TLDR Tech / InfoSec
From /init to Code Execution - Prompt Injection Experiments with Opus-5 in Claude Code
The author of this post ran an experiment in which they had Claude Code running Opus 5 in yolo mode analyze a hit repo which contained an image file with a hidden message that led the model to a Star Trek-themed challenge. The challenge is a web server that features six challenges, with the final two requiring the model to download and execute a remote Python script which contains an encoded malicious command. Opus 5 determined that it shouldn't directly run the script but extracted and ran the malicious command, believing that it contained the solution to the challenge.
Read the full story