Simulating Vulnerability
LLM-driven agents for policy-controlled audits under exposure constraints
Why this project
Auditing recommender systems in safety-critical domains is hard for a simple reason: you often need to express audit intent (seek, avoid, stop) while also respecting exposure constraints. Scripted bots scale, but they are typically open-loop and cannot condition behavior on what is actually on screen. This project introduces a framework for dynamic audits under exposure constraints, where intent is explicitly defined as an interaction policy and executed in a closed loop.