How to Prevent AI Agents from Leaking Sensitive Data with Fiddler Guardrails and LiteLLM

Learn how to secure your AI applications and coding agents from leaking sensitive data like PII, PHI, and credentials. In this video, we demonstrate how to set up inline guardrails using Fiddler Guardrails and the LiteLLM gateway to redact sensitive data in real-time before it leaves your corporate boundary. 👉 Explore How to Prevent AI Agents from Leaking Sensitive Data with Fiddler Guardrails and LiteLLM: https://www.fiddler.ai/blog/prevent-ai-agent-data-leakage-guardrails-?utm_source=youtube&utm_medium=organic_social&utm_campaign=202609-prevent-ai-agent-leakage AI agents require data access to perform tasks, but without proper inline guardrails, sensitive information can easily leak to external LLM providers (like Anthropic) and downstream observability tools. Watch as we walk through a hands-on demo using Claude Code, LiteLLM, and Fiddler to detect, redact, and monitor data leakage. What You'll Learn: – Guardrails vs. Evaluations: Understand why inline enforcement is critical compared to post-hoc evaluations. – Real-Time Redaction: See how Fiddler Centor Models redact PII/secrets in milliseconds. – LiteLLM Gateway & Fiddler SDK Integration: Learn how to proxy requests between your AI agent and model providers. – Preventing Data Leakage: Ensure sensitive data doesn't propagate to downstream logging and monitoring systems. – Guardrail Monitoring Dashboards: Use dashboards to track policy enforcement on an agent, including blocked inputs and outputs, redaction metrics, and tool calls. Drill down into individual spans for deeper analysis. Timestamps: 0:00 Guardrails vs. Evals: The Difference Explained 1:01 Our Demo Setup (Claude Code, LiteLLM, & Fiddler Guardrails) 1:43 Demo: Retrieving Customer Records Without Guardrails 2:21 Demo: Real-Time PII Redaction Using Centor Models 2:58 Guardrail Monitoring Dashboard 3:43 Understanding Traces