AI Security · Testing
TrustBoundary Lab
TrustBoundary Lab is a defensive security evaluation workbench for exploring how an LLM-enabled application handles hostile prompts, untrusted retrieved documents, sensitive canaries, and unauthorized tool calls.
Features
- Versioned JSON scenarios: direct injection, indirect injection, canary leakage, RAG poisoning simulation and unauthorized tool use
- Deterministic mock provider for repeatable runs, with an optional Gemini provider
- Assertions: exact-string exclusion, regex checks and heuristic checks reported as inconclusive
- Per-scenario JSON evidence and an aggregate summary
- FastAPI service and React dashboard, with pytest and CI on every push
Technologies
PythonPydanticFastAPIReactTypeScriptVitepytestGemini APIGitHub Actions