AI agents are increasingly being deployed in various businesses and organisations. Protocols such as the Model Context Protocol (MCP) facilitate the connection of AI systems to tools and data sources, often using existing credentials, access tokens, or delegated permissions. As with human employees, these agents should only have access to systems, data, and other assets when strictly necessary, and it must be transparent whether access was performed by an agent or directly by a human. However, the way such agents are integrated may make it difficult for an administrator or auditor to distinguish between actions performed directly by humans and those performed by AI agents. Moreover, employees might be tempted to authorise an agent without carefully considering the potential for misbehaviour.
This thesis project is therefore aimed at analysing shortcomings in current IAM practices with agentic AI. Possible outcomes of the thesis include, e.g.:
- a systematic review or empirical study of authorisation and delegation practices for AI agents
- an analysis of audit-log requirements for distinguishing actions performed by employees or agents
- an observational or interview-based study of how organisations use AI agents and which security risks arise in practice
- the design and evaluation of a concept, prototype, or potential MCP extension for more secure delegated authorisation and auditable agent actions
The project is suitable for students interested in cybersecurity, identity and access management, AI security, distributed systems, or human factors in security.
Recommended read: