AI APPLICATION SECURITYPROPOSED STUDY
When an AI agent trusts the wrong instruction
A local lab for tracing prompt injection across retrieval, tool calls, and application permissions.
PLANNED LAB · NOT YET COMPLETEDThis is a proposed lab brief. Findings, evidence, and reproducible steps will be added after the work is completed.
The question
Where should an AI application stop trusting the information it receives?
The lab
A deliberately vulnerable local assistant, synthetic documents, and a restricted set of tools.
What the work will produce
- A map of inputs and trust boundaries
- Reproducible tests with expected and observed behavior
- Mitigations and a regression check
This is a proposed lab brief. Findings, evidence, and reproducible steps will appear after the work is completed.
END OF LAB BRIEFBACK TO RESEARCH ↗