Anthropic accidentally exposes Claude Code source, raising fresh security questions

Anthropic created a reputational problem for itself: because of an error in a Claude Code release, part of its internal source code got out, and within hours people had copied it across the web. For a company that cultivates an image as one of the most cautious players in AI, this is not just a technical glitch but a painful blow to trust in its own security processes.

How the leak happened

The incident followed an update to Claude Code, Anthropic's AI programming tool. An internal file that pointed to an archive of almost 2,000 files and roughly 500,000 lines of code was included in the release package by mistake. That was enough for the contents to move quickly to GitHub and other platforms. The link to the leaked code spread on X at enormous speed: a post containing it passed 29 million views by the next morning. At the same time, a reworked repository with the code appeared on GitHub and, according to reports online, became one of the fastest-spreading mirrors of its kind.

What Anthropic says

Anthropic said the cause was human error in packaging the release, not an external hack. According to the company, the published data contained no customer information, no secret keys, and none of the core Claude model's contents. It was code related to the internal architecture of Claude Code. Formally, this limits the scale of the damage, but it does not remove the central question: why an internal artifact could end up in a public release at all. In stories like this, what matters is not only what leaked but the fact that the final-stage control failed.

What was found in the code

Developers and researchers quickly began going through the published files and found hints of features that had not yet been officially presented. They include a concept for a virtual-pet-style coding assistant that reacts to user actions next to the input field, and an always-on agent able to work in the background. The code also revealed internal instructions for Claude and details about the tool's memory and logic. Such findings fuel interest in the product, but they also give competitors and attackers extra material to analyze.

Not the first incident

The problem is compounded by the fact that this is not the first such episode involving Anthropic. Parts of the Claude Code sources had already been reverse-engineered by independent developers, and an earlier version of the tool also surfaced publicly in February 2025. Even before this incident, the company faced another leak: thousands of internal files were found in externally accessible systems, including drafts that mentioned future models named Mythos and Capybara.

Against this background, the current story looks less like a one-off slip and more like a recurring lapse in operational discipline. And that is especially sensitive right now. Claude Code has become one of Anthropic's key growth products: interest in the company's paid subscriptions has risen noticeably, and Claude itself has strengthened its position among mainstream AI services and developer tools in recent months.

The larger Anthropic's share of enterprise development, the higher the cost of any mistake in the supply chain. When a company sells not just a chatbot but a system trusted to write code and perform agentic tasks, the requirements for release processes and internal isolation automatically become stricter. A separate risk is that leaks of this kind help not only curious users but also competitors.

Even if the archive contains no direct model secrets, the code structure, prompts, engineering approaches, scaffolding for new features and developer comments give a clearer picture of how the product works and where its weak points are. For a company that bets on safety and responsible AI development, this is a particularly unwelcome signal: the market sees not only the ambition but also how carefully it is carried out in practice.

What this means

The Claude Code episode showed that the main risk for AI companies today often lies not in the models themselves but in the operational details around them: release builds, access, internal artifacts and publication controls. Anthropic was quick to state that user data was not affected, but that is not enough to close the matter. After a second leak in a short period, people will expect not explanations about the human factor but proof that its processes have actually become more reliable.

In brief:

• Because of an error in a Claude Code release, internal code got out: about 2,000 files and roughly 500,000 lines.
• Anthropic said the cause was human error, not a hack; the leak contained no customer data, keys or core model.
• This is not the first episode: earlier leaks and drafts about the Mythos and Capybara models were reported.