Basilisk: An Evolutionary AI Red-Teaming Framework for Systematic Security Evaluation of Large Language Models
Abstract
This paper presents Basilisk, an open-source evolutionary framework for automated security evaluation of large language models (LLMs). Unlike static prompt-injection databases, Basilisk applies genetic algorithms to natural language attack payloads — evolving adversarial prompts across generations using 15 mutation operators, 5 crossover strategies, and multi-signal fitness scoring. The Smart Prompt Evolution engine (SPE-NL) discovers novel jailbreaks, guardrail bypasses, and data exfiltration vectors that no human-curated list would contain.
Basilisk provides 32 attack modules mapped to the OWASP LLM Top 10, including multi-turn attack chains (prompt cultivation, authority escalation, sycophancy exploitation), differential multi-model testing, guardrail posture assessment, and forensic audit logging. The framework supports GPT-4, Claude, Gemini, Llama, and any custom LLM endpoint, with native C/Go extensions for 10-100x speedup on performance-critical operations.
Archived On
Key Contributions
Smart Prompt Evolution
First application of genetic algorithms to natural language attack payloads for LLM security testing, with 15 mutation operators and population diversity tracking.
Multi-Turn Attack Chains
Novel multi-phase conversational attacks including prompt cultivation (5-phase guardrail conditioning) and sycophancy exploitation leveraging LLM agreement bias.
Systematic OWASP Coverage
32 attack modules providing complete OWASP LLM Top 10 coverage with automated differential testing and guardrail posture grading.
Native Performance
C and Go extensions providing 10-100x speedup for token analysis, pattern matching, and mutation operations via a ctypes bridge with Python fallbacks.
Citation
@misc{regaan2026basilisk,
author = {Regaan},
title = {Basilisk: An Evolutionary AI Red-Teaming
Framework for Systematic Security Evaluation
of Large Language Models},
year = {2026},
version = {1.0.8},
publisher = {ROT Independent Security Research Lab},
doi = {10.5281/zenodo.18909538},
url = {https://doi.org/10.5281/zenodo.18909538}
}