A manual deployment error by Anthropic accidentally exposed the source code of its popular AI coding tool, Claude Code, to the public. The leak revealed internal project codenames and a proactive background agent named KAIROS. Anthropic clarified it was a packaging issue, not a security breach, and is enhancing automation to prevent future incidents. No employees were fired.

A missed step in a manual deployment process exposed the internal workings of one of AI’s hottest coding tools—and briefly handed the rest of the industry a detailed map of how Anthropic builds it. Claude Code, the agentic coding assistant that crossed $2.5 billion in annualised revenue just a year after launch, had its source code out in the open for anyone to download, fork, and study.Earlier this week, Anthropic accidentally bundled a debug source map file into a routine Claude Code update, pushing it to a public npm registry. Within hours, someone on X had posted a link to the full archive. The post crossed 30 million views. GitHub lit up with forks. The code—512,000 lines across nearly 1,900 files—was everywhere.Claude Code creator Boris Cherny confirmed the cause on X: a manual deploy step that didn’t get done. “Our deploy process has a few manual steps, and we didn’t do one of the steps correctly,” he wrote. Anthropic called it a packaging issue, not a security breach, and said no customer data or credentials were exposed.

Claude Code’s creator says nobody got fired, and the fix is more automation—not less

Early speculation pointed to Bun, the JavaScript runtime Anthropic acquired, citing a known bug where source maps get inadvertently served. Cherny shut that down quickly: unrelated, just developer error.When someone on X asked whether the person responsible was “still breathing,” Cherny didn’t flinch. “Full trust,” he wrote. “In this case the problem wasn’t the person, it was infra that was error prone. Anyone could have made this same mistake by accident.” No one was fired. The fix, somewhat counterintuitively, is to move faster: more automation, with Claude itself checking deployment results before anything goes out.

What developers found inside the Claude Code leak

Once the code was out, developers went through it methodically. They found references to unreleased model versions—Opus 4.7, Sonnet 4.8—and internal codenames like “Capybara” and “Tengu.” The most-discussed discovery was KAIROS: an always-on background agent that handles tasks proactively, logs actions daily, and runs a nightly routine called “autoDream” that reorganizes what it has learned. Cherny said Anthropic is still undecided on whether to ship it.There was also a Tamagotchi-style coding companion that sits beside the input box and reacts to your work, and a system that flags swear words as negative sentiment signals. Cherny confirmed the latter is real. “We put it on a dashboard and call it the ‘fucks’ chart,” he wrote on X.

Anthropic’s takedown sweep caught the wrong repos

Containment didn’t go cleanly either. Anthropic filed sweeping DMCA takedown requests that ended up hitting legitimate forks of its own open-source Claude Code repository—code that had nothing to do with the leak. Developer Theo, whose repo was caught in the sweep, called it “absolutely pathetic.” Cherny said it was unintentional, and Anthropic worked with GitHub to walk back the overbroad requests.