Pebblo is an AI security platform designed to secure customer-facing and internal AI chatbot interactions that use Retrieval-Augmented Generation (RAG) architectures. It operates through a dual-layer architecture that addresses threats at both the model interaction layer and the data retrieval layer. Safe Infer (Model Layer): Provides policy-based model routing by user or group to approved models, real-time sensitive data redaction before prompts reach the model, and validation of model completions for appropriateness and compliance prior to display. Safe RAG (Retrieval Layer): Secures data connectors with fine-grained permissions, enforces authorization-aware retrieval against vector databases, and defends against data-borne prompt/command injection attacks by stripping malicious instructions at the retrieval stage before content reaches the model. Injection Defense: Specifically targets command-injection attacks hidden within retrieved data (as demonstrated by attacks like EchoLeak), neutralizing embedded malicious instructions before they influence model behavior. Unified Policy and Observability: A single policy plane governs both model and retrieval layers, providing centralized policy definition, full audit trails, alerting, and dashboards for security and operations teams. Compliance Controls: Built-in enforcement for HIPAA, GDPR, CCPA, and data residency requirements, with per-group model policies. Data ingestion classifies content dynamically based on confidentiality, compliance, and semantic context.