What problem does it solve? LLM-powered chat and agent applications can leak data across tenants, invoke tools without proper authorization, or render model output unsafely, and these risks are hard to verify without a structured, evidence-based testing method. ## Core Features & Use Cases - Trust Boundary Modeling: Maps user input, system prompts, retrieval sources, tool declarations, authorization identity, approval points, and output rendering to separate what the model says from what the application actually executes. - Minimal Safe Validation: Uses self-owned test documents, test tenants, and harmless canaries to verify cross-tenant isolation, tool-call authorization and confirmation, source isolation, and safe output rendering. - Oracle and False-Positive Rules: Defines strict validity criteria so that model claims, prompt-injection text echoes, or simulated tool returns are not mistaken for real findings. - Use Case: A security researcher testing an AI assistant with RAG ingestion plants a canary in one tenant's documents and checks whether another tenant's session ever retrieves it, capturing conversation hashes and tool audit logs as evidence. ## Quick Start Ask the agent to run the hunt-llm-ai route against my authorized LLM chat application to verify tenant isolation and tool-call authorization using test canaries.
