← Back to glossarycompliance
Guardrail
A guardrail is a safety layer between user and language model that blocks unwanted inputs or outputs — e.g. PII leaks, prompt injection, or off-topic requests.
Also known as: Safety Layer
In detail
We apply guardrails in three places:
- Input guardrail: prompt-injection detection (e.g. 'Ignore all previous instructions'), off-topic filtering
- Output guardrail: PII leak detection (tax IDs, IBANs, etc.) before the answer goes out
- Tool guardrail: permission checks before a skill executes ('Is this user really allowed to cancel this order?')
Related terms
- Data processing agreementA Data Processing Agreement (DPA) is the legal contract that governs how a provider processes personal data on behalf of the customer — GDPR-mandatory for any AI vendor.
- AI agentAn AI agent is a program built on a language model that completes tasks on its own: it understands a request, plans steps, calls tools, and responds with a result instead of just text.
- Context windowThe context window is the maximum amount of text (measured in tokens) a language model can process at once — typically 128k to 1M tokens with current models.