User Access Provisioning and Security Configuration

Overview
User Access Provisioning and Security Configuration is a key activity in the Operationalize AI Solution phase of the PMI Cognitive Project Management for AI (CPMAI) methodology. Once a model has been trained, evaluated and approved for deployment, the project team must decide who can use the AI system, what they can do with it, and how the system, its data and its models are protected. A technically excellent model that is deployed with weak access controls or poor security can harm the organization, its customers and the credibility of the whole AI program.

Why It Is Important
1. Protecting sensitive data: AI systems often consume and produce personal, financial, health or proprietary data. Unauthorized access can lead to privacy breaches and regulatory penalties (for example under GDPR, HIPAA or CCPA).
2. Protecting the model as an asset: Trained models represent major investment and intellectual property. Attackers may try model theft or extraction, model inversion (reconstructing training data from outputs), membership inference, or tampering with model weights.
3. Preventing misuse and manipulation: AI systems are exposed to AI-specific threats such as adversarial inputs, prompt injection (for generative AI), data poisoning through feedback loops, and abuse of APIs.
4. Supporting trustworthy and responsible AI: Access control, auditability and accountability are core parts of trustworthy AI. Stakeholders need confidence that only appropriate people make decisions with, or changes to, the AI system.
5. Enabling compliance and governance: Auditors and regulators expect documented access policies, logging, segregation of duties and evidence of control over who changed what and when.
6. Ensuring operational continuity: Proper configuration reduces downtime, prevents accidental changes to production models, and supports incident response.

What It Is
User access provisioning is the process of creating, granting, modifying, reviewing and revoking access rights to the AI solution and its components. Security configuration is the set of technical and procedural controls that protect the AI solution in production. Together they cover:
- Identity management: establishing who users and systems are (human users, service accounts, applications calling model APIs).
- Authentication: verifying identity, for example single sign-on (SSO), multi-factor authentication (MFA), API keys or tokens, and certificates.
- Authorization: deciding what each identity may do, using role-based access control (RBAC), attribute-based access control (ABAC) and least privilege.
- Data security: encryption at rest and in transit, data masking, anonymization or pseudonymization, and data residency controls.
- Model and pipeline security: protecting model registries, training pipelines, feature stores and deployment environments, with version control and change approval.
- Monitoring and auditing: logging access and predictions, alerting on anomalies, and periodic access reviews.
- Lifecycle management: onboarding, role changes and offboarding, sometimes described as joiner, mover and leaver processes.

How It Works in Practice
Step 1: Identify stakeholders and user personas. Build on the business understanding work from CPMAI Phase I. Typical personas include end users who consume predictions, business analysts, data scientists, ML engineers, MLOps or IT operations staff, administrators, auditors and external systems or partners.

Step 2: Define roles and permissions. Map each persona to a role with only the permissions it needs (principle of least privilege). For example:
- End users can view predictions but cannot change models.
- Data scientists can experiment in development environments but cannot push directly to production.
- MLOps engineers can deploy approved model versions.
- Auditors have read-only access to logs and documentation.
Apply segregation of duties so that no single person can build, approve and deploy a model alone.

Step 3: Classify data and assets. Use the data understanding and data preparation work (CPMAI Phases II and III) to classify data sensitivity, such as public, internal, confidential or restricted. Sensitivity drives the strength of controls. Models trained on sensitive data inherit sensitivity concerns.

Step 4: Configure authentication and authorization. Integrate with enterprise identity providers. Enforce MFA for privileged roles. Secure model APIs with tokens, rate limiting and throttling, and network restrictions such as private endpoints, VPNs and firewalls.

Step 5: Apply security controls to the AI stack.
- Encrypt training data, feature stores and model artifacts.
- Secure the model registry and require approval workflows for promotion to production.
- Validate and sanitize inputs to reduce adversarial and prompt injection risk.
- Filter outputs to prevent leakage of sensitive information.
- Isolate environments (development, test, staging, production).
- Manage secrets with vaults rather than hard-coded credentials.

Step 6: Establish logging, monitoring and alerting. Record who accessed the system, what was queried, what the model returned (where lawful and appropriate) and any configuration changes. Feed logs into security monitoring and tie them to model monitoring for drift and anomalies.

Step 7: Test and validate. Conduct security testing before go-live: penetration testing, access control testing, red-teaming for AI-specific attacks, and verification that roles behave as designed.

Step 8: Document and hand over. Produce access policies, role matrices (for example a RACI or permission matrix), runbooks and incident response procedures. Train users and support teams.

Step 9: Operate, review and iterate. Perform periodic access recertification, revoke access promptly when people leave or change roles, rotate keys, patch vulnerabilities, and revisit configurations whenever the model is retrained or the scope changes. CPMAI is iterative, so security configuration is not a one-time event.

Connection to the CPMAI Methodology
CPMAI has six phases: Business Understanding, Data Understanding, Data Preparation, Model Development, Model Evaluation and Model Operationalization. Access and security needs should be identified early (Phase I requirements and Phase II data governance) but are fully implemented and validated in operationalization. CPMAI emphasizes trustworthy AI, data-centric thinking and iteration. Expect exam scenarios that link security to governance, privacy, transparency and ongoing monitoring rather than treating it as a purely IT task.

Common Pitfalls
- Granting broad admin access to everyone for convenience.
- Letting data scientists deploy directly to production without approval.
- Leaving default credentials or hard-coded API keys.
- Ignoring AI-specific threats and relying only on traditional IT controls.
- Failing to revoke access for departed team members or vendors.
- No logging, which leaves no audit trail for decisions made with AI.
- Treating security as an afterthought added after deployment.

Exam Tips: Answering Questions on User Access Provisioning and Security Configuration
1. Default to least privilege. When asked how to grant access, the best answer usually gives the minimum permissions needed for the role. Avoid answers that grant broad or administrative access to speed things up.
2. Prefer role-based, policy-driven approaches. Answers that define roles and assign users to roles beat ad hoc, individual permission grants.
3. Look for segregation of duties. If a scenario has one person developing, approving and deploying models, the correct response introduces separation and approval workflows.
4. Security is planned early, implemented in operationalization. If a question asks when to consider security, the answer is from the beginning, with requirements captured in Business and Data Understanding and controls verified before go-live. Do not choose answers that postpone security until after deployment.
5. Recognize AI-specific threats. Know model inversion, model extraction or theft, membership inference, data poisoning, adversarial examples and prompt injection. Correct answers often add AI-specific controls such as input validation, output filtering, rate limiting and red-teaming on top of standard IT security.
6. Data sensitivity drives controls. When sensitive or regulated data is involved, select answers involving encryption, masking or anonymization, data minimization and compliance review.
7. Audit trails matter. For questions on accountability, compliance or investigating incidents, choose options that include logging, monitoring and periodic access reviews.
8. Engage the right stakeholders. The project manager coordinates with information security, legal or compliance, data governance and IT operations rather than deciding security policy alone. Answers that collaborate with these groups are usually preferred.
9. Think lifecycle, not one-time. Correct answers often mention ongoing review, deprovisioning, key rotation and re-evaluation after retraining or scope changes.
10. Balance usability and security. The best answer protects the system without making it unusable, for example SSO with MFA rather than excessive manual approvals for read-only users.
11. Watch for distractors. Be wary of options that rely only on user training, only on perimeter firewalls, or that skip testing to meet deadlines. Also avoid options that share credentials or service accounts among many people.
12. When in doubt, pick the governance-aligned answer. CPMAI strongly emphasizes trustworthy, responsible AI. Options that improve transparency, accountability and compliance while enabling business value are typically correct.

Sample Question
An AI fraud detection model is ready for deployment. Data scientists request production admin rights so they can quickly fix issues. What should the project manager do?
A. Grant admin rights to all data scientists to reduce downtime.
B. Deny all data scientist access to production.
C. Define role-based access with least privilege, provide read-only monitoring access to data scientists, and route changes through an approved MLOps deployment workflow.
D. Share a single admin account among the team and log its use.
Answer: C. It applies least privilege, segregation of duties and controlled change management while still enabling the data scientists to support the model.

Key Takeaways
User access provisioning and security configuration ensure that the right people have the right access to the right AI components for the right reasons, and that data, models and pipelines are protected against both traditional and AI-specific threats. On the exam, favor least privilege, RBAC, segregation of duties, early planning, AI-aware controls, comprehensive logging, stakeholder collaboration and continuous review.