4 comments

[ 2.7 ms ] story [ 15.6 ms ] thread
I’m surprised it’s seemingly so easy for folks to get the LLMs to divulge their system prompts.
I just put an HTTP proxy in front of Claude Code.

Surprisingly, it just accesses their `/v1/messages` endpoint - nothing hidden at all.

Wow. I assumed you had to break its safety system to get it to reveal the prompt.
(comment deleted)