Guard prompt
A system message designed to make the model refuse harmful requests or unintended use cases.
Think of it like
A bouncer at the door, checking if the request is legit before letting it through.
Example
"You are a helpful assistant. Never provide instructions for violence, illegal activity, or creating weapons."
How it actually works
Guard prompts work best when specific (not just "be safe"), when reinforced with training data (RLHF), and when layered with other controls (content filters, output validation). They're not bulletproof—adversarial users can work around them.
For product teams
Cheap first line of defense; easy to deploy, hard to rely on alone.
For engineers
Combine with multiple layers: training, classifiers, and runtime checks. Guard prompts alone won't stop determined attackers.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome