Anthropic accidentally leaked a significant portion of its Claude Code AI product, revealing a 59.8 MB JavaScript source map file on the public npm registry. This incident, discovered by an intern and quickly shared online, exposed approximately 512,000 lines of TypeScript code. For Anthropic, which boasts a $19 billion annual revenue run-rate, this leak is more than a simple error; it risks their competitive edge by providing rivals, from established companies to nimble startups, a detailed blueprint of their technology. The leaked code highlights Anthropic’s innovative approach to memory management with a three-layer architecture designed to minimize confusion during complex tasks. Their "Self-Healing Memory" system uses a lightweight index to track information without cluttering the model's context. It also emphasizes a "Strict Write Discipline," ensuring updates only occur after successful data writes. Another key feature is "KAIROS," allowing Claude Code to function as an always-on background agent, which optimizes user experience during idle time. Moreover, the leak sheds light on unreleased internal models and performance metrics, revealing ongoing development struggles, such as a high false claims rate. The source code also includes a feature called "Undercover Mode," which allows Claude Code to contribute to open-source projects without revealing its internal workings. This leak not only exposes Anthropic’s intellectual property but also poses security risks for users, who may be vulnerable to attacks exploiting the newfound knowledge of Claude Code's architecture. Immediate caution is advised for users who updated their packages around the time of the leak, as a separate attack on the axios package may have introduced malicious software into their projects.