Prompt Injection Protection
Prompt Injection Protection is a security testing framework designed to help teams evaluate the resilience of LLMs and AI applications against prompt-injection attacks. The platform provides structured and custom testing workflows for probing model behaviour under adversarial inputs, identifying weaknesses in system prompts, and assessing different mitigation strategies. Alongside its testing capabilities, the platform includes guidance for hardening system prompts and supporting material on common attack vectors, giving teams a central environment for testing and improving the security of AI systems.
In collaboration with AvanadeMy Contributions
- I contributed to the design and development of the full-stack application, working across the React/TypeScript frontend and Python/FastAPI backend.
- My work included implementing the user-facing testing workflows, application logic, and functionality required to run and explore prompt-injection tests and mitigation strategies.
Tech Stack
- React 18 & TypeScript - Frontend architecture and interactive testing workflows
- Vite - Development environment and production build tooling
- Tailwind CSS v4 - Responsive interface styling
- Radix UI / shadcn/ui - Accessible and reusable interface components
- Python & FastAPI - Backend APIs, testing logic and server-side processing
- CSS variables & theme system - Light and dark theme support
Features
- Standardised prompt-injection testing - Repeatable testing workflows for evaluating AI models against known prompt-injection techniques
- Custom adversarial testing - Tools for designing and running custom prompt-injection scenarios against different AI systems
- Model security evaluation - Workflows for assessing how models and AI applications respond to adversarial inputs
- System prompt hardening - Guidance and techniques for designing more resilient system prompts
- Attack and mitigation exploration - Interactive tools for analysing prompt-injection techniques and testing defensive approaches
- Full-stack testing environment - Backend-powered infrastructure supporting security tests and experimentation
- Responsive interface - Modern testing UI with light and dark mode support
Industry Collaboration
- The project was developed as part of UCL's IXN programme in collaboration with Avanade, working alongside Teo Montero Bonet, David Thomas, and Nathaniel Gibbon.
- The project was supported by Josh McDonald from Avanade, with Zekun Wu contributing guidance during the research process, and was supervised through UCL by Professor Dean Mohamedally and Yun Fu.
- The team also had the opportunity to present the project to the Microsoft Foundry Community in a webinar titled "Closing the Gap: A Unified Approach to Prompt Injection Testing", with the opportunity facilitated by Lee Stott.
Project Status
Used internally by Avanade - The framework is currently used internally by Avanade as a tool for exploring and testing prompt-injection security. Following its development, the project was also presented to the Microsoft Foundry Community as an approach to bringing prompt-injection testing into a more unified and practical framework.
Project Screens
Short placeholders for now. Replace these notes with a fuller explanation later.
Prompt Injection Catalogue
Overview screen for browsing vulnerability types, testing attacks and reviewing mitigation ideas inside a structured prompt-injection security workflow.
Garak Framework Scan
Scan results view showing blocked, handled and compromised prompts, with a risk breakdown for evaluating model behaviour under adversarial inputs.
