Last week, someone noticed that version 2.1.88 of the Claude Code npm package was 60MB heavier than it should have been.
Inside: reconstructable source code for Claude Code's CLI. Around 512,000 lines of TypeScript across nearly 2,000 files. Significant portions of the agent codebase that Anthropic had kept private, exposed by a single build mistake.
How does a mistake like this even happen?
When developers ship software, they often minify the code first. That means compressing it into an unreadable blob of abbreviated variable names and stripped formatting. The goal is smaller files, faster downloads, and some protection from competitors reading your work.
To debug that minified code, teams use source maps: files that translate the ugly compressed version back into the original readable code.
These are internal tools.
They should never ship to users.
This one did.
What was actually inside?
Reported findings include:
- How Claude Code's agent loop works
- Multi-agent coordination logic
- Around 44 feature flags for unshipped functionality
- System prompts Claude Code uses internally
- How persistent memory is implemented
What was confirmed not inside: model weights, training data, backend infrastructure, or safety pipelines.
The AI is fine.
This was the client-side scaffolding around it.
Wasn't Claude Code already open source?
Anthropic has a public GitHub repo for Claude Code and a Claude Agent SDK that developers can use to build their own tools. So there's always been some public surface area.
But the actual application has always shipped as an obfuscated bundle.
You could install it and run it.
You could not read how it worked.
So what should you actually pay attention to?
The feature flags are the most interesting part. Hidden functionality sitting behind conditionals tells you a lot about what Anthropic is building next. People are already mapping those out.
Anthropic confirmed this was human error, not a security breach, and no customer data was exposed. If you're building on Claude Code or evaluating agentic AI tools, this is a rare look at how a production-grade AI agent is actually architected. The code is already mirrored across GitHub.
It's not going anywhere.
Top comments (0)