Claude Code Leak Was a ‘Manual Error’, and ‘No one Was Fired’

Boris Cherny, the creator and Head of Claude Code at Anthropic, has confirmed that there was a leak of the company’s internal Claude Code source code. The accidental release exposed more than 500,000 lines of code to the public. 

Cherny attributed the incident to a breakdown in the release process rather than a security breach, and offered a series of explanations on X.

“It was a human error,” confirmed Cherny. “Mistakes happen, as a team, the important thing is to recognise it’s never an individual’s fault—it’s the process, culture or the infrastructure.”

Cherny also confirmed that no individual has been fired. 

The leak occurred when a debug or source map file was mistakenly bundled into a routine software update and pushed to a public package registry used by developers.

The exposed code revealed Claude Code’s internal architecture, feature flags, and even unreleased capabilities. 

Developers quickly spotted it after the update went live, with some sharing links and screenshots on X and GitHub. Within hours, the code was widely circulated, reverse-engineered, and even forked into public repositories, giving a detailed look at how Claude Code works internally.

The Claude Code creator explained that Anthropic’s deployment process entails a few manual steps, and the company didn’t perform some of them correctly. “We have landed a few improvements and are digging in to add more sanity checks,” said Cherny. 

“Like with any other incident, the counter-intuitive answer is to solve the problem by finding ways to go faster, rather than introducing more process,” he added, suggesting that the process may now involve more automation and Claude itself checking the results. 

Separately, some developers initially speculated that the issue may have been caused by the JavaScript runtime Bun, the open-source framework that was recently acquired by Anthropic. Cherny directly addressed this, stating that the leak was unrelated to Bun and was purely a developer error during deployment.

He also discussed the findings with developers on X who were curious about what they had spotted in the leaked code. 

A developer pointed to an internal feature flagged as “KAIROS.” Based on leaked code, it appears to be an always-on, proactive version of Claude that can run continuously in the background, taking action without explicit user prompts.

Responding to the speculation, Cherny said, “We’re always experimenting with new ideas. 90% don’t ship because we don’t think they’re good enough experiences. Still on the fence about this one—should we ship it?”

Developers suggested KAIROS could autonomously fix bugs, respond to messages, update files, and even monitor workflows such as GitHub pull requests. The feature also appeared to include capabilities like push notifications, file delivery, and persistent logging of actions, along with a nightly process referred to in the code as “autoDream” that reorganises learned information.

Separately, developers noted that the code includes regex-based detection for frustrated or abusive user inputs, which are tagged internally as negative signals.

Cherny confirmed this, explaining that Anthropic uses such signals as part of its product feedback loop. 

“This is one of the signals we use to figure out if people are having a good experience,” he said, adding that the company tracks instances of explicit language from developers as a direct indicator of frustration.