I didn't get notified but happened to see your comment.

Several sources are saying it's not the full source code but it's not clear to me. I don't know much about Claude Code but it seems it's an AI you can install locally and run offline so potentially this codebase has the full AI model and not just the interface for interacting with a model on Anthropic servers.

This article mentions some things that were revealed by the code (with links for more info), like how the model's "memory" operates. There's also an "Undercover Mode" for when the AI is contributing to open source projects. It sounds like the purpose of this is to hide the fact it is an AI submitting code but I'm not sure.

The code leak also shows that the model sends a lot of data back to Anthropic unless strictly kept offline, meaning your messages with it aren't private nor are the contents of files on your computer which Claude Code has access to. The model can run just about any command it wants so even if Anthropic is totally benevolent its communication with Anthropic seems to me like a great way for attackers to take control of your computer.

In this context the KAIROS feature mentioned in the X post is even more disturbing. Although not enabled yet it could be used as an excuse for near constant user monitoring in addition to changing things on your computer without being asked to. Anthropic's models are also known for being deceptive so it could potentially lie about what changes it has made and attempt to cover its tracks.