When an AI agent trusts the wrong instruction
A local lab for tracing prompt injection across retrieval, tool calls, and application permissions.
OPEN THE LAB BRIEF RESEARCH / 01ALL FOUR PATHS ↗
BEHIND THE BLACK BOX
VULNERABILITY HUNTING.
Exploring how AI applications fail, where trust breaks, and what makes a finding worth fixing.
The small changes that fool a model
An experiment in model robustness, controlled perturbations, and the limits of accuracy.
OPEN THE LAB BRIEFProposed research directions. Findings and walkthroughs will be added as the work is completed.