Safety
Guardrails
Rules and filters that constrain what an AI model can output — used to block harmful, off-topic, or non-compliant responses.
01 ——
In plain English
Guardrails are the safety layers built around an AI model to keep its outputs appropriate, on-topic, and within policy. They're what stops a customer-support bot from giving medical advice, or a kids' app from producing adult content.
Where guardrails sit:
- In the prompt — system instructions ("Refuse to discuss competitors")
- In the model itself — RLHF training that builds in refusals
- Around the model — input/output classifiers that filter requests and responses
- At the orchestration layer — rules about which tools the model can call
Common guardrail goals:
- Block harmful content (violence, illegal advice, CSAM)
- Prevent prompt injection attacks
- Keep the bot on-topic for its use case
- Protect privacy (redact PII)
Off-the-shelf libraries (Guardrails AI, NeMo Guardrails) and major providers (OpenAI moderation, Anthropic Constitutional AI) ship guardrail tooling.
02 ——
Related terms
Alignment
The challenge of making AI systems behave in ways that match human values and intentions — not just their literal instructions.
Prompt Injection
A security attack where malicious instructions hidden in user input or external content trick an AI model into ignoring its real instructions.
Jailbreak
A prompt or technique that tricks an AI model into ignoring its safety rules and producing content it would normally refuse.
RLHF
Reinforcement Learning from Human Feedback — the training technique that teaches AI models to be helpful, harmless, and honest.
Bias
When an AI model's outputs systematically reflect unfair patterns from its training data — about gender, race, age, or other groups.
Back to glossaryLast reviewed May 2026