To allow unrestricted execution of external commands.<\/span><\/li>\n<\/ol>\nCorrect Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Least privilege limits an AI agent’s access to the minimum resources and actions required to perform its authorized functions. This is particularly important for agents because they may be capable of interacting with databases, APIs, files, applications, or other external systems. If an agent is compromised, manipulated, or simply makes an incorrect decision, excessive permissions can increase the potential impact. Organizations should define permissions based on the agent’s intended purpose and periodically review them as capabilities or workflows change. Least privilege does not prevent every security incident, but it limits the agent’s potential reach. It should be combined with authentication, monitoring, logging, sandboxing, and appropriate human approval controls.<\/span><\/p>\nQuestion 382<\/b><\/p>\n
Why should tool calls made by an AI agent be logged?<\/b><\/p>\n\n- To increase the agent’s creativity.<\/span><\/li>\n
- To provide traceability of actions performed by the agent.<\/span><\/li>\n
- To eliminate the need for access controls.<\/span><\/li>\n
- To prevent every possible security incident.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Logging agent tool calls provides a record of actions taken by an AI system when interacting with external resources. Depending on the system, logs may capture the requested action, time, identity or session information, target resource, result, and relevant security events. These records can support monitoring, incident investigation, troubleshooting, auditing, and accountability. Logging does not prevent every incident, but it provides evidence that can help an organization understand what happened and determine whether controls operated as expected. Logs should themselves be appropriately protected because they may contain sensitive information. Organizations should define suitable retention, access, integrity, and review requirements based on the risk and purpose of the agent.<\/span><\/p>\nQuestion 383<\/b><\/p>\n
What is an important safeguard when an AI agent can perform high-impact actions?<\/b><\/p>\n\n- Require appropriate authorization or human approval before specified actions.<\/span><\/li>\n
- Give the agent unlimited permissions.<\/span><\/li>\n
- Disable all activity logging.<\/span><\/li>\n
- Allow the agent to modify its own security controls.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
High-impact actions can create significant consequences if performed incorrectly or without authorization. Examples may include changing financial records, modifying important infrastructure, approving transactions, deleting information, or making decisions that materially affect individuals. Organizations should identify which actions require additional safeguards and establish appropriate approval or authorization mechanisms. Depending on the risk, this may involve human confirmation, transaction limits, dual approval, or other controls. The objective is to ensure that the agent’s ability to act is proportional to the potential consequences. Unlimited permissions and self-modification of security controls increase risk, while disabling logs removes valuable evidence needed for oversight and investigation.<\/span><\/p>\nQuestion 384<\/b><\/p>\n
What is the main purpose of implementing transaction limits for an AI agent?<\/b><\/p>\n\n- To increase the number of actions the agent can perform simultaneously.<\/span><\/li>\n
- To limit potential financial or operational impact if the agent behaves incorrectly.<\/span><\/li>\n
- To eliminate authentication requirements.<\/span><\/li>\n
- To guarantee that all transactions are legitimate.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Transaction limits provide a preventive control that restricts the maximum financial or operational impact an AI agent can create within a defined period or workflow. For example, an organization may restrict the value, frequency, or number of transactions an agent can initiate without additional approval. These limits are particularly useful when agents have authority to perform actions automatically. They do not guarantee that every transaction is legitimate because an incorrect action can still occur within the permitted boundary. However, they can significantly reduce the potential impact of mistakes, compromised credentials, prompt manipulation, or unexpected agent behavior. Limits should be reviewed periodically to ensure they remain appropriate for the agent’s approved purpose and risk level.<\/span><\/p>\nQuestion 385<\/b><\/p>\n
Which control can help protect an AI agent from prompt injection leading to unauthorized tool use?<\/b><\/p>\n\n- Allowing all retrieved instructions to override system controls.<\/span><\/li>\n
- Separating untrusted content from trusted instructions and enforcing tool authorization.<\/span><\/li>\n
- Disabling authentication for external tools.<\/span><\/li>\n
- Giving retrieved documents unrestricted execution privileges.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Prompt injection can occur when untrusted content attempts to manipulate an AI system into ignoring its intended instructions or performing unauthorized actions. For an agent with tool access, the consequences can be significant because a manipulated instruction may result in data access, external communication, or other actions. A layered defense should distinguish trusted instructions from untrusted retrieved or user-provided content and enforce authorization independently of the model’s interpretation. Tool permissions, input validation, sandboxing, confirmation requirements, and monitoring can provide additional protection. The model should not be treated as the sole security boundary. Even if an agent is manipulated, independent authorization controls should prevent it from performing actions outside its approved permissions.<\/span><\/p>\nQuestion 386<\/b><\/p>\n
Why is retrieval access control important in a retrieval-augmented AI system?<\/b><\/p>\n\n- It ensures the system retrieves only information the requesting user or process is authorized to access.<\/span><\/li>\n
- It guarantees that generated answers are always factually correct.<\/span><\/li>\n
- It eliminates all prompt injection risks.<\/span><\/li>\n
- It makes every document publicly available.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Retrieval-augmented systems can connect AI models to internal documents, databases, knowledge bases, or other information sources. Without appropriate access controls, a system might retrieve confidential information that the requesting user should not be allowed to see. This can create serious confidentiality and privacy risks even if the underlying model itself is functioning correctly. Authorization should therefore be enforced at the retrieval layer and should reflect the permissions of the requesting user or process. Additional controls such as document classification, tenant isolation, logging, and output monitoring may also be appropriate. Access control does not guarantee factual accuracy, but it helps ensure that AI responses are generated from information that the requester is authorized to access.<\/span><\/p>\nQuestion 387<\/b><\/p>\n
What is a major risk of placing confidential organizational information into an unauthorized generative AI service?<\/b><\/p>\n\n- The information may be exposed, retained, or processed outside approved organizational controls.<\/span><\/li>\n
- The information automatically becomes more accurate.<\/span><\/li>\n
- The service will always delete the information immediately.<\/span><\/li>\n
- Confidentiality requirements no longer apply after submission.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Submitting confidential information to an unauthorized AI service can create risks involving disclosure, retention, secondary use, access by unauthorized parties, or loss of organizational control over the information. The organization may also be unable to determine how the provider stores or processes submitted content. Appropriate governance should therefore define which AI services are approved, what information may be entered, and what controls are required. Employees should receive practical guidance so they understand the risks of entering sensitive information into unapproved systems. Technical controls such as data-loss prevention or access restrictions may provide additional protection. Authorization, provider due diligence, data classification, and acceptable-use policies should work together to reduce this risk.<\/span><\/p>\nQuestion 388<\/b><\/p>\n
What is the purpose of tenant isolation in a multi-tenant AI application?<\/b><\/p>\n\n- To prevent one customer’s data or activity from being improperly accessible to another customer.<\/span><\/li>\n
- To allow all customers to share the same confidential information.<\/span><\/li>\n
- To eliminate the need for authentication.<\/span><\/li>\n
- To ensure every customer receives identical AI outputs.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Multi-tenant AI applications serve multiple customers or organizational groups through shared infrastructure or services. Tenant isolation helps ensure that one tenant’s data, requests, configurations, or results cannot improperly cross into another tenant’s environment. Weak isolation can create serious confidentiality and privacy problems, particularly when AI systems retrieve information or maintain conversation context. Appropriate controls may include tenant-aware authorization, logical or physical data separation, secure identifiers, access testing, and monitoring for cross-tenant leakage. Tenant isolation does not guarantee identical outputs and does not replace authentication. Instead, it is a fundamental security and privacy control for ensuring that shared AI infrastructure does not unintentionally expose information between customers.<\/span><\/p>\nQuestion 389<\/b><\/p>\n
Why should an organization test an AI system for unintended data leakage?<\/b><\/p>\n\n- Because AI systems may sometimes expose information through outputs, retrieval, logs, or integrations.<\/span><\/li>\n
- Because data leakage only occurs in traditional databases.<\/span><\/li>\n
- Because testing guarantees that leakage can never occur.<\/span><\/li>\n
- Because leakage testing eliminates the need for access controls.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
AI systems can create data leakage risks through generated outputs, retrieval mechanisms, logs, prompts, integrations, error messages, or other components. A system may unintentionally disclose confidential information when a user provides a particular prompt, exploits a weakness, accesses another user’s context, or causes the system to retrieve unauthorized material. Testing can help identify these weaknesses before or after deployment. Examples include checking for unauthorized retrieval, cross-user information exposure, sensitive information appearing in outputs, and inappropriate logging. Testing should complement strong access controls, data minimization, authorization, and monitoring. It cannot guarantee that every future leakage scenario will be prevented, but it can identify weaknesses and improve the overall security posture.<\/span><\/p>\nQuestion 390<\/b><\/p>\n
What is an appropriate response when an AI security test identifies a critical vulnerability before production deployment?<\/b><\/p>\n\n- Deploy immediately and address it later.<\/span><\/li>\n
- Document the vulnerability, assess its risk, and remediate or otherwise formally manage it before deployment.<\/span><\/li>\n
- Delete all security test results.<\/span><\/li>\n
- Disable security monitoring.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
A critical vulnerability identified before production should be treated as a significant release decision rather than ignored. The organization should document the finding, assess its potential impact, determine its cause, and implement appropriate remediation or risk treatment. Depending on the situation, remediation may involve changing code, restricting permissions, modifying architecture, adding security controls, or delaying deployment. Any residual risk should be explicitly evaluated and accepted by the appropriate authority when justified. Deleting test results removes important evidence and prevents effective follow-up. A disciplined security testing process allows organizations to identify weaknesses before they affect real users and provides a structured basis for determining whether the system is ready for deployment.<\/span><\/p>\nQuestion 391<\/b><\/p>\n
What is the purpose of secure model artifact management?<\/b><\/p>\n\n- To protect model files and related artifacts from unauthorized modification, replacement, or access.<\/span><\/li>\n
- To ensure all models are publicly downloadable.<\/span><\/li>\n
- To eliminate version control.<\/span><\/li>\n
- To allow anyone to replace production models.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Model artifacts, including model weights, configuration files, evaluation packages, and related components, can be valuable and sensitive assets. Unauthorized modification or replacement of these artifacts could cause security, performance, fairness, or operational problems. Secure artifact management can include access controls, integrity checks, versioning, controlled storage, provenance records, and approval requirements before deployment. These controls help organizations determine which artifact was approved and whether the artifact deployed to production matches the tested version. Public availability is not inherently required and may increase risk depending on the system. Strong artifact management also supports incident investigation and rollback because teams can identify and restore known-good versions when necessary.<\/span><\/p>\nQuestion 392<\/b><\/p>\n
Why should AI models and software components be scanned for known vulnerabilities?<\/b><\/p>\n\n- To identify known security weaknesses that may affect the system or its dependencies.<\/span><\/li>\n
- To guarantee that the AI system is free of every possible vulnerability.<\/span><\/li>\n
- To eliminate the need for patching.<\/span><\/li>\n
- To increase model creativity.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
AI applications often rely on software libraries, frameworks, containers, operating-system packages, model-serving components, and other dependencies. Known vulnerabilities in these components can provide attackers with opportunities to compromise the system or its surrounding infrastructure. Vulnerability scanning helps identify known weaknesses so organizations can prioritize patches, upgrades, compensating controls, or other remediation. Scanning is not a guarantee of complete security because unknown vulnerabilities and configuration weaknesses may remain. It should therefore be part of a broader vulnerability-management process that includes inventory, severity assessment, patch management, testing, monitoring, and incident response. Regular scanning is especially important when systems depend on frequently updated open-source or third-party components.<\/span><\/p>\nQuestion 393<\/b><\/p>\n
What is the main purpose of change impact assessment for an AI system?<\/b><\/p>\n\n- To determine how a proposed change may affect system behavior, risks, controls, and affected stakeholders.<\/span><\/li>\n
- To prevent all future changes.<\/span><\/li>\n
- To guarantee that the updated system will perform better.<\/span><\/li>\n
- To eliminate documentation requirements.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
A change to an AI system can affect more than technical performance. Changes to models, training data, prompts, infrastructure, integrations, intended use, or business processes may alter privacy, security, fairness, reliability, or operational risks. A change impact assessment helps an organization identify these potential effects before implementing the change. The assessment can determine whether additional testing, approvals, documentation, or monitoring are required. It does not guarantee that the new version will perform better, and it does not mean that every change must be rejected. Instead, it provides a structured mechanism for understanding consequences and ensuring that governance controls remain appropriate after the change.<\/span><\/p>\nQuestion 394<\/b><\/p>\n
Why is rollback capability important for AI deployments?<\/b><\/p>\n\n- It allows an organization to return to a known-good version when a new release creates unacceptable problems.<\/span><\/li>\n
- It guarantees that no future incident can occur.<\/span><\/li>\n
- It prevents all model updates.<\/span><\/li>\n
- It eliminates the need for testing.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Rollback capability provides an important operational safeguard when a new AI version produces unexpected or unacceptable behavior. If monitoring identifies severe performance degradation, security issues, harmful outputs, or other significant problems, the organization can return to a previously validated version while investigating the cause. A rollback process should be tested and should clearly identify which version is considered reliable. Organizations should also preserve relevant records so they can understand what changed between versions. Rollback does not replace testing or monitoring, but it reduces the time required to respond to serious problems. For high-impact systems, the ability to restore a known-good configuration can be an important part of operational resilience.<\/span><\/p>\nQuestion 395<\/b><\/p>\n
What is the purpose of a canary deployment for an AI system?<\/b><\/p>\n\n- To expose a new version to a limited population or traffic segment before wider deployment.<\/span><\/li>\n
- To deploy the new version to every user immediately.<\/span><\/li>\n
- To disable monitoring during release.<\/span><\/li>\n
- To prevent comparison between old and new versions.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
A canary deployment introduces a new AI version to a limited portion of users, traffic, or operational activity before expanding deployment more broadly. This allows organizations to observe real-world behavior and compare relevant performance, reliability, security, and other indicators against the existing version. If serious problems appear, the organization can stop the rollout or revert to the previous version before the impact becomes widespread. Canary deployment is especially useful when production behavior may differ from controlled testing. It should be supported by clearly defined monitoring thresholds, escalation procedures, and rollback capability. It does not eliminate risk, but it can reduce the scale and duration of problems associated with a new release.<\/span><\/p>\nQuestion 396<\/b><\/p>\n
What is the purpose of defining release approval criteria for an AI system?<\/b><\/p>\n\n- To establish measurable conditions that must be satisfied before a version is approved for deployment.<\/span><\/li>\n
- To allow every model version to enter production automatically.<\/span><\/li>\n
- To remove accountability from the deployment process.<\/span><\/li>\n
- To prevent all future model improvements.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Release approval criteria establish expectations that an AI system must meet before it moves into production. Criteria may address performance, security, privacy, fairness, robustness, documentation, testing results, unresolved findings, operational readiness, or other risk factors. Clear criteria help prevent deployment decisions from being based solely on informal judgment or schedule pressure. They also provide evidence that appropriate controls were considered before release. Criteria should be proportionate to the system’s risk and intended use, and exceptions should be documented and approved by the appropriate authority. Establishing release gates does not prevent improvements; rather, it ensures that improvements are evaluated before they become operational dependencies.<\/span><\/p>\nQuestion 397<\/b><\/p>\n
Why should AI evaluation datasets be version controlled?<\/b><\/p>\n\n- To track which data was used for a particular evaluation and understand changes over time.<\/span><\/li>\n
- To guarantee that every dataset is unbiased.<\/span><\/li>\n
- To prevent all dataset updates.<\/span><\/li>\n
- To eliminate the need for evaluation documentation.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Evaluation datasets can change because records are added, removed, corrected, relabeled, or otherwise modified. Without version control, an organization may not know exactly which dataset produced a particular performance result. This can make comparisons between model versions unreliable and complicate incident investigations. Versioning allows teams to identify the data used during testing and understand what changed between evaluations. It also supports reproducibility and auditability. Version control does not guarantee that the dataset is unbiased or representative. Those qualities still require assessment. However, maintaining identifiable dataset versions provides an important foundation for understanding evaluation results and determining whether changes in model performance may be related to changes in the evaluation data.<\/span><\/p>\nQuestion 398<\/b><\/p>\n
What is a key governance concern when an AI provider changes its underlying model without notifying customers?<\/b><\/p>\n\n- The change may alter system behavior, performance, risks, or compliance assumptions.<\/span><\/li>\n
- The change can never affect customer applications.<\/span><\/li>\n
- Customers automatically become responsible for all provider activities.<\/span><\/li>\n
- Model changes are always harmless.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Changes to an underlying provider model can affect outputs, performance, security behavior, latency, supported capabilities, or other characteristics that customers rely on. If the customer is not informed, it may be difficult to determine whether a change caused a new issue or whether existing testing remains valid. Organizations using third-party AI services should therefore consider contractual and operational controls for provider changes. These may include notification requirements, version transparency, testing opportunities, service-level expectations, and mechanisms for escalation. The appropriate controls depend on the risk and service arrangement. Provider changes do not automatically create an incident, but ungoverned changes can undermine assumptions made during the customer’s initial assessment.<\/span><\/p>\nQuestion 399<\/b><\/p>\n
What is the purpose of establishing service-level expectations for a third-party AI provider?<\/b><\/p>\n\n- To define measurable expectations for important aspects of the provider’s service and support.<\/span><\/li>\n
- To guarantee that the provider will never experience an outage.<\/span><\/li>\n
- To eliminate the need for vendor oversight.<\/span><\/li>\n
- To give the provider unlimited control over customer data.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Service-level expectations help an organization establish measurable requirements for important characteristics of a third-party AI service. Depending on the service and risk level, these may address availability, response times, incident notification, support, security responsibilities, or other operational commitments. Clearly defined expectations make it easier to determine whether the provider is meeting agreed requirements and what escalation procedures apply when service levels are not achieved. Service-level expectations cannot guarantee that outages or incidents will never occur. They also do not replace vendor due diligence or ongoing monitoring. Instead, they provide a contractual and operational framework that supports accountability and helps organizations manage dependencies on external AI providers.<\/span><\/p>\nQuestion 400<\/b><\/p>\n
What is the primary purpose of an AI governance control self-assessment?<\/b><\/p>\n\n