Indirect prompt injection lets attackers bypass LLM supervisor agents by hiding malicious instructions in profile fields and ...