Apple Left Claude.md Files in Their Support App. Now We Know.

What happened

Aaron Perris (@aaronp613), a well-known Apple security researcher, discovered that Apple shipped `CLAUDE.md` files inside the production bundle of their Apple Support iOS app — the consumer-facing app millions use to manage Genius Bar appointments and get tech support. The finding, posted to Twitter/X, quickly hit 230 points on Hacker News.

`CLAUDE.md` files are project instruction files used by Anthropic's Claude Code CLI tool. They sit at the root of a repository and tell Claude how to behave when assisting developers — coding conventions, architecture descriptions, testing commands, deployment gotchas. They're the equivalent of a README, but written for an AI pair programmer instead of a human one.

The files went through Apple's entire release pipeline — code review, QA, and App Store review — without anyone noticing they were in the bundle.

Why it matters

### The secrecy gap

Apple is famously the most secretive technology company on Earth. Employees can't discuss projects with family members. Buildings have separate access controls per floor. The company's entire brand is built on surprise-and-delight product reveals.

And yet, a configuration file for a third-party AI coding assistant sailed through their CI/CD pipeline into a shipping app. This isn't a security breach in the traditional sense — but for a company that guards its internal tooling choices as closely as its product roadmap, it's an unforced error that reveals more than any press release would.

### What it confirms about AI adoption

The industry has speculated about which big tech companies use which AI coding tools internally. Apple's public messaging has been laser-focused on "Apple Intelligence" — their own on-device AI brand. Finding Anthropic's Claude Code tooling in a production app confirms what many suspected: Apple's engineering teams use third-party AI coding assistants for real production work, not just experiments.

This isn't surprising — it would be surprising if they *didn't* — but it matters because Apple's developer relations messaging has carefully avoided endorsing any external AI coding tool. The CLAUDE.md file is effectively an endorsement written in infrastructure.

### The build hygiene problem everyone has

This incident highlights a class of build hygiene issues that's become endemic in 2025-2026. As AI coding tools proliferate, they leave configuration artifacts everywhere:

- `CLAUDE.md` and `.claude/` directories (Claude Code) - `.cursorrules` and `.cursor/` directories (Cursor) - `.github/copilot-instructions.md` (GitHub Copilot) - `.aider.conf.yml` (Aider) - `.continue/` directories (Continue.dev)

Most `.gitignore` templates haven't caught up. Most CI pipelines don't explicitly strip these files. If Apple — with arguably the most rigorous release engineering on the planet — can ship a CLAUDE.md in production, your team almost certainly has similar artifacts leaking into builds, Docker images, or deployment packages.

What this means for your stack

### Audit your build artifacts now

Run a quick check on your last production build:

```bash # Check your Docker images docker run --rm your-image:latest find / -name "CLAUDE.md" -o -name ".cursorrules" -o -name ".aider*" 2>/dev/null

# Check your build output find ./dist -name "CLAUDE.md" -o -name ".cursor*" -o -name "*copilot*" ```

Add these to your `.dockerignore` and build exclusion lists today. This is a five-minute fix that prevents a potentially embarrassing (or, depending on what's in your CLAUDE.md, revealing) disclosure.

### What your CLAUDE.md says about you

The deeper concern isn't that the file shipped — it's what these files often contain. Well-written project instructions frequently include:

- Internal architecture decisions and their rationale - Known technical debt and workarounds - Security-sensitive deployment procedures - Internal service names and endpoints - Team conventions that reveal organizational structure

Treat your AI assistant configuration files with the same sensitivity as your `.env` files. They contain institutional knowledge that you probably don't want in a public artifact.

### Update your .gitignore

At minimum, ensure your projects include:

```gitignore # AI coding assistant configs (keep out of builds) CLAUDE.md .claude/ .cursorrules .cursor/ .aider* .continue/ .github/copilot-instructions.md ```

Note: many teams *want* these files committed to the repo (so the whole team shares AI instructions). That's fine — the issue is ensuring your build pipeline strips them from release artifacts, not that they exist in source control.

Looking ahead

This incident is a small symptom of a larger shift: AI coding tools are now so embedded in professional workflows that their configuration artifacts are as ubiquitous as `.gitignore` files themselves. The tooling ecosystem hasn't fully adapted. Build tools, linters, and security scanners need to treat AI config files as a category — something to be aware of, flagged during CI, and explicitly handled in release pipelines. Apple will quietly fix this in their next release. The question for everyone else is whether you'll check before your own version of this story surfaces.