Google Translate is vulnerable to prompt injection

Prompt injection in Google Translate reveals base model behaviors behind task-specific fine-tuning