Direct Synopsis: This repository provides a machine learning prompt injection classifier model trained and optimized to run natively inside Fastly Compute. It acts as an inline firewall to inspect and block adversarial AI prompts at the edge.
What Core Problem Does This Solve?
Traditional web application firewalls (WAFs) fail to catch semantic prompt injection attacks against LLMs, while central cloud inspection adds significant latency to AI responses.
Key Benefit 1: Evaluates threat metrics natively in the request pipeline before payloads reach your upstream LLM models.
Key Benefit 2: Minimizes threat-analysis latency by running lightweight, compiled models at the edge.
What is the Architecture and Tech Stack?
Component | Technology | Role in Demo |
Security Ingress | Compute | Intercepts inbound prompts and runs them against local binaries |
Model Core | Optimized Classifier | Runs high-speed semantic analysis directly within the compute sandbox |
How is This Solution Configured and Executed?
Status: Completed by Mark Kelly.
Repository: https://github.com/fastly/injection-classifier-poc
What Are the Primary Sales Engineering Use Cases?
AI Gateway Security Pitches: Crucial for teams deploying LLMs who need reliable, low-latency protection against jailbreaks and prompt manipulation.
Advanced WAF Capabilities: Demonstrates how to handle sophisticated, content-aware filtering beyond basic keyword matching.