2026-06-30

Claude Code’s leaked source and the cost of “100% AI-written”

If you build with Claude Code, this story lands a little too close to home. It’s not just about a source leak; it’s about what happens when an AI-first engineering culture is allowed to set its own standards, define its own metrics, and then congratulate itself for speed while the codebase quietly turns into a maintenance hazard.

Key Points

Anthropic leadership repeatedly described Claude Code and other Anthropic products as increasingly, and eventually fully, AI-written.
A packaging mistake later exposed a large chunk of Claude Code’s source, and the leak revealed some ugly realities.
The leaked code included a 3,167-line function in print.ts, plus very large files such as QueryEngine.ts, Tool.ts, and commands.ts.
One internal prompt/sentiment check used regex for frustration detection: /\b(wtf|shit|fuck|horrible|awful|terrible)\b/i.
A comment in autoCompact.ts said 1,279 sessions had 50+ consecutive failures, wasting about 250K API calls per day globally.
The response to that bug was apparently a small code change: cap failures and disable compaction for the session.
Community reports also described serious memory issues, including processes using multiple gigabytes of RAM and, in some cases, far more.
The issue tracker was heavily automated, with bots handling deduplication, staleness, and closure; the article cites an estimate that 49 to 71% of issue closures were bot-driven.
Anthropic’s reaction to the leak focused on the deploy mistake and speeding up the process, with more automation and Claude checking Claude’s work.
The article argues that this creates a self-reinforcing loop: AI writes the code, AI reviews the code, AI checks the deployment, and when things break, the answer is more AI.

My Take

What strikes me is not that the code leaked. Leaks happen. The interesting part is that the leak let outsiders verify what “100% AI-written” actually looks like in production, and the answer seems to be: very large files, awkward structure, and a willingness to ship known problems because the system can absorb the waste. I think that’s the real story here, not the packaging error.

The regex sentiment check is the detail that made me wince. Sure, regex is cheap and fast, and I can see why someone would reach for it. But in a company selling frontier model quality, it reads as a little embarrassing. That doesn’t mean it’s wrong in isolation. It means the rhetoric and the reality are drifting apart.

The more uncomfortable bit is the cultural signal. If the philosophy is “don’t review, regenerate,” then code review stops being a quality gate and becomes a kind of ceremonial checkpoint. That might work for a while when you have lots of compute and a lot of money. I think it gets fragile fast once the defects pile up and the costs stop being abstract.

As a Claude or Claude Code user, I’d be curious whether this approach actually scales beyond a narrow slice of product engineering. Maybe it does in some cases. Perhaps the team really can move faster by leaning hard on AI-generated code and AI-generated review. But if the leaked source is representative, I wouldn’t call the result inspiring. I’d call it aggressive, clever in spots, and visibly under-disciplined.

If I were using Claude Code seriously, I’d still use it. I just wouldn’t mistake generated velocity for engineering maturity. The takeaway here is pretty simple: AI can amplify discipline, but it also amplifies sloppiness. And sloppiness at machine speed is expensive.