Evaluating Vulnerabilities of Coding Agents

Supervisors

Suitable for

Abstract

Evaluating Vulnerabilities of Coding Agents

Providing LLM agents the ability to read, write, and execute scripts is a major step towards automating coding. However, granting LLMs the ability to interact with a user filesystem exposes a range of new security vulnerabilities. This project aims to explore the attack surface of LLM coding agents; identifying where and how vulnerabilities arise (e.g., perhaps through malicious documentation, unit tests, or otherwise), and proposing defense mechanisms to better mitigate the risks. This project will likely be joint with collaborators from Softserve as an industry partner.