Adversarial testing and runtime monitoring that establishes where an AI system fails

Advai was founded in London in 2020 by David Sully and Chris Jefferson, and its method comes from adversarial machine learning rather than from compliance checklists. The premise is that an AI system's accuracy score says nothing useful about where it breaks, so the platform deliberately applies perturbations, distribution shifts, corrupted inputs, and adversarial examples to find the boundary conditions under which performance collapses. Those boundaries become numbers an organization can operate against: thresholds for acceptable input quality, decision confidence, and drift, defined before deployment rather than inferred after an incident.

The platform covers six dimensions in its assessments, spanning performance, security, safety, robustness, auditability, and stability, and applies the same tests comparatively when a team is choosing between model options so the selection rests on measured failure behavior. After deployment, connectors feed production data back in for continuous monitoring against the established thresholds, tracking drift, reliability, and operating cost, with configurable test packages and policies mapped to an organization's own risk framework. Advai works predominantly in defence, national security, and regulated finance, and its early research was funded through multiple UK Defence and Security Accelerator projects on adversarial attack and defence methods for computer vision and natural language systems.