Security for AI agents extends well beyond mitigating prompt injection. Production agents require permission boundaries, controlled tool access, output validation, and human checkpoints to address broader operational risks.