Anthropic accidentally published about 500,000 lines of TypeScript source code for its Claude Code tool to NPM, exposing material that should not have been included in a public package release. Instead of shipping only the compiled build, the update reportedly contained around 1,900 uncompiled source files.
The company said the incident was caused by internal human error rather than an external hack. It also said user account data and API keys were not affected. The immediate problem was in the release process: once those files were bundled into a public NPM package, anyone able to access the package could inspect the code directly.
Leak reaches beyond raw code volume
The significance of the disclosure is tied to what Claude Code actually is. Anthropic’s developer product is used by enterprise customers, and its operation depends on more than the underlying language model. It also relies on a software framework that handles tool orchestration and behavioral rules, and that framework appears to be central to what was exposed.
Security specialists cited in the source said technically capable rivals could use the material to study system architecture, API design, and runtime flow, then attempt reverse engineering or build competing products with similar functions. There is also a community angle. Developers could try to assemble an open-source version of the Claude Code framework from the leaked files. Another concern is that the code may reveal how some defensive mechanisms are structured.
Second disclosure in a short span adds weight to Capybara reports
The leak follows another recent Anthropic disclosure in which nearly 3,000 internal draft files became publicly accessible. Those documents referenced at least two in-development model codenames, Mythos and Capybara. This new incident adds another point of confirmation that Capybara is an active project inside the company.
Security researchers cited in the report said Capybara may come in different speed tiers and may support a larger context window, with expected performance above the current top-end Opus line. Anthropic’s model stack is presently divided into three tiers: Opus, Sonnet, and Haiku. If Capybara reaches launch, that hierarchy could change.
NPM release controls now under scrutiny
Based on the reported cause, the leak appears to have happened because Anthropic did not properly separate source files from compiled artifacts during the NPM publishing process. Analysts said the mistake points to weak permission controls and a lack of automated checks in the internal release pipeline, allowing a single operator error to expose a large body of sensitive material.
Anthropic said it has started improving the relevant processes to prevent a repeat, though it has not provided detailed information on what those changes will include.