I conducted an authorized AI security assessment of Lily Cybersecurity 7B to evaluate how effectively a hardened system prompt could resist jailbreak and prompt injection attacks.
Using Prompt Fuzzer, I ran 15 attack techniques against the model. The system prompt successfully blocked 8 attempts, including most roleplay and social engineering attacks. The successful bypasses mainly used translated or altered wording to disguise the intent of the request.
This project demonstrates why system prompts should be supported by input validation, moderation, output filtering, and continuous AI red team testing.
I conducted an authorized AI security assessment of Lily Cybersecurity 7B to evaluate how effectively a hardened system prompt could resist jailbreak and pro...