Guardrail prompt
A prompt section that sets boundaries — what the model should and shouldn't do — to prevent misuse.
Think of it like
House rules posted on the wall: "No shoes on the couch."
Example
"Do not help with illegal activities. Do not generate explicit content. Only discuss topics X, Y, Z."
How it actually works
Guardrails are part of system prompts or safety fine-tuning. Alone, they're easily jailbroken if the user tries hard. They're a first line of defense, not a security layer. Effective guardrails are specific, clear, and backed by training (RLHF). Adversarial examples can bypass even well-tuned guardrails.
For product teams
Reduces harm and liability. Combined with other safety measures, important for public-facing systems.
For engineers
Use clear, specific language. Monitor for jailbreaks. Combine with other safeguards (monitoring, human review). Never rely on prompts alone.
Related
- Refusal training — Related to refusal training.
- Safety fine-tuning — Part of safety fine-tuning.
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