A Comparative Evaluation of AI Agent Security Guardrails

2026-04-27

AI Summary

This study systematically evaluates the effectiveness of mainstream AI safety guardrails in mitigating two critical risks: attacks targeting AI agents—such as prompt injection and tool misuse—and the induction of harmful content, including hate speech and violence. Leveraging a human-annotated unified benchmark, the work presents the first fine-grained, dual-dimensional empirical comparison among DKnownAI Guard, AWS Bedrock Guardrails, Azure Content Safety, and Lakera Guard. Results demonstrate that DKnownAI Guard significantly outperforms its counterparts, achieving a recall of 96.5% and a true negative rate of 90.4%, thereby exhibiting superior balance between high detection capability and low false positive rates. This evaluation fills a critical gap in the literature by providing a standardized, multi-guardrail assessment framework.

Abstract

This report presents a comparative evaluation of DKnownAI Guard in AI agent security scenarios, benchmarked against three competing products: AWS Bedrock Guardrails, Azure Content Safety, and Lakera Guard. Using human annotation as the ground truth, we assess each guardrail's ability to detect two categories of risks: threats to the agent itself (e.g., instruction override, indirect injection, tool abuse) and requests intended to elicit harmful content (e.g., hate speech, pornography, violence). Evaluation results demonstrate that DKnownAI Guard achieves the highest recall rate at 96.5\% and ranks first in true negative rate (TNR) at 90.4\%, delivering the best overall performance among all evaluated guardrails.

Problem

• AI agent security
• guardrails
• harmful content detection
• instruction override
• content safety

Innovation

• AI agent security
• guardrails
• comparative evaluation
• harmful content detection
• instruction override