Guardrail model

When a company runs an AI chatbot, the chat model itself is rarely the only thing in the loop. A second, much smaller model reads each incoming message and often each outgoing answer, and flags anything that violates the operator’s policy. That checker is the guardrail model, and it is why a chatbot sometimes stops mid-sentence or declines to continue.

Older guardrail models were trained on a fixed list of categories, so adding a new rule meant retraining. Newer ones, such as Mistral’s Shieldstral or OpenAI’s gpt-oss-safeguard, take the policy as plain text at the moment of the check, which makes them adjustable without touching the weights.