AI Accuracy, Robustness and Cybersecurity Assurance

Define and test accuracy, robustness and cybersecurity measures for high-risk AI, including AI-specific attack resilience.

For a high-risk AI provider, define and document appropriate performance, resilience and security criteria for the intended use. Assess likely errors, failures and attacks, including AI-specific threats where relevant. Test suitable safeguards and recovery measures, and document the limits of the evaluation. Keep instructions aligned with validated performance and revisit weaknesses after changes or incidents. Where a presumption of conformity is relied on, verify its exact scope and conditions; it does not establish compliance with unrelated obligations.

Before release; throughout the lifecycle as appropriate; after relevant model, infrastructure or threat changes.

Performance metrics and validation reports Fault, resilience and security test plans and results Threat assessments and safeguard decisions Remediation and retest records Published performance information and any relied-on conformity evidence

Trace a significant performance or security risk to a safeguard and test result. Check that failures were investigated and that the deployer-facing information matches the validated system. Include one changed component in the sample.