The Chrome Web Store and the extension stores of other browsers contain a growing number of AI-powered tools intended to support users with tasks such as summarising, writing, translating, or analysing web content. However, these tools often have broad access to any web content accessed by the user and share it with a model, while it may be unclear whether it is processed further, e.g., for training models or other purposes. In addition, AI-powered browser extensions may be vulnerable to prompt-injection attacks. A malicious website could include instructions in its content that influence an extension’s AI component and potentially cause unintended actions or data disclosure. Previous work has demonstrated such risks for website-summarisation tools [1]. As the number of AI-powered extensions continues to grow while security research in this area remains limited, this thesis aims to identify and assess vulnerabilities and privacy risks in AI-powered browser extensions.
The goals of this thesis may therefore include:
- Analysing a different/wider set of AI-based browser extensions for prompt injection vulnerabilities
- Examining whether privacy declarations and permission descriptions match the extensions’ observed data-handling behaviour
- Developing a test framework for organisations and consumers to verify certain security criteria
- Proposing or evaluating mitigations against prompt injection and unintended data disclosure in browser extensions
This thesis is particularly suitable for students interested in web-application security, user privacy, and the security of AI-powered applications. General knowledge of web programming and an understanding of large language model (LLM) applications are desirable. However, as the project focuses on the practical security of AI-based applications, detailed knowledge of LLMs is not necessarily required.
References: