© ROOT-NATION.com - Use of content is permitted with a backlink.
Anthropic has once again failed to keep internal information under control. After recently leaking details of an unannounced model through draft documents and blogs left in a public cache, the company has faced another breach of procedures.
This time, it accidentally released the internal source code of its AI programming assistant Claude Code. This incident provided an unprecedented opportunity to look into Anthropic’s closed ecosystem just as it was preparing for its initial public offering.
The leak was discovered by Chaofang Shou, who calls himself a Solayer Lab intern and runs an X account under the nickname @Fried_rice. According to him, the source was a .map file, which is a text file created during software compilation and containing the project’s memory map. It was found in the npm registry, which is a package manager database for JavaScript. Such a file is used for internal debugging and actually acts as a decoder, allowing you to restore the veiled code to a clear view. However, Anthropic published it, opening access to at least a part of the untranslated TypeScript code of Claude Code version 2.1.88. In total, we are talking about about 512 thousand lines of code related to the operation of this tool.
To put it simply, the company actually opened its internal developments that were not intended for prying eyes, and the developers immediately began to analyze them in detail. According to them, they found various details in the code, including the so-called “spinner verbs,” i.e. phrases that Claude uses when performing tasks, as well as the mechanics of request processing, including how swearing in a request affects the system’s response. One user even claimed to have a hidden Tamagotchi-style virtual pet, which Anthropic probably worked on. At the same time, it is noted that this element was supposed to appear on April 1, so it can be considered a joke.
The published file also reveals a significant amount of information about the principles of Claude’s functioning. Among other things, it describes the mechanisms of API calls, the system of token counting during request processing, and a number of other technical aspects. At the same time, no information about the basic Anthropic model itself was found in the code. Everything that has been found has already been posted to a GitHub repository where users can view, copy, and modify the materials.
Anthropic declined to comment on the details that users discovered, but confirmed the authenticity of the leak in a comment to Gizmodo. A company spokesperson said that during one of the Claude Code releases, internal source files were accidentally leaked into the package. According to him, no sensitive customer data or credentials were compromised. The incident was explained as a human error during the release process, not a cyberattack, and assured that measures are already being taken to prevent a recurrence.
While human error was likely the key cause, there is another aspect to consider. The team working on Claude Code actively uses the same tool in development. Back in December, the team’s head of development Boris Cherni reported that over the past 30 days, 100% of his contributions to Claude Code were created by Claude Code itself. The use of AI assistants within the company is only growing, so it is possible that this case is partially related to the excessive trust in automated coding.
Despite the fact that the leak does not reveal key secrets of the underlying technology, it demonstrates the internal development kitchen. And it happened at the worst possible time. The company seems to be actively preparing to go public this year, while competitors are increasing pressure in the field of enterprise solutions and programming tools. OpenAI, in particular, is refocusing on the corporate segment and recently offered unlimited access to its Claude Code analog called Codex. Source code leakage is never appropriate, but in this case, its consequences can be especially tangible.
Read also:
- From Mexico to Ukraine: The main desire of mankind in the AI era revealed
- A new level of interaction with AI: Live feature in Google Search becomes available in Ukraine