XGuardLM by Zakariae FanziXGuardLM by Zakariae Fanzi

XGuardLM

Portfolio project
Zakariae Fanzi
XGuardLM — A Security Boundary for LLM Applications
A defensive layer designed to reduce prompt-injection and sensitive-data risks around language-model applications.
The problem
An application can expose an LLM to untrusted instructions, confidential values or malicious prompt content. Basic keyword filters are not enough for every scenario.
What I built
XGuardLM explores a controlled request and response workflow that:
• Inspects incoming prompts • Detects potentially hostile instructions • Pseudonymizes sensitive information • Sends controlled content to the model • Validates the generated output • Restores protected values only after the response passes validation
The objective is to create multiple security checkpoints instead of relying on a single prompt or filter.
Technology
Python, LLM Guard, language models and policy-learning experiments.
XGuardLM is a security-focused prototype and should be combined with application-specific authorization, logging and testing before production use.

0 likes · 0 views · Nov 20, 2026

A defensive layer designed to reduce prompt-injection and sensitive-data risks around language-model applications.

Tags