LLM Guard is a configurable security toolkit for identifying, editing, and cleansing prompts and model responses. It provides real-time protection for production LLM applications, helping enforce application-specific safety, security, and regulatory requirements. section 10.4
Protection is organized around input and output scanners. Input scanners can block unwanted substrings, competitor names, selected topics, toxic language, and prompt injection attempts. Output scanners check factual consistency, malicious URLs, and sensitive entities such as people and email addresses, with optional redaction. section 10.4
Financial-services chatbot
A customer-onboarding chatbot can restrict conversations to product information, risk-tolerance assessment, and account setup. A safeguard function scans the prompt, records validity and scores, blocks failed inputs with a fixed response, and otherwise continues toward model inference and output scanning. Example tests compare a prompt mentioning Fidelity, an ordinary question about dogs, toxic language, and a request for personal data with and without safeguards. section 10.4
