OpenAI Guardrails: A Framework to Keep LLM Apps Safe and Reliable (openai.github.io)

OpenAI released Guardrails, a safety framework for LLM apps that automatically validates inputs and outputs with configurable, pipeline-based checks. Developers can create rules visually via a no-code Guardrails Wizard or write pipeline configs, then swap their OpenAI client for GuardrailsAsyncOpenAI (a drop-in replacement for AsyncOpenAI). Guardrails runs validations on every API call (input, pre-flight, output) and surfaces results on the response object (response.guardrail_results), making it easy to block, flag, or log problematic content while still returning model output (e.g., response.llm_response.output_text). Technically, Guardrails bundles built-in modules for content safety (moderation, jailbreak detection), data protection (PII detection, URL filtering), and content quality (hallucination and off-topic detection). It’s positioned as production-ready infrastructure for real-world LLM deployments and includes quickstart examples (e.g., client = GuardrailsAsyncOpenAI(config="guardrails_config.json"); response = await client.responses.create(model="gpt-5", input="Hello")). Important caveats: Guardrails may call third-party tools like Presidio and is not a substitute for developer-side safeguards. Teams remain responsible for storage, logging, and legal compliance around sensitive or illegal content and should avoid persisting blocked material.