Masters Thesis: An LLM Agent for Expert-Level Privacy Audits

AI Sweden

21h • Hybrid

About The Position

As Sweden's national center for applied AI, AI Sweden is looking for a master's thesis student to join their team. The project focuses on developing an LLM-based planning layer around the LeakPro privacy auditing framework. The goal is to investigate whether LeakPro can be driven by a controlled auditing agent without modifying its existing attack implementations. This involves building a Model Context Protocol (MCP) server to expose LeakPro's attack catalog, creating an audit-scenario benchmark with expert-authored reference solutions, and developing an auditor-agent to evaluate its performance against expert attack selection, configuration, and compute cost. The work is dual-use, meaning the agent audits only models the operator controls, and candidate attack code must be sandboxed and reviewed by a human before registration.

Requirements

• Ongoing Master’s studies in Computer Science, Data Science, Engineering Physics, Complex Adaptive Systems, Machine Learning, or a related field.
• Comfortable with Python and deep learning.
• Comfortable with the reality that an experiment might yield unexpected results.
• Curious, independent, and self-driven.

Responsibilities

• Build a Model Context Protocol (MCP) server that derives tool definitions from LeakPro's attack schemas.
• Expose the attack catalog without modifying individual attack implementations.
• Assemble an audit-scenario benchmark with 10-15 scenarios, including target models, handlers, compute budgets, and expert-authored reference solutions.
• Fix scoring metrics, measure expert run-to-run variance, and release the suite as a benchmark.
• Develop an agent that receives a target model, handler, budget, and goal, and scores it on the suite for attack selection, configuration, and outcome.
• Log GPU-hours, LLM latency, and token cost per run.
• Plot risk estimate against compute for the agent and LeakPro's existing Optuna search.
• Explore an AutoMIA-style propose-implement-evaluate loop if time permits and the existing catalog underperforms.

What This Job Offers

• Career Level: Entry Level