LLM Security and Safety: Insights from Homotopy-Inspired Prompt Obfuscation

TL;DR

This paper introduces a homotopy-inspired prompt obfuscation framework to study security and safety vulnerabilities in Large Language Models, connecting topology-inspired prompt deformation with jailbreak strategies in code-generation contexts. It details a five-stage methodology for eliciting, transforming, generating, verifying, and reporting malware-like artifacts across diverse LLMs, and provides a 7,374-specimen malware dataset for cybersecurity research, with 9,725 samples used for analysis. The study reports model-dependent jailbreak performance, with an overall precision around 0.758 and notable variability across LLaMA, DeepSeek, and KIMI when verified by Claude or KIMI, highlighting gaps in current safeguards. The work offers actionable defensive insights—adversarial stress-testing, multi-stage safety pipelines, and reinforcement of alignment—to strengthen safety, reliability, and trustworthiness of AI systems in real-world cybersecurity contexts.

Abstract

In this study, we propose a homotopy-inspired prompt obfuscation framework to enhance understanding of security and safety vulnerabilities in Large Language Models (LLMs). By systematically applying carefully engineered prompts, we demonstrate how latent model behaviors can be influenced in unexpected ways. Our experiments encompassed 15,732 prompts, including 10,000 high-priority cases, across LLama, Deepseek, KIMI for code generation, and Claude to verify. The results reveal critical insights into current LLM safeguards, highlighting the need for more robust defense mechanisms, reliable detection strategies, and improved resilience. Importantly, this work provides a principled framework for analyzing and mitigating potential weaknesses, with the goal of advancing safe, responsible, and trustworthy AI technologies.

Figures (12)

• Figure 1: Topological deformation of a doughnut into a coffee cup.
• Figure 2: Homotopy in LLMs, Malware and the definition have the same semantics. This is, how the circle fits into the cylinder in the unit interval $[0,1]$.
• Figure 3: KIMI trying to produce a metaphorical version of a malicious prompt.
• Figure 4: Copilot metaphorical version of a malicious prompt, the new version is a benign prompt.
• Figure 5: KIMI performing a Homotopy deformation of a malicious prompt.