← All themesOpen in explorer →
Accuracy, Robustness & Security across ISO 42001, NIST AI RMF and the EU AI Act
// theme · accuracy-robustness
Accuracy, Robustness & Security
Performance, resilience to errors and adversarial inputs, cybersecurity.
// Do once → satisfies all three
ONE pre-deployment test report (accuracy, robustness, adversarial, cybersecurity) plus a live drift-monitoring dashboard.
Art.15 demands declared levels; ISO/NIST demand the evidence underneath. One report + one dashboard maintains both.
ISO 42001
Annex A.6.2.4 · Cl.9.1
NIST AI RMF
MEASURE 2.5 · MEASURE 2.6 · MEASURE 2.7
EU AI Act
Art.15
// Evidence auditors expect
- ✓ Test report covering accuracy, robustness, cybersecurity
- ✓ Adversarial / red-team test results
- ✓ Drift-monitoring dashboard with thresholds
- ✓ Pen-test / vulnerability scan against model serving stack
// Common pitfalls
- ⚠ Test set drawn from training distribution only - no out-of-distribution evaluation.
- ⚠ Accuracy reported aggregate, hiding worst-case performance on subgroups.
- ⚠ No cybersecurity testing of the model serving / prompt-injection surface.
ISO 42001
3Annex A.6.2.4 covers verification and validation; Cl.9.1 covers ongoing measurement.
Clause 9.1
Monitoring, measurement, analysis and evaluation
Define what to monitor, methods, frequency, and evaluate AI performance.
Annex A.4.5
System and computing resources
Document the system and computing resources (infrastructure, compute) supporting AI systems.
Annex A.6.2.4
AI system verification and validation
Define and apply measures to verify and validate AI systems against requirements.
NIST AI RMF
6MEASURE 2.5–2.7 cover validity, reliability, safety, security and resilience.
MAP 2.3
Scientific integrity and TEVV
Scientific integrity and Test, Evaluation, Verification & Validation considerations documented.
MEASURE 1.1
Metrics and methods selected
Approaches and metrics for measuring AI risks are selected.
MEASURE 2.1
Test sets, metrics, details documented
Test sets, metrics and methodology details documented for evaluation.
MEASURE 2.5
Validity & reliability assessed
AI system is regularly evaluated for validity, reliability and performance.
MEASURE 2.6
Safety risks evaluated
AI system is evaluated for safety risks.
MEASURE 2.7
Security & resilience evaluated
Security and resilience including adversarial robustness evaluated.
EU AI Act
2Art.15 requires high-risk AI to achieve appropriate accuracy, robustness and cybersecurity, declared in technical documentation.