware attack alone: Security researchers at Sysdig say they've spotted the first ransomware campaign driven start to finish by an AI agent, no human at the wheel. The agent, called JadePuffer, breached a server through a Langflow vulnerability, moved laterally, and locked up 1,342 database configs before demanding Bitcoin. Sysdig warns that these "agentic threat actors" sharply lower the skill needed to pull off damaging attacks.

IBM Bob works directly in the codebase to generate and refine code in context.
Describe what you need in natural language and get implementation where work is
already happening, without switching tools.

As development progresses, Bob can surface complexity and suggest refactors early, before issues build up. Across IDE, terminal and pipeline, it can support the full workflow, helps reduce rework to keep delivery on track.

Get to know Bob

A week inside the top AI labs. Veteran engineer Gergely Orosz recently spent a week inside OpenAI, Anthropic, and Cursor. He wanted to know one thing: how is their daily work changing? The most striking answer came from Jarred Sumner, an MTS at Anthropic. This is the same Jarred who went viral on X for porting Bun from Zig to Rust. He explains how a massive migration that usually takes a year can be finished in just two months. 

Dozens of agents, one branch. Sumner pulled it off by spinning up dozens of parallel coding agents, each owning a slice of the migration. His hardest problem was keeping them from stepping on one another's toes: he skipped Git worktrees after finding them too slow, put every agent on the same branch, and divided the work so no two agents ever touched the same file.

History keeps rhyming. Compilers automated machine instructions and memory-safe languages handled memory. Both shifts pushed engineers toward higher-leverage work. Cursor cofounder Sualeh Asif told Orosz the next step up is already here: with code generation automated, the leverage sits in designing environments where agents execute well.

A skill worth building early. Most teams still run one agent at a time and babysit it. Anthropic's Claude Code best-practices guide shows Sumner's playbook in miniature: generate the list of files to migrate, fan out one scripted agent per file with scoped permissions, test on two or three files, then run at scale. Orosz's field notes cover the rest.

Agentic coding config from scratch (by Ex-Meta Principal): In this tutorial, you will learn how to set up a completely reproducible "agentic engineering" development environment from scratch. By using Nix and tools like Nix Darwin and Home Manager, they'll be able to write configuration files that can set up a system consistently every time, starting from a freshly installed Mac.

Render: It provides the infrastructure developers need to deploy apps and agents that scale without limits. Just push your code to instantly set up web apps, databases, agent loops, and workers with automatic networking.

Davidondrej-skills (7K interactions): This repository contains reusable skills for AI coding agents, research agents, and workflow agents. Each skill packages a focused workflow into instructions that an agent can load when the task calls for it.

How to handle rate limits: You’ll learn how to prompt Codex effectively for agentic coding tasks, including structuring instructions, guiding code changes, and getting more reliable, high-quality results from the model.

Most "can you change this" requests in Slack get lost because acting on them requires switching back to your editor. Cursor just released a fix in v3.8. Now, a simple emoji reaction triggers an agent.

The Cursor team even uses this workflow internally. Just connect Slack via your Cursor dashboard and run /automate in any agent session. Make sure to replace “your-repo” with your actual repository: 

Reacting with the eyes “👀” emoji now hands the message off to a cloud agent. You’ll see the run pop up in your Agents window.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Anthropic's top exec Mike Krieger and the Every team just shared 13 prompts you can use right away. Make sure you bookmark these before moving on to the rest of these emails.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ 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