REGAAN RGenetic Prompt Mutations on LLMs
Applying genetic algorithms to automate safety evaluation vectors against boundaries in Large Language Models.
🔍 The Limits of Static Boundaries
Most LLM safety guardrails use simple filters. By mutating adversarial prompts using genetic operators (like homoglyph swaps and token smuggling), we can automate the search for safety bypasses.
This write-up covers the design of the SPE-NL (Smart Prompt Evolution for Natural Language) system. I explain how scoring model outputs via semantic similarity embeddings allows us to dynamically optimize prompts, discovering jailbreak vulnerabilities across cloud-hosted models.