What it means
Guardrails are the checks wrapped around a model rather than baked into it: screening inputs for prohibited requests, screening outputs for unsafe or off-topic content, validating format, restricting which tools can be called, and enforcing rate and spend limits.
The critical distinction is between guardrails and instructions. Telling a model in its system prompt not to discuss competitors is a preference it can be talked out of. A separate classifier checking output before it reaches the user is a control that runs regardless of what the model was persuaded to produce.
The tuning problem is real in both directions: strict guardrails block legitimate use — medical questions refused for a healthcare product, security topics refused for a security team — while loose ones let harmful output through. Over-refusal is the more common complaint in practice.
Why it matters
Guardrails are where AI safety becomes engineering rather than aspiration. Anything that genuinely must not happen belongs in code, because a model's compliance with instructions is probabilistic and adversarially defeatable.
In practice
Enforce the rules that matter outside the model. Log what guardrails block and review it — the blocked set is where you discover you have made the product unusable for its actual audience.
Where this shows up
Tools and models in our catalog.
ShieldstralMistral's 3 billion parameter open-weights safety classifier. Judges text and images against moderation policies written in plain language at inference time, under Apache 2.0.
Patronus AIAI evaluation and agent-testing platform: scores LLM outputs for hallucinations and safety, benchmarks them on custom criteria, and stress-tests AI agents in simulated "Digital World" environments before they reach production.