Description:
The Model Context Protocol (MCP) enables AI development assistants to interact with repositories, documentation, and external services through a common interface. This thesis investigates how malicious instructions in tool descriptions, requests or responses can manipulate an assistant into performing unauthorized actions across connected MCP servers. The student will develop a reproducible test environment and compare prompt-based safeguards, static tool permissions, and task-aware authorization enforced outside the language model. Controlled experiments will evaluate their effectiveness in preventing attacks while preserving the assistant’s ability to complete legitimate software engineering tasks.
AI Website Generator