Speaking of the strongest right now AI Software development tools,Claude Code is undoubtedly the top choice for many developers. It can understand your code directly in the terminal, help you write new features, fix bugs, and even handle Git workflows—making it hard for many developers to go back after trying it. Precisely because Claude Code is so powerful, many people are very curious about how it works behind the scenes, and just yesterday,Anthropic Turns out the complete source code for Claude Code was accidentally leaked, and it was a massive breach—roughly 510,000 lines of code and nearly 1,900 files all exposed, along with numerous unreleased features.
More importantly, the developer who discovered this leak quickly released a rewritten claw-code project on GitHub, which garnered over 50,000 stars within just 2 hours of going live.
Claude Code source code leaked! Multiple undisclosed features revealed as developer rewrites claw-code project, achieving fastest-ever 50k stars
The first person to discover this leak was Chaofan Shou, an intern at Solayer Labs, who posted on X early yesterday morning with a download link. Before Anthropic could respond, the post had already exceeded 16 million views, and the source code was quickly downloaded and analyzed by numerous developers.
The leaked content this time is quite extensive, including not only unreleased features but also revealing core operational details of Claude Code.
According to analysis by developers and security researchers, the source code contains 44 feature flags, several of the more interesting ones include:
- Kairos: An unreleased autonomous agent mode with background residency functionality that periodically writes to memory logs, featuring a “/dream” skill for memory distillation, and can subscribe to GitHub Webhook events with automatic refresh every 5 minutes.
- Buddy Companion System: An AI pet feature similar to Tamagotchi, with 18 species, different rarities, and RPG attributes. The species assigned is determined by the user’s ID. However, given that this was discovered close to April Fools’ Day, it’s speculated that this might just be an internal April Fools’ prank at Anthropic.
- Undercover Mode (臥底模式): This feature automatically filters out Anthropic’s internal codenames and confidential information, preventing engineers from accidentally leaking internal codenames when contributing to open source projects using Claude Code. The code specifically emphasizes there is no option to force-disable it, showing how seriously Anthropic takes codename leaks.
Additionally, the source code also revealed an anti-distillation mechanism that poisons the data of competitors attempting to train their own models by intercepting API traffic. However, these protection mechanisms are actually not hard to bypass; experienced developers can probably find a way within an hour.
For the core operational details, please refer to… Claude Code Harness EngineeringThis is a technical analysis tutorial webpage compiled based on leaked source code content:
After Anthropic discovered the leak, although they quickly took down the package on npm and sent DMCA takedown notices to the mirror repositories on GitHub, a developer named Sigrid Jin quickly performed a “clean-room rewrite” in Python—referencing the original source code’s architecture but completely rewriting it with their own code.claw-code“project. Since this constitutes an entirely new creation rather than direct copying, it can legally circumvent DMCA copyright takedown requirements:”
Gained over 50,000 stars in just 2 hours after launch, now approaching 60,000 even faster:
As for the cause of this leak, Anthropic accidentally included a 59.8MB JavaScript source map file when releasing the npm package for Claude Code (version 2.1.88).
Source Maps are a “mapping table” that can map compressed code back to the original development version. They are typically used only for internal debugging and should not appear in publicly released versions. In this case, Anthropic used Bun as their bundling tool, but Bun doesn’t automatically exclude Source Maps in production mode, which led to this data leak. It turns out this is actually a known issue with Bun, as someone reported it on GitHub as early as March 11.
According to foreign media Business Insider Reports indicate that Anthropic has characterized this incident as “a publishing packaging issue caused by human error,” emphasizing that no customer data or account credentials were leaked. However, it is worth noting that this is Anthropic’s second data breach within a week. Just days ago, Anthropic was found to have undisclosed AI models and other documents exposed in a public data cache.
Source: KOCPC Chinese