1.1 | AI security content in the cyber security training programme | 0 |
1.2 | Ongoing staff awareness of AI threats, with secure AI coding training for developers | 0 |
10.1 | Tell end-users where and how their data is used, accessed and stored | 0 |
10.2 | Accessible guidance for end-users, including limitations and security updates | 0 |
10.3 | Support end-users and affected entities through incidents | 0 |
11.1 | Provide security updates and patches, with contingency where updates are impossible | 0 |
11.2 | Treat major updates as a new version and support operators through model changes (11.2 and 11.3) | 0 |
12.1 | Log system and user actions | 0 |
12.2 | Analyse logs for anomalies, breaches and unexpected behaviour | 0 |
12.3 | Monitor internal states of the AI system where useful for security | 0 |
12.4 | Monitor model and system performance over time for security-relevant change | 0 |
13.1 | Secure disposal when training data or a model is transferred | 0 |
13.2 | Secure deletion when a model or system is decommissioned | 0 |
2.1 | Assess the need, the risks and the mitigations before building an AI system | 0 |
2.2 | Design for resilience against adversarial attack, unexpected input and failure | 0 |
2.3 | Create an audit trail for the AI system | 0 |
2.4 | AI-specific risk assessment and due diligence for external components | 0 |
2.5 | Match intended use to data sensitivity, and encourage staff to report AI security risks | 0 |
2.6 | Least-privilege, risk-assessed permissions for the AI system on other systems | 0 |
2.7 | Due diligence on external providers | 0 |
3.1 | Threat modelling and risk management for AI systems | 0 |
3.2 | Pass unresolved AI threats down the chain to operators and end-users | 0 |
3.3 | Assurance over external parties responsible for AI risks | 0 |
3.4 | Continuous monitoring and review of system infrastructure against risk appetite | 0 |
4.1 | Build in and maintain capabilities for human oversight | 0 |
4.2 | Make outputs easy for responsible humans to assess | 0 |
4.3 | Technical measures where human oversight is a risk control | 0 |
4.4 | Verify the data custodian's specified controls are implemented | 0 |
4.5 | Tell end-users which uses of the system are prohibited | 0 |
5.1 | Comprehensive asset inventory including interdependencies | 0 |
5.2 | Processes and tools to track, authenticate, version and secure AI assets | 0 |
5.3 | Disaster recovery plans tailored to attacks on AI systems | 0 |
5.4 | Protect sensitive data and sanitise inputs, with proportionate protection of weights | 0 |
6.1 | Evaluate access control frameworks for APIs, models, data and pipelines | 0 |
6.2 | Controls on externally offered APIs against attacks on the AI system | 0 |
6.3 | Dedicated, segregated environments for development and model tuning | 0 |
6.4 | Publish a clear vulnerability disclosure policy | 0 |
6.5 | AI system incident management plan and recovery plan | 0 |
6.6 | Cloud service contracts that support these requirements | 0 |
7.1 | Secure software supply chain processes for AI development | 0 |
7.2 | Justify, mitigate and disclose use of poorly documented or unsecured components | 0 |
7.3 | Re-run evaluations on released models before use | 0 |
7.4 | Tell end-users before models are updated | 0 |
8.1 | Document system design and maintenance plans, and publish component hashes | 0 |
8.2 | Record the provenance and use of public training data | 0 |
8.3 | Audit log of changes to system prompts and model configuration | 0 |
9.1 | Security assessment testing before release of models, applications and systems | 0 |
9.2 | Operator testing before deployment, using independent skilled testers | 0 |
9.3 | Share testing and evaluation findings with system operators | 0 |
9.4 | Evaluate outputs for leakage of model internals and unintended influence | 0 |
G | The Implementation Guide: worked scenarios and example measures | 0 |
R | Stakeholder roles | 0 |
S | Scope and status of the Code of Practice | 0 |