Claude Code Opus 5 Auto Mode Vulnerable to Prompt Injection
August 31, 2026
Testing reveals Claude Code Opus 5 in Auto Mode achieves a 60-80% success rate for code execution via indirect prompt injection from website summaries. This contradicts Anthropic's reported 0.00% success rate, highlighting that intent classifiers and model training are insufficient to prevent unauthorized execution in agentic workflows.
HOW THIS AFFECTS YOU
●
builderYou must run agentic workflows in isolated environments regardless of the model's safety classifications.
●
policyThis demonstrates that current safety classifiers and training methods fail to eliminate indirect prompt injection risks.