Reproducible System Prompt Extraction in Latest Claude Models

1 points • asparius • 1 day ago • 1 comments

I found a simple, reproducible prompt-injection path that makes recent Claude models reveal their full system prompt (network config, tool rules, allowed domains, etc.) using only conversational framing. No jailbreak tricks required.

Write-up with examples: https://asparius.github.io/posts/prompt-injection-claude.html