Basilisk: An Evolutionary AI Red-Teaming Framework for Systematic Security Evaluation of Large Language Models

Basilisk: An Evolutionary AI Red-Teaming Framework for Systematic Security Evaluation of LLMsThis research paper introduces Basilisk, an offensive security framework designed to automate the discovery of vulnerabilities in Large Language Models (LLMs). At the core of the framework is Smart Prompt Evolution for Natural Language (SPE-NL), a novel genetic algorithm that treats adversarial prompts as organisms under selection pressure.SPE-NL evolves payloads through sophisticated mutation operators (syntactic, semantic, and multilingual) and a multi-signal fitness function to systematically bypass production-grade guardrails. Empirical evaluation demonstrates that Basilisk achieves a 92% improvement in Attack Success Rate (ASR) compared to static adversarial payload libraries. This work aims to provide security researchers and AI safety engineers with a standardized methodology for continuous, automated red-teaming of evolving AI systems.