# Engineering Overview

## Building an In-Process Generative Adversarial Guardrail

When deploying autonomous LLM agents at scale, protecting customer data from leakage and endpoints from unintended operation is paramount to the success of your business.
Next generation attackers and evolving threat landscape leave enterprise models exposed to insider threats and public facing exploits. Protecting your data in the age of 
AI is an evolving challenge with new attacks being published constantly.

**Enterprise Shield** aims to solve emerging threats through a generative adversarial training method that uses the Arcanum Taxonomy model developed by the security 
research pioneered by Jason Haddix. Trained on a robust combination of diverse datasets, including real-world user-generated attacks across over 100 languages, the engine 
achieves deep cross-lingual threat generalization. By developing models using a near exhaustive mapping of attacks, the system delivers next generation defense without the 
performance penalty of traditional architectures.

## The Arcanum Taxonomy: Redefining Threat Detection

Standard industry solutions rely on static blocklists and rigid token classifiers that degrade when confronted with novel, syntactically mutated adversarial framing. 
Grounded in the comprehensive Arcanum threat taxonomy and backed by extensive multi-lingual real-world attack data, Enterprise Shield introduces a multi-tiered security 
model designed around adversarial simulation and real-time semantic deconstruction.

### Dynamic Adversarial Training

Instead of matching known signatures, the Arcanum model maps inputs against an internal latent manifold trained via generative adversarial simulations and real 
user-generated exploit streams spanning the prompt injection attack surface. Leveraging Haddix's structured threat classifications, the system proactively identifies structural intent 
anomalies, neutralizing prompt injections and obfuscated payloads before they touch underlying LLM weights.


## Try It Locally

Payloads can be tested directly in the live sandbox, or packages can be downloaded to run local benchmarks against application logs:

* **Python:** `pip install icephi-python`
* **TypeScript:** `npm i @ice_phi/icephi-ts`
* **API Docs & Sandbox:** [icephi.com/api/agentic-guardrail](https://www.google.com/search?q=https://icephi.com/api/agentic-guardrail)