Othmane Belmoukadam, Jiri De Jonghe, Sofyan Ajridi, Amir Krifa, Joelle Van Damme, Maher Mkadem1 and Patrice Latinne, AI LAB, Belgium

AdversLLM is a comprehensive framework designed to help organizations tackle security threats associated with the use of Large Language Models (LLMs), such as prompt injections and data poisoning. As LLMs become integral to various industries, the framework aims to bolster organizational readiness and resilience by assessing governance, maturity, and risk mitigation strategies. AdversLLM includes an assessment form for reviewing practices, maturity levels, and auditing mitigation strategies, supplemented with real-world scenarios to demonstrate effective AI governance. Additionally, it features a prompt injection testing ground with a benchmark dataset to evaluate LLMs' robustness against malicious prompts. The framework also addresses ethical concerns by proposing a zero-shot learning defense mechanism and a RAG-based LLM safety tutor to educate on security risks and protection methods. AdversLLM provides a targeted, practical approach for organizations to ensure responsible AI adoption and strengthen their defenses against emerging LLM-related security challenges.

Large Language Models, Natural Language Processing, Prompt Injections, Responsible AI, AI guardrails