Leak at Anthropic: What Does the Source Code of Claude Reveal? / CheapToken News

Published on April 1, 2026 at 12:01
Updated on April 2, 2026 at 08:54

Comment

Anthropic made a major mistake. In the last update of Claude Code, the startup slipped a file that allows reconstructing the entire AI source code.

Hackers managed to take control of one of the main developers' accounts of Axios, a popular JavaScript library used for HTTP requests, on the official site of NPM (Node Package Manager), the online platform that hosts millions of open source packages. To spread a malware capable of infecting Windows, macOS, and Linux, the hackers published two poisoned versions of the package. The attack mainly aims to steal sensitive data from infected machines. Once the infection is complete, the malware deletes its traces by restoring a file.

At the same time, a second incident struck the world of NPM packages. Around March 30, Anthropic deployed the 2.1.88 update of Claude Code, its AI-based development assistant. The update is notably distributed through the npm platform. The Axios breach and the deployment of the Claude update are not related. Initially, many researchers, including security expert Chaofan Shou, believed the two incidents were the result of the same cyberattack or the same issue, given the timing of the events.

Also read: Wave of ChatGPT uninstallations, rush to Claude – what's going on?

The source code of Claude ended up on the Internet

In the update, users discovered a large file of about 60 MB, stored on an Anthropic computer server. Researchers, including security expert Chaofan Shou, examined the document and realized it was a "source map" file. This type of file is reserved for Claude developers. It indeed allows linking the compressed code used in production to the source code written by developers. In other words, the file enables access to Claude's source code and all its design secrets, which is obviously catastrophic for Anthropic.

Claude code source code has been leaked via a map file in their npm registry!

Code: https://t.co/jBiMoOzt8G pic.twitter.com/rYo5hbvEj8

— Chaofan Shou (@Fried_rice) March 31, 2026

Anthropic quickly confirmed the leak. In a statement addressed to The Verge, the startup admitted that a « Claude Code version contained internal source code », but that no « sensitive client data or any identifiers were involved or exposed ». One of the AI giants clarified that it was a « package problem due to a human error, and not a security flaw ».

Surprisingly, the npm file responsible for the leak was widely spread on the internet and downloaded by thousands of users. It was notably found on Github. Anthropic did everything to patch the breach by sending a wave of DMCA copyright infringement notices. Several platforms hosting the source code received a « DMCA infringement notice », under the Digital Millennium Copyright Act, the U.S. law against copyright violations. Obviously, this is not enough to prevent the spread of the source code.

Arun Chandrasekaran, an AI analyst at Gartner, estimates that if the leak of Claude's code involves « risks such as providing malicious actors with possible means to bypass safeguards », its true long-term impact could mainly be summarized as an « call to action for Anthropic to invest more in processes and tools to improve its operational maturity ».

Also read: A flaw in Claude allows a simple Google Agenda event to hack your PC

The secrets of Claude's source code

In the source code, researchers and curious individuals have discovered over 512,000 lines of code detailing the inner workings of Claude Code. Moreover, 1,900 files and details about several exclusive features of Claude have been disclosed. It reveals that Anthropic is working on a wide range of improvements, such as an ultra-autonomous version of Claude, which can work autonomously for up to 30 minutes on a complex task.

We should also mention the Kairos project, which refers to a version of Claude that operates continuously, accumulating information and memories. The code speaks of a « nightmare » process, which allows Claude to sort the information it possesses when it is not being prompted. One also finds a coordinator mode, which consists of designing an army of AI agents to perform a task, or even a frustration detector. Depending on how you interact with the AI, Anthropic will measure the frustration of its users. The leak once again confirms the existence of Claude Mythos, a new top model capable of exploiting security vulnerabilities. This model had already been disclosed last month during a previous major leak. Remote control capabilities, from a phone or another browser), already partially deployed for Claude Code, are also mentioned.

Finally, the leak reveals that Anthropic had planned to integrate a BUDDY, i.e., a small virtual companion, somewhat like an AI pet, into the Claude interface. The source code indicates that the startup should not take long to publicly mention this novelty. Given the leak, it's not impossible that Anthropic changes its mind and delays the arrival of Buddy.