ABCDEFGHIJKLNOPQRSTUVWXYZAAABACADAE
1
AI SECURITY AUDIT CHECKLIST
2
Name of Organisation: CryptoVault Nigeria Limited
3
Address : 24 Bishop Aboyade Cole Street, Victoria Island, Lagos, Nigeria
4
RC No: 148552
5
Sector:Cryptocurrency Exchange & Digital Asset Management
6
AUDIT PARAMETERS
7
Audit AreaDomainsReferences to standards and frameworksResponse
8
9
Domain 1: AI Governance and Accountability
10
Organizational StructureIs there a formally approved AI Governance Policy that defines the organization's risk appetite, ethical principles, and accountability structure for AI systems?ISO/IEC 42001:2023 (Clause 5.2, 5.3 - Policy and Roles); NIST AI RMF (Govern function).
No formal policy yet. Draft in progress, expected approval by end of April 2026.
11
Roles & ResponsibilitiesAre roles, responsibilities, and authorities for AI system ownership, risk management, and compliance clearly defined, documented, and communicated across the organization?ISO/IEC 27002:2022 (5.2 - Roles and responsibilities); SOC 2 (Control Environment).
Partially defined informally. No formal RACI matrix. Teams know their roles but not documented.
12
Ethical PrinciplesHas the organization established and documented a set of ethical principles for AI use, and are these principles integrated into the AI system design and review process?NIST AI RMF (Govern function, specifically addressing societal and ethical risks); ISO/IEC 42001:2023 (6.1.2 - AI system impact assessment).No documented ethical principles. Not yet considered in design/review process.
13
14
Domain 2: AI Risk Management and Oversight
15
Risk AssessmentIs a formal, documented AI-specific risk assessment performed for all new and significantly modified AI systems, covering technical, ethical, legal, and operational risks?ISO/IEC 23894:2023 (AI risk management principles); NIST AI RMF (Map and Measure functions).
No formal AI risk assessments. Risks discussed in product meetings but not documented.
16
Risk TreatmentAre identified AI risks treated with appropriate controls, and is the effectiveness of these controls periodically reviewed and documented?ISO/IEC 27001:2022 (6.1.3 - Statement of Applicability); NIST SP 800-37 (Risk Management Framework).
No formal risk treatment. Some ad-hoc controls exist but effectiveness not reviewed.
17
Impact AssessmentDoes the organization conduct an AI System Impact Assessment (AIA) to evaluate potential negative impacts on individuals, groups, and fundamental rights before deployment?NDP Act (Section 28 - Data Privacy Impact Assessment, extended for AI); ISO/IEC 42001:2023 (6.1.2 - AI system impact assessment).
Not conducted. Team unaware of NDP Act requirements for AI impact assessments.
18
19
Domain 3: Model Lifecycle Security and Change Management
20
Secure MDLCAre security requirements (e.g., threat modeling, secure coding practices) integrated into each phase of the Model Development Lifecycle (MDLC), from design to deployment?NIST SP 800-218 (Secure Software Development Framework); ISO/IEC 27002:2022 (8.28 - Secure coding).
Not formally integrated. Security team consulted ad-hoc for major releases only.
21
Model VersioningIs a robust model versioning and artifact management system in place to ensure traceability, reproducibility, and integrity of all model components (code, data, configuration)?ISO/IEC 42001:2023 (8.3.3 - Traceability of AI systems); SOC 2 (System Operations).
Yes, we use MLflow for versioning. But integrity controls like checksums not implemented.
22
Change ManagementIs there a formal change management process for model updates, retraining, and deployment that includes security review, testing, and rollback capabilities?ISO/IEC 27002:2022 (8.32 - Change management); NIST SP 800-53 (CM-3 - Configuration Change Control).
Informal process. Changes approved by lead data scientist. Security review not mandatory. Rollback possible manually.
23
24
Domain 4: Training, Validation, and Inference Data Security
25
Data ProvenanceIs the provenance (source, collection method, licensing) of all training and validation data documented, and are controls in place to prevent the use of unauthorized or poisoned data?ISO/IEC 42001:2023 (8.3.2 - Data for AI systems); OWASP Top 10 for LLMs (TDP - Training Data Poisoning).

Partial documentation. KYC data sourced internally OK. Some external sentiment datasets have unclear licensing.
26
Data PrivacyAre appropriate privacy-enhancing technologies (PETs) or anonymization techniques applied to sensitive data before it is used for model training, and is access strictly controlled?NDP Act (Principle of Data Minimisation and Pseudonymisation); ISO/IEC 27002:2022 (8.10 - Data masking).
Anonymization applied inconsistently. KYC images stored raw. Access controls strong but data not masked.
27
Inference DataAre input and output data at the inference stage validated for integrity and sanitized to prevent injection attacks or data leakage?OWASP Top 10 for LLMs (PI - Prompt Injection; IO - Insecure Output Handling); NIST SP 800-53 (SI-10 - Information Input Validation).Basic input validation implemented. No specific sanitization for prompt injection risks.
28
29
Domain 5: Model Integrity, Robustness, and Adversarial Resistance
30
Adversarial TestingAre AI systems subjected to dedicated adversarial robustness testing (e.g., evasion, poisoning, model inversion attacks) before deployment, and are mitigation strategies documented?OWASP Top 10 for LLMs (MA - Model Abuse; MI - Model Denial of Service); NIST AI RMF (Measure function - Robustness).
Never performed. Team unaware of adversarial testing techniques.
31
Model ExplainabilityAre model explainability (XAI) techniques implemented and validated to ensure that model decisions can be understood, debugged, and audited for bias or security flaws?ISO/IEC 42001:2023 (8.3.4 - Transparency and explainability); NIST AI RMF (Measure function - Interpretability).
Not implemented. Models treated as black boxes. Unable to explain specific decisions.
32
Bias and FairnessAre systematic bias and fairness assessments conducted on the model and its training data, and are documented efforts made to mitigate unfair outcomes?NIST AI RMF (Measure function - Fairness); Regulatory Compliance (Emerging AI-specific regulations).
No formal assessments. Assumed models are fair without validation.
33
34
Domain 6: Access Control, Identity, and Privilege Management
35
Least PrivilegeIs the principle of least privilege strictly enforced for all human and service accounts accessing AI training environments, model repositories, and production endpoints?ISO/IEC 27002:2022 (5.15 - Access control); NIST SP 800-53 (AC-6 - Least Privilege).
Mostly yes. RBAC implemented in AWS. Quarterly access reviews conducted. Some developer accounts have excess permissions.
36
Service Account SecurityAre service accounts and API keys used by AI pipelines (e.g., for data access, model deployment) managed securely, including rotation, vaulting, and non-use of default credentials?ISO/IEC 27002:2022 (5.18 - Access rights review); SOC 2 (Security - Logical Access Controls).

Yes. AWS Secrets Manager used. Keys rotated every 90 days. Default credentials disabled.
37
AuthenticationIs multi-factor authentication (MFA) enforced for all administrative and privileged access to AI development and production environments?ISO/IEC 27002:2022 (5.17 - Authentication information); NIST SP 800-63B (Digital Identity Guidelines).
Yes. MFA mandatory for all AWS access. Enforced via IdP.
38
39
Domain 7: Infrastructure, Cloud, and MLOps Security
40
Network SegmentationAre AI development, training, and production environments logically and physically separated (e.g., via VPCs, subnets, firewalls) from the corporate network and each other?ISO/IEC 27002:2022 (7.2 - Network security); NIST SP 800-53 (SC-7 - Boundary Protection).
Yes. Dev, training, prod in separate AWS VPCs. Strict firewall rules. No direct corporate network access.
41
Configuration HardeningAre all compute resources (e.g., VMs, containers, Kubernetes clusters) used for AI workloads hardened according to established security baselines (e.g., CIS benchmarks)?ISO/IEC 27002:2022 (8.9 - Configuration management); NIST SP 800-70 (Security Configuration Checklists).

Partially. EC2 instances follow CIS. EKS clusters not fully hardened. Vulnerability scans run monthly.
42
MLOps Pipeline SecurityIs the MLOps CI/CD pipeline secured against tampering, ensuring that only authorized, scanned, and tested code/models can be deployed to production?OWASP Top 10 for LLMs (SCV - Supply Chain Vulnerabilities); NIST SP 800-218 (Secure Software Development Framework).
Basic controls in place. SAST/DAST implemented. Model scanning not yet integrated. Artifacts not signed.
43
44
Domain 8: Monitoring, Logging, and Incident Response
45
Model MonitoringAre AI models continuously monitored in production for performance drift, data drift, concept drift, and security anomalies (e.g., sudden changes in input patterns)?ISO/IEC 42001:2023 (8.3.5 - Monitoring and measuring of AI system performance); NIST AI RMF (Maintain function).
Yes. Transaction Monitoring AI monitored for drift. Alerts configured. KYC and Sentiment models not monitored.
46
Security LoggingAre comprehensive security logs (including input/output data, access attempts, and administrative actions) generated, protected from tampering, and retained according to policy?ISO/IEC 27002:2022 (8.15 - Logging); SOC 2 (Security - Monitoring Activities).
Yes. All logs sent to SIEM. Log retention 12 months. Integrity controls not implemented.
47
Incident ResponseDoes the incident response plan include specific playbooks and procedures for handling AI-specific incidents, such as model poisoning, adversarial attacks, or bias-related failures?ISO/IEC 27002:2022 (5.26 - Incident management); NIST SP 800-61 (Computer Security Incident Handling Guide).
No. General IR plan exists. No AI-specific playbooks. Never tested AI incident scenarios.
48
49
Domain 9: Third Party, Open Source, and Supply Chain Risk
50
Third-Party ModelsIs a due diligence process performed on all third-party or pre-trained models (e.g., LLMs, foundation models) to assess their security, provenance, and compliance with organizational standards?ISO/IEC 27002:2022 (5.21 - Managing information security in the supply chain); NIST AI RMF (Govern function - Supply Chain).
No. Sentiment analysis LLM imported from Hugging Face. No security review performed.
51
Open Source SecurityAre automated tools used to scan open-source libraries and dependencies for known vulnerabilities (CVEs) and license compliance before they are integrated into the AI system code base?OWASP Top 10 for LLMs (SCV - Supply Chain Vulnerabilities); NIST SP 800-161 (Supply Chain Risk Management).
Not yet. Tool identified but not implemented. Current process relies on manual review.
52
Data Provider RiskAre contractual and technical controls in place to ensure that external data providers maintain the security and integrity of the data they supply for AI training?ISO/IEC 27002:2022 (5.23 - Information security for use of cloud services); NDP Act (Data Processor requirements).

No formal agreements. External sentiment data purchased via online transaction. No security clauses.
53
54
Domain 10: Regulatory Compliance, Ethics, and Responsible AI
55
Regulatory MappingIs there a current inventory of all applicable AI-specific regulations (e.g., EU AI Act, state-level laws) and a documented mapping of these requirements to internal controls?NDP Act (Lawfulness, fairness, and transparency); Emerging AI-specific regulations (e.g., EU AI Act requirements for high-risk systems).



Yes. SEC, NDP Act, NFIU mapped. EU AI Act not applicable. Controls partially mapped.
56
Responsible AIAre processes in place to continuously monitor AI systems for unintended consequences, societal impacts, and deviations from responsible AI principles post-deployment?NIST AI RMF (Govern and Map functions - Societal Impact); ISO/IEC 42001:2023 (6.1.2 - AI system impact assessment).
No formal processes. No responsible AI principles defined. Ad-hoc reviews only.
57
Audit TrailIs a complete and immutable audit trail maintained for all significant decisions and actions related to the AI system (e.g., model selection, parameter tuning, data filtering)?ISO/IEC 42001:2023 (8.3.3 - Traceability of AI systems); SOC 2 (Control Activities)

Partial. MLflow tracks experiments. Audit trail exists but not immutable. Some decisions undocumented.
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100