Katharsis Secure AI Prompting System
Faculty Mentor
Sanmeet Kaur
Presentation Type
Poster
Start Date
4-14-2026 9:00 AM
End Date
4-14-2026 11:00 AM
Location
PUB NCR
Primary Discipline of Presentation
Computer Science
Abstract
Large Language Models (LLMs) have grown in popularity in both personal and enterprise use over the last five years. Many businesses have put an immense focus on integrating AI technology into internal workflows. These systems can be coerced and manipulated just like humans. When AI is given access to critical business infrastructure, massive security vulnerabilities can appear. Specific attack vectors include DANs (Do anything now), Smuggling, Jailbreak and Prompt Injection. Katharsis is a Secure AI Prompting system developed under the mentorship of Pattern Agentic AI, a Spokane local AI startup. The system sits between an upstream prompt source and a downstream LLM. Katharsis is designed to work with any upstream prompt source, equipped to process input from humans or a separate agentic system. The system is equipped with rule based and AI driven flagging abilities. Rule based abilities include word blocking, regex pattern matching and levenshtein distance calculation. For AI defense, Katharsis uses a fine tuned Qwen 3 model to act as a “Security Advisor”, it analyzes prompts and categorizes threat severity. The software demonstrated an ability to block many known attack methods. Along with this, AI security was able to detect and block coercive and nuanced attacks. The aim of this project was not simply to develop and deploy a useful piece of technology, but to research and improve the topic of AI security.
Recommended Citation
Tucker, Andrew and McMahon, Brayden, "Katharsis Secure AI Prompting System" (2026). 2026 Symposium. 29.
https://dc.ewu.edu/srcw_2026/ps_2026/p1_2026/29
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Katharsis Secure AI Prompting System
PUB NCR
Large Language Models (LLMs) have grown in popularity in both personal and enterprise use over the last five years. Many businesses have put an immense focus on integrating AI technology into internal workflows. These systems can be coerced and manipulated just like humans. When AI is given access to critical business infrastructure, massive security vulnerabilities can appear. Specific attack vectors include DANs (Do anything now), Smuggling, Jailbreak and Prompt Injection. Katharsis is a Secure AI Prompting system developed under the mentorship of Pattern Agentic AI, a Spokane local AI startup. The system sits between an upstream prompt source and a downstream LLM. Katharsis is designed to work with any upstream prompt source, equipped to process input from humans or a separate agentic system. The system is equipped with rule based and AI driven flagging abilities. Rule based abilities include word blocking, regex pattern matching and levenshtein distance calculation. For AI defense, Katharsis uses a fine tuned Qwen 3 model to act as a “Security Advisor”, it analyzes prompts and categorizes threat severity. The software demonstrated an ability to block many known attack methods. Along with this, AI security was able to detect and block coercive and nuanced attacks. The aim of this project was not simply to develop and deploy a useful piece of technology, but to research and improve the topic of AI security.
