1 | Scope of the document | 0 |
2 | Normative references of the standard | 0 |
4 | Abbreviated terms and acronyms | 0 |
5.1 | Security controls applied to AI systems | 0 |
5.2 | Establishing an AI governance programme | 0 |
5.3 | Zero trust principles for AI systems | 0 |
5.4 | Software engineering practices for AI | 0 |
5.5.1 | Purpose of supply chain management | 0 |
5.5.2 | AI supply chain threat assessment | 0 |
5.5.3 | Complexity of AI supply chains | 0 |
5.5.4 | Supply chain attack vectors | 0 |
5.5.5 | AI supply chain security measures | 0 |
5.6 | Threat modelling for AI systems | 0 |
5.7 | Red teaming of AI systems | 0 |
6.1 | Types of AI system compromise | 0 |
6.10.1 | Output injection attacks description | 0 |
6.10.2 | Failures from output injection attacks | 0 |
6.10.3 | Detect output injection attacks | 0 |
6.10.4 | Mitigate output injection attacks | 0 |
6.11.1 | General recommended practices pointer | 0 |
6.11.2.1 | Limiting data to reduce impact | 0 |
6.11.2.2 | Data minimization | 0 |
6.11.2.3 | Minimal retention | 0 |
6.11.2.4 | Data encryption and obfuscation | 0 |
6.11.2.5 | Limit exposure of AI technical details | 0 |
6.11.3.1 | Determine cause of unwanted model behaviour | 0 |
6.11.3.2 | Human oversight and guardrails | 0 |
6.11.3.3 | Least privilege for AI agents | 0 |
6.11.3.4 | Tell stakeholders how the system works | 0 |
6.11.3.5 | Continuous validation against test data | 0 |
6.11.3.6 | Explain individual system decisions | 0 |
6.11.3.7 | Test for unwanted bias | 0 |
6.11.4 | Provenance for AI systems and data | 0 |
6.2.1 | How data poisoning works | 0 |
6.2.2 | Failures caused by data poisoning | 0 |
6.2.3 | Detect poisoned training data | 0 |
6.2.4 | Mitigate data poisoning attacks | 0 |
6.3.1 | Evasion attack description | 0 |
6.3.2 | Failures from evasion attacks | 0 |
6.3.3 | Detect evasion attacks | 0 |
6.3.4 | Mitigate evasion attacks | 0 |
6.4.1 | Membership inference description | 0 |
6.4.2 | Failures from membership inference | 0 |
6.4.3 | Detect membership inference | 0 |
6.4.4 | Mitigate membership inference | 0 |
6.5.1 | Model exfiltration description | 0 |
6.5.2 | Failures from model exfiltration | 0 |
6.5.3 | Detect model exfiltration | 0 |
6.5.4 | Mitigate model exfiltration | 0 |
6.6.1 | Model inversion description | 0 |
6.6.2 | Failures from model inversion | 0 |
6.6.3 | Detect model inversion | 0 |
6.6.4 | Mitigate model inversion | 0 |
6.7.1 | Direct model poisoning description | 0 |
6.7.2 | Failures from direct model poisoning | 0 |
6.7.3 | Detect direct model poisoning | 0 |
6.7.4 | Mitigate direct model poisoning | 0 |
6.8.1 | Sensitive model output description | 0 |
6.8.2 | Failures from sensitive model output | 0 |
6.8.3 | Detect sensitive model output | 0 |
6.8.4 | Mitigate sensitive model output | 0 |
6.9.1 | Prompt injection description | 0 |
6.9.2 | Failures from prompt injection | 0 |
6.9.3 | Detect prompt injection | 0 |
6.9.4 | Mitigate prompt injection | 0 |
7.1 | Assess mitigations as a system | 0 |
7.2 | Check for conflicting mitigation effects | 0 |
7.3 | Keep mitigations consistent across the life cycle | 0 |
7.4 | Monitor for mitigation deterioration | 0 |
7.5 | Log and monitor model use | 0 |
7.6 | Protect the development environment | 0 |
7.7.1 | Detect out of distribution inputs | 0 |
7.7.2 | Detect adversarial inputs | 0 |
7.8 | Throttle model use per user | 0 |
7.9 | Limit model size and memorisation | 0 |
Annex A | Pair threat mappings with conventional controls | 0 |
Annex B | Protect training data in development and operation | 0 |
Annex C | Federated learning risks and protections | 0 |