PI-XM Vulnerability Assessment

Enable safer multimodal AI by systematically measuring when and how image embedded prompts (and other steganographic payloads) cause downstream multimodal models to produce harmful or disallowed content — and by producing clear mitigations and reproducible tests that model builders, researchers, and instructors can run locally.

Build a reproducible report that:

• uses Hugging Face models to synthesize images containing encoded/embedded malicious prompts,
• feeds those images into alternate multimodal models to see whether the malicious instruction is executed,
• automates scoring and triage with LLM a judge and evaluation tools (JailBreak Bench),
• produces well documented code, experiments, trials, and an instructor friendly demo.

Runs and experiments will be executed on the university GPU server cec.cap.gpu1.csi.miamioh.edu under approved access and with explicit permission from the university.