# Guaranteeing AI Robustness against Deception
GARD was a DARPA program addressing the robustness of artificial-intelligence systems against deception and adversarial manipulation.
## Outputs identified by the article
The program’s public tooling lineage includes the Armory evaluation testbed, the Adversarial Robustness Toolbox, and APRICOT datasets. These support repeatable evaluation of models under adversarial conditions.
## Architectural role
GARD is a foundation of [[AI Immune Systems]]: the model is treated as a component whose failure modes must be deliberately elicited, measured, and mitigated.
## Boundary
Robustness against tested attacks does not establish general safety. Threat coverage, model version, deployment conditions, data drift, and untested attack classes remain material.
## Source
[[articles/Cognitive-Cyber Warfare Measures and Countermeasures#DARPA and IARPA Research Leads|Cognitive-Cyber Warfare — GARD]].