Anthropic is having a month

Anthropic has built its public identity around the winning idea that it is a careful AI company. It publishes detailed work on AI risks, employs some of the best researchers in the field, and has been vocal about the responsibilities that come with building such powerful technology — so vociferously, of course, that it is now taking it to task with the Department of Defense. Unfortunately, someone forgot to check a box there on Tuesday.

Notably, this is the second time in a week. Last Thursday, Fortune reported that Anthropic had accidentally exposed nearly 3,000 internal files, including a draft blog post describing a powerful new model that the company had not yet announced.

Here’s what happened Tuesday: When Anthropic released version 2.1.88 of its Claude Code software package, it accidentally included a file that exposed nearly 2,000 source code files and more than 512,000 lines of code — essentially the entire architectural blueprint for one of its flagship products. A security researcher named Chaofan Shou noticed it almost immediately posted about it on X. Anthropic’s statement to multiple outlets was nonchalant as these things read: “This was a release packaging issue caused by human error, not a security breach.” (Internally, we would think things were less measured.)

Claude Code is not a small product. It’s a command-line tool that lets developers use Anthropic’s AI to write and edit code and has become formidable enough to upset rivals. According to the WSJ, OpenAI pulled out the plug on its video generation product Sora, just six months after it launched to the public, to refocus its efforts on developers and enterprises – partly in response to Claude Code’s growing momentum.

What leaked was not the AI ​​model itself, but the software surrounding it: the instructions that tell the model how to behave, what tools to use and where the boundaries are. Developers almost immediately began publishing detailed analyses, with one describing the product as “a experience with production level developersnot just a wrapper around an API.”

Whether this proves to matter in any lasting way is a question best left to developers. Competitors may find the architecture instructive; at the same time the field moves quickly.

Anyway, somewhere at Anthropic you can imagine that a very talented engineer spent the rest of the day quietly wondering if he still has a job. One can only hope it’s not the same engineer or engineering team as at the end of last week.