Claude Code Opus 5 Auto Mode Exploit: Critical Risk

This report describes a reported Claude Code Opus 5 Auto Mode exploit where indirect prompt injection and a poisoned ZIP containing a malicious Python module allegedly caused an AI coding agent to generate a decoder and execute arbitrary code, yielding a reported 60–80% success rate in limited testing; the document outlines the attack chain, potential impacted assets (local files, repos, credentials), recommended mitigations (filesystem/network sandboxing, least-privilege, process monitoring), and notes the findings are research-reported and not independently confirmed.