AI guardrails
Guardrails constrain and check AI application behaviour. Model instructions and technical enforcement perform different tasks.
Guardrails are rules and controls intended to constrain AI application behaviour. They can check inputs, outputs or proposed actions. They may include model instructions, technical checks and appropriate approvals.
A message example
A hypothetical assistant creates an email draft. Sending remains subject to separately checked permission and the intended approval. The model considering the text harmless does not open the sending path.
An output filter can miss disallowed content and block permitted content. A prompt instruction is no fixed access boundary either. Controls must fit the task and be tested with permitted and prohibited cases. The guardrails label alone is no security demonstration.
Sources
Related terms
- Human-in-the-loopHuman-in-the-loop actively involves people in an automated process. Useful review requires understandable information and ways to intervene.
- Prompt injectionPrompt injection attempts to redirect an AI application through inserted instructions. Retrieved documents can contain them too.
- Least privilegeLeast privilege limits permissions to the required task and data scope. Necessary work remains possible while unnecessary powers are removed.
- Model cardA model card documents an AI model’s purpose, foundations, evaluation conditions and known limitations for a particular version.
- AI readinessAI readiness concerns the foundations of an AI project: objectives, usable data, skills, responsibilities and an operating workflow.
- Context engineeringContext engineering prepares information for individual model calls: relevant instructions, messages, sources and current states.
Articles on this topic
AI & Automation with ASES
AI assistants and automated workflows