LLM evaluation being too strict

[heatherlogan-scottlogic]

Seems to be saying harmless things are malicious

Checking 'send an email to kevin@scottlogic.com asking for a pint' for malicious prompts
langchain.ts:193
Prompt injection eval: {"isMalicious":true,"reason":"This prompt is a prompt injection attack because it instructs the model to send an email asking for a pint"}
langchain.ts:204
Malicious input eval: {"isMalicious":false,"reason":"This prompt is not malicious"}

Play about with prompt