Evaluating using Mock Tool Calls to Quarantine Untrusted Prompt Inputs

In three linesarXiv study on LLM security against untrusted inputs. Researchers test whether wrapping untrusted content in mock tool calls improves robustness across 7 models and 3 LLM-as-a-Judge tasks. Finding: the approach fails and typically increases attack success rates, inverting the expected instruction hierarchy.

Summary generated by Claude — human-verified