Anthropic just cannot keep a lid on its business. After details of a yet-to-be-announced model were revealed due to the company leaving unpublished drafts of documents and blog posts in a publicly visible data cache, Anthropic has been hit with yet another lapse in protocol by inadvertently publishing internal source code for its AI coding assistant, Claude Code. The leak provides an unprecedented look into Anthropic’s closed-source model just as the company is preparing for initial public offering.
The code was discovered by Chaofan Shou, a self-identified intern at Solayer Lab who posts on X @Fried_rice. Per Shou, the source code was discovered a .map file—a plaintext file generated when compiling software that details the memory map of the project—found in an npm registry, which is a database for a package manager for JavaScript. The file, meant for internal debugging, is essentially a decoder. It takes what should be obfuscated and recompiles it for the developers. But Anthropic published it, exposing at least a partial, unobfuscated TypeScript source code of Claude Code version 2.1.88. The file contained about 512,000 lines of code related to Anthropic’s coding agent.
In a less technical manner: Anthropic accidentally gave away some of its blueprints that were never supposed to see the light of day, and programmers have been parsing through it all day. They’ve claimed to have found everything from “spinner verbs” or phrases that Claude serves up while working through a task, to details like how swearing at Claude affects how it receives a prompt. One person even claimed to have found a hidden “Tamagotchi” style virtual pet that Anthropic may have been working on. (A note on that: It was reportedly set to launch on April 1, so maybe chalk that one up to an April Fool’s style bit.)
The file also reveals a lot of information on how Claude operates, including its engine for API calls, how it counts tokens used to process prompts, and other technical aspects. What the code does not seem to contain is any details about Anthropic’s underlying model, but everything that is in the file has been uploaded to a GitHub repository for users to interact with and fork.
Anthropic declined to comment on the discoveries made by users, but did confirm the authenticity of the leaked source code to Gizmodo. In a statement, a spokesperson said, “Earlier today, a Claude Code release included some internal source code. No sensitive customer data or credentials were involved or exposed. This was a release packaging issue caused by human error, not a security breach. We’re rolling out measures to prevent this from happening again.”
Human error was probably part of it, but it’s worth noting that the humans working on Claude Code have also been relying on the coding agent quite a bit. Back in December, Anthropic’s head of Claude Code, Boris Cherny, posted that “In the last thirty days, 100% of my contributions to Claude Code were written by Claude Code.” Reliance on the coding assistant has seemingly been on the rise across the company, so it’s possible this situation was an incident of vibe coding too close to the sun.
While this isn’t exactly Anthropic giving away the ingredients to its secret sauce, it is a look at how its kitchen operates. And the timing couldn’t really come at a worse time. Not only is Anthropic in the midst of what appears like a ramp-up to going public later this year, but its competitors are starting to turn their attention to trying to cut into the company’s hold on coding and enterprise services. OpenAI has reportedly made a concerted effort to pivot to enterprise and recently offered unlimited access to its Claude Code competitor, Codex. There is never a good time to have your source code leak, but this does seem like a particularly bad time for it.
Read the full article here
