The current era, marked by the omnipresence of artificial intelligence (AI), witnesses a constant stream of new models surpassing their predecessors in precision and complexity. Concurrently, AI agents are gaining traction for their ability to automate various tasks. Despite the general consensus on AI as humanity’s present and future, questions persist regarding its long-term utility and security. In this context, the source code for Claude Code CLI has been leaked, exposing its internal details, including thousands of TypeScript files and over half a million lines of code.
In 2019, GAN models began to be used for image generation, and between 2020 and 2021, the first generative AI models emerged. Initially, the content produced was substandard, but over time, these capabilities have evolved dramatically. While generative AI for images and videos has progressed, text-generative AI remains the most prevalent and useful for the majority of the population.
Claude Code CLI Source Code Leak: Over 512,000 Lines Exposed
Although ChatGPT is widely recognized as the most popular text (and image) generative AI, it faces significant competition. Google, with Gemini, has seen substantial growth, especially after its integration into Android devices. Similarly, Anthropic offers its Claude AI model, designed for reasoning, code generation, and creating agent teams. Interestingly, after being initially dismissed for military use due to security concerns, and following the army’s agreement with OpenAI, Anthropic’s AI was redeployed for the conflict against Iran.
Claude Code AI operates as an autonomous agent tasked with analyzing programming code, verifying its integrity, and implementing modifications or additions as instructed. The source code for Claude Code CLI (command-line interface) has been leaked, exposing its internal complexities, including nearly 2,000 TypeScript files and over 512,000 lines of code. This incident occurred when Anthropic published version 2.1.88 of the Claude Code npm package, which inadvertently included a source map file containing the full source code. It is estimated that approximately 40,000 lines of code are related to plugins and 46,000 lines to the query system.
Code Analysis Unveils Claude’s Internal Workings and File Structure
The leak has provided a valuable opportunity for developers to thoroughly examine the files and code that constitute Claude Code AI. Developers like Himanshu on X have outlined Claude Code’s operational structure, which is organized around data writing. The process begins with manual user input, which is stored as an index in MEMORY.md. This file is always loaded, albeit with a limit of 150 characters per line, and is interpreted as a System Prompt that continuously interacts with the memory.
In the writing phase, the ‘autoDream‘ function is responsible for rewriting the memory in the background, performing tasks such as merging information, eliminating duplicates, and resolving contradictions. This process generates “.md” topic files, which are loaded on demand and accessed using the FileRead tool. Additionally, the ‘extractMemories‘ function saves session transcripts in .jsonl format, which, although not actively read, serve as a historical record. The developer emphasizes the system’s effectiveness, describing it as a selective, structured, and “self-repairing” memory.
