A Comparative Evaluation of AI Agent Security Guardrails
Abstract
DKnownAI Guard demonstrates superior performance in AI agent security by achieving the highest recall rate and true negative rate compared to AWS Bedrock Guardrails, Azure Content Safety, and Lakera Guard.
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.
Community
This benchmark is directly aligned with a gap we keep running into: guardrails for agents need to be evaluated at the runtime/action boundary, not only as generic content filters. Instruction override, indirect injection, unsafe tool use, and exfiltration all look different once they become tool arguments or planned actions.
We have been building Armorer Guard as a local Rust runtime scanner for that boundary. It returns JSON verdicts, scores, and reasons for prompt injection, sensitive-data requests, exfiltration-style text, destructive-command risk, safety bypass, and system-prompt extraction. The intent is to combine fast semantic scoring with deterministic policy and least-privilege tool design, rather than treating any single guardrail as the whole defense.
Demo: https://huggingface.co/spaces/armorer-labs/armorer-guard-demo
Repo: https://github.com/ArmorerLabs/Armorer-Guard
One evaluation split I would love to see more often is pre-context scanning vs pre-tool-call gating vs post-action audit. Those stages have very different latency and false-positive costs.
Get this paper in your agent:
hf papers read 2604.24826 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 1
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper