An adversarial scan is a scheduled execution of probe templates against an LLM endpoint, an AI agent, or a RAG pipeline. Each probe is a deliberately crafted input designed to elicit a known failure mode: prompt injection, sensitive-information disclosure, jailbreak, tool overuse, overreliance, fairness regression, and so on. The response is scored and the result becomes a finding.
Adversarial scans differ from a one-shot pentest in two ways: cadence (daily or weekly rather than annual) and structure (the probe catalogue is versioned, mapped to control identifiers, and reproducible). Scheduled cadence is required to satisfy the post-market monitoring obligations under EU AI Act Article 72 and the continuous improvement loop in ISO/IEC 42001.
A finding from an adversarial scan ships with: the probe identifier, the verdict and the tier that decided it (a deterministic check, or the model judge when the checks could not settle the response), the framework control identifiers it maps to, and a remediation pointer.