ISO 42001 A.11.6: AI Robustness and Security
AI systems face unique security threats that traditional controls don't address. A.11.6 requires organizations to build resilience against adversarial attacks, data poisoning, and model inversion attacks. This control ensures your AI infrastructure withstands both known and emerging threats throughout its lifecycle.
What this means
A.11.6 mandates implementing security measures specifically designed to protect AI systems from attacks that exploit model vulnerabilities. This includes hardening your data pipelines against poisoning, securing model APIs from unauthorized access, and protecting inference infrastructure from adversarial inputs. You must conduct robustness testing during development and maintain periodic security assessments during operations to identify and remediate emerging threats.
How to comply
- 1.Map your AI system's attack surface, including data ingestion points, model training pipelines, APIs, and inference endpoints
- 2.Implement data validation and sanitization controls to prevent poisoning attacks during training and inference
- 3.Conduct adversarial robustness testing before deployment and establish a recurring testing schedule (quarterly or semi-annually)
- 4.Deploy API authentication and rate limiting to prevent unauthorized model access and inference manipulation
- 5.Establish monitoring and alerting for anomalous model behavior, prediction patterns, and performance degradation
- 6.Document all robustness tests, findings, and remediation actions for audit trails
- 7.Review and update threat models quarterly as new AI-specific attack vectors emerge
Evidence auditors look for
- Adversarial robustness test reports showing FGSM, PGD, or other attack method results
- Data pipeline validation logs demonstrating input sanitization and anomaly detection
- Model API access logs with authentication records and rate-limiting configuration
- Model performance monitoring dashboards tracking inference accuracy and drift
- Security assessment reports specifically addressing AI attack surface
- Incident response logs documenting any attempted or successful model attacks
- Testing schedules and sign-offs from development and security teams
Frequently asked questions
When will FAQs be available?
The FAQ for this control is currently being prepared.
GRCWatch automates robustness testing workflow tracking, stores adversarial test evidence, maintains a continuous audit trail of security assessments, and flags when periodic robustness testing is due—eliminating manual spreadsheet management for A.11.6 compliance.
See how GRCWatch handles this control automatically
Start free trial