The problem
Prompt injection is often tested as a one-off demonstration. That makes it difficult to compare applications, repeat an assessment after a model or retrieval change, or turn results into evidence that engineering and assurance teams can use.
What I built
I designed and built an end-to-end evaluation framework for retrieval-augmented and LLM applications. It combines a threat model, a structured prompt-injection payload taxonomy, and a scoring rubric for grading application and model responses.
Assurance mapping
Evaluation output maps to NIST AI RMF and ISO/IEC 42001 controls. The result is auditable assurance evidence rather than an isolated pass-or-fail result, connecting technical behavior to governance obligations.
Designed for repeatability
The framework can be rerun as models, prompts and retrieval configurations change, making injection resistance a measurable and regression-testable property. The repository is private and available on request.