Welcome back. Earlier this month, we witnessed the first documented case of agentic ransomware: a cyberattack driven completely by an LLM. And now Hugging Face reported a breach carried out by OpenAI’s latest models to score on benchmarks. Today we cover what exactly happened.

Also: Give your agents a company brain, cut your token usage in GitHub Copilot, and the 10 AI repos blowing up on GitHub.

OpenAI's models hacked Hugging Face to cheat a benchmark: A few days ago, we reported that an AI agent breached Hugging Face's servers. The ChatGPT maker just revealed that GPT-5.6 Sol and an unreleased model were the attackers. These models escaped their sandbox during a cyber evaluation and broke into Hugging Face's servers looking for test answers.

Google's new models make AI agents cheaper to run: The search giant just dropped Gemini 3.6 Flash and 3.5 Flash-Lite to reduce high token costs for production agents. The company claims its Gemini 3.6 Flash model uses 17% fewer output tokens than its predecessor and performs better at coding and computer tasks. They also released Flash Cyber, a specialized security model that finds and patches vulnerabilities, though the company currently limits its rollout to government agencies and trusted partners.

Anthropic teaches Claude to watch and learn: The AI lab just shipped two major screen-focused updates. In Claude Cowork, you can now record yourself performing a task and narrate your steps. Claude then converts that workflow into a reusable skill. Simply click “Record a skill” in the desktop app to start. Meanwhile, Claude Code now supports iOS simulators. This allows Claude to run your app, navigate through it, and iterate until the code works perfectly.

AI is in your engineering workflow. While the token spend shows it, the throughput doesn't. The human is very much still in the loop, and that's a context problem.

Join live and for free on July 23 to learn:

Register Now.

The lights-out gamble. "Dark factory" is the phrase taking over engineering circles this week. It refers to AI agents writing and shipping code that no human ever reads. Former Google director Addy Osmani just mapped the idea in his viral essay. His warning is that a team that stops reading its code slowly stops understanding it.

Free everywhere but one place. The rot starts with a basic imbalance. Generating code, running tests, and security scans now cost almost nothing. Human review is the only stage that stays expensive, because attention doesn't scale. So Osmani sets a strict limit. An agent gets only as much freedom as you can cheaply verify, and not one inch more.

Proof from the floor. The essay draws on the insights from HumanLayer co-founder Dex Horthy. In a talk at the AI Engineer World's Fair, he explored why software factories fail. Horthy has seen it happen himself. He once ran a fully automated factory for four months without anyone reading a single line of code. Even though the tests stayed green, the team slowly lost their understanding of the system. Fixing it took painstaking manual debugging. Osmani calls this gap comprehension debt.

Read the blueprint first. Osmani's answer is to keep humans wherever mistakes are expensive. Review the plan before the agent builds. That way, you're judging the decision itself. If you skip it, you'll be digging that same decision out of thousands of lines of code later. His outer loop essay goes deeper, and Horthy's 12-factor agents guide shows how to build loops that can safely run in the dark.

Recently, a vibe-coded app leaked 1.5M API keys because nobody reviewed the code.

Your team ships AI-written code every day. Superblocks keeps it governed:

Book a demo.

4 repos to cut token usage (by a Microsoft engineer): You'll learn how to slash token costs while using AI coding agents. The tutorial walks through setting up four tools: RTK for trimming command outputs, Headroom for compressing chat history, Ponytail for writing more concise code, and Graphy for quickly mapping out projects for the AI. 

FogLamp Scan: This tool creates an interactive map of your repository with just one prompt. You don't need to install anything or create an account to get started.

Show me your work (by a Cursor engineer): A plugin that helps you navigate large repositories. It establishes a standard for maintaining a clear audit trail of decisions during long or automated agent sessions. 

How to run large code migrations with Claude Code (by Anthropic): Code migrations are getting faster. Bun ported 1M lines of Zig to Rust in under two weeks with all tests passing, while another Python codebase was moved to 165k lines of TypeScript over a single weekend. Anthropic has now shared a six-step workflow to help manage these large-scale migrations.

Reviewers often repeat the same PR comments, like warnings about field renames or logging customer data. Codex Code Review now fixes this by reading your custom rules from AGENTS.md and citing them directly in its feedback.

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

Check out Coinbase's new playbook on their AI-native engineering interview process. 

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