Adam Gleave

The paper tested the hypothesis that wrapping untrusted prompt inputs in mock tool calls would improve LLM robustness, but found that this technique generally fails and can even increase vulnerability.