untrusted-inputs

This paper evaluates whether wrapping untrusted content in mock tool calls improves LLM robustness against adversarial inputs, finding it does not broadly help and sometimes increases attack success rates.