GUARD-AI

As agentic systems gain real-world autonomy — browsing the web, executing code, managing files — the attack surface expands dramatically. GUARD-AI addresses this with a model-agnostic framework designed to protect agentic systems from adversarial attacks, prompt injection, and unsafe tool use.

The architecture introduces layered defense checkpoints: an input sentinel that detects prompt-injection and jailbreak attempts, a tool-call auditor that validates every proposed action against a configurable policy before execution, and an output verifier that scans responses for data leakage and harmful content. Each layer is model-agnostic, meaning it can protect GPT, Claude, or open-source agents equally.

A technical paper was written to detail the GUARD-AI framework, its implementation, and evaluation results demonstrating that safety and capability are not a zero-sum trade-off.