International Journal of Scientific Research and Engineering Development

As large language models (LLMs) evolve from chatbots into integral parts of autonomous agents within enterprise ecosystems, attack surfaces grow exponentially, both in number and complexity. This paper provides an up-to-date survey of the landscape of LLM jailbreaking and prompt injection, with a specific focus on indirect and multimodal injection methods, which pose a significant threat to the safety alignment of LLMs. We investigate the practical efficacy of different LLM firewalls, which aim to offer runtime guardrails between the user querying the model and the model itself by performing input and output sanitization. We find that many existing approaches offer only a limited safety guarantee, due to the fact that most of them apply only one detection layer to classify potentially hazardous inputs. We propose a multi-layer sequential defense pipeline, which employs both light-weight pattern classification and semantic analysis based on an LLM, in which each layer is evaluated individually.

Shreyas G C, Mr. Raghu M T, Monisha R, Nishanth K L, Punith M, "LLM Jailbreak & Prompt Injection Firewall" International Journal of Scientific Research and Engineering Development, V9(4): Page(1069-1075) July-August 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.