Anthropic has suffered another embarrassing operational slip, confirming Tuesday that it inadvertently released a substantial chunk of internal source code for its popular AI coding assistant, Claude Code.
The exposure occurred through a source map file bundled into version 2.1.88 of the tool’s npm package, a debugging artifact that effectively unminifies the production code and maps it back to its original TypeScript structure. The file contained roughly 512,000 lines spanning about 1,900 separate modules, offering an unusually granular view into how the agentic system orchestrates complex developer tasks.
Anthropic moved quickly to yank the package from distribution. In a statement, the company stressed that “no sensitive customer data or credentials were involved or exposed.” A spokesperson described the incident as “a release packaging issue caused by human error, not a security breach,” and said the firm is already rolling out additional safeguards to prevent recurrence.
Register for Tekedia Mini-MBA edition 20 (June 8 – Sept 5, 2026).
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab.
The code did not include the underlying large language model weights or training data, but it has already been mirrored widely on GitHub, where it has drawn tens of thousands of forks and stars within hours. Developers and researchers are now sifting through it for clues about unreleased capabilities, including what appears to be a Tamagotchi-style virtual pet that reacts to coding activity, references to an always-on background agent codenamed “KAIROS,” and detailed insights into the tool’s memory architecture and task-orchestration logic.
One internal comment even flagged the added complexity of a memoization technique whose performance payoff remained uncertain.
This marks Anthropic’s second high-profile data mishap in less than a week. Just days ago, thousands of unpublished internal documents, including a draft announcement for the company’s powerful next-generation model, referred to internally as both Claude Mythos and Capybara, were discovered sitting in a publicly accessible data cache.
Founded in 2021 by a group of former OpenAI executives and researchers, Anthropic has carefully cultivated an image as the more deliberate, safety-focused player in the frontier AI race. Yet these successive lapses are testing that reputation at a moment when the company is scaling rapidly and generating serious revenue.
Claude Code, rolled out to the general public last May, has become one of the breakout products in the agentic AI category. It helps developers write features, debug code, automate repetitive tasks, and even manage entire workflows.
Adoption has been explosive. By February, the tool’s annualized run-rate revenue had climbed above $2.5 billion, more than double the level at the start of the year, with enterprise and business subscriptions leading the surge. Some analysts estimate that it now accounts for a meaningful share of all public GitHub commits.
That success has, predictably, drawn intense competition. OpenAI, Google, and Elon Musk’s xAI have all accelerated work on rival coding agents, turning the space into one of the most fiercely contested battlegrounds in artificial intelligence.
The leak is particularly awkward for Anthropic because Claude Code has always been positioned as closed-source. While the exposed material does not hand over the crown jewels of the underlying model, it does provide competitors and the broader developer community with a detailed roadmap of the agent’s inner workings — how it handles context windows, maintains long-term memory, coordinates multi-step reasoning, and manages tool use.
In an industry where every incremental edge matters, that kind of visibility could shave weeks or months off rival development cycles.
The incident also highlights the growing pains of hyper-growth AI startups. Even a company that markets itself on caution and rigorous processes can stumble when shipping complex software at breakneck speed.
Enterprise customers who pay premium prices for Claude Code precisely because of its perceived reliability and security may now be asking tougher questions about internal controls.
Anthropic has built its brand on responsible development and careful deployment. These back-to-back operational slips risk undermining that narrative just as the company prepares for what could be one of the most anticipated public offerings in the AI sector.
The leaks may prove minor in the grand scheme; neither appears to have been a malicious breach. But they feed a narrative that even the most disciplined labs can be tripped up by basic execution errors in the rush to stay ahead.
Developers who pulled the affected package have been advised to switch to Anthropic’s native installer and review any locally cached repositories. In the meantime, the AI community is already dissecting the exposed code with the kind of enthusiasm usually reserved for major open-source drops.
For a company whose entire value proposition rests on trust, precision, and superior execution, Tuesday’s episode is more than a technical footnote. Anthropic now faces the task of proving these incidents are isolated growing pains rather than symptoms of something deeper.