Free CompTIA CY0-001 Practice Questions 2026 - Page 3

Explanation :

Storage retrieval and prompt processing are the primary cost drivers in generative AI systems, making them the best criteria for cost monitoring. Generative AI costs are typically calculated based on token consumption—the number of input tokens (prompts) and output tokens (completions) processed by the model. Prompt processing encompasses the computational expense of encoding user inputs, while storage retrieval applies to systems using Retrieval-Augmented Generation (RAG), where vector databases are queried to fetch relevant context before generating responses. Each retrieval operation consumes compute resources and incurs storage costs. Together, these metrics directly correlate with the volume of API calls, the length of prompts, and the complexity of retrieval queries—making them the most accurate financial indicators.

Why the other options are incorrect:

A. Connection access and exchange gateway:
This refers to network connectivity and API gateway traffic, which are minor operational costs compared to model inference. Gateways handle routing and authentication, not the core generative processing that drives billing.

B. Encryption and decryption processing:
Encryption/decryption is a security control for data in transit and at rest. While it consumes minimal CPU cycles, it is not a significant cost factor in generative AI billing. Most providers do not charge for encryption separately.

D. Catalog servicing and exchange processing:
This terminology suggests data catalog management or API exchange services, which are not relevant to generative AI cost structures. It does not align with any standard AI billing model.

References:

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 3.1 (Securing AI Systems – Cost Monitoring and Resource Optimization) and Domain 1.3 (AI Lifecycle – RAG and Token-Based Pricing).

Cloud AI Provider Billing Models: AWS Bedrock, Azure OpenAI, and Google Vertex AI all bill based on input/output token counts and vector database retrievals.

NIST AI RMF: Recommends monitoring resource utilization, including storage and compute, for cost efficiency.

Explanation:

Pattern recognition is the best technique to mitigate a DoS attack that exploits the context window by hiding thousands of obfuscated characters behind an emoji. This attack—often called a token smuggling or compressed prompt attack—uses Unicode tricks, homoglyphs, or layered encoding to bypass simple length restrictions. Pattern recognition algorithms (e.g., regex-based heuristics, entropy analysis, or ML-based anomaly detection) can identify abnormal patterns such as excessive encoded characters, unusual Unicode sequences, or statistical outliers in token distribution. By analyzing the structure and composition of the input rather than just its raw length, pattern recognition detects obfuscation attempts before they reach the model's context window, allowing the system to block, truncate, or sanitize the malicious payload.

Why the other options are incorrect:

A. Fraud detection: Fraud detection systems monitor for anomalous transactions or user behavior in financial contexts. They are not designed to analyze input text structure or detect obfuscated content embedded in prompts.

B. Large language model (LLM)-as-a-judge:
This technique uses another LLM to evaluate the quality, safety, or relevance of a primary model's outputs. It is used for content moderation and output validation, not for detecting obfuscated input patterns at the ingestion stage. Using an LLM to pre-filter inputs would be computationally expensive and latency-heavy.

D. Prompt filter:
Prompt filters typically block specific keywords, harmful phrases, or known jailbreak patterns. However, the attacker in this scenario is using obfuscation (hiding characters behind an emoji) to bypass keyword-based filters. A simple prompt filter would not recognize the encoded content as malicious because the keywords themselves are not present in plaintext.

References:

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 3.2 (Implementing AI Security Controls – Input Validation and Anomaly Detection) and Domain 2.2 (Threats to AI Systems – Denial-of-Service via Context Window Overflow).

OWASP Top 10 for LLMs (LLM04 – Unbounded Consumption): Highlights DoS attacks through excessive token usage and obfuscated inputs.

Explanation:

Configuring least privilege controls for model access directly addresses the root cause of this incident—an unauthorized user or process successfully modified the system prompt. The system prompt is a privileged configuration that defines the AI's core behavioral instructions. Its modification indicates that the attacker had excessive permissions—likely administrative or developer-level access—to the AI system's configuration interface or API. By implementing least privilege access controls, the administrator restricts system prompt modification to only authorized personnel (e.g., AI architects, security administrators) and prevents regular users, automated processes, or compromised accounts from altering core system instructions. This aligns with the principle of separation of duties and privileged access management (PAM), ensuring that configuration changes require proper authorization and audit trails.

Why the other options are incorrect:

A. Limit the number of evaluations that a user can send to the model:
This is a rate-limiting or quota control that prevents excessive API usage. It does not address system prompt modification because the attacker only needed one successful modification to cause harm—not multiple requests.

B. Check for model hallucination and recommend fine-tuning:
Hallucination refers to the model generating factually incorrect information. In this scenario, the model correctly followed the maliciously modified system prompt—it was not hallucinating. Fine-tuning would not prevent unauthorized prompt changes.

D. Encrypt all data going to and coming from the model:
Encryption protects data confidentiality and integrity during transmission (TLS) and storage (at-rest encryption). However, the system prompt was modified at the application configuration level—not intercepted during transmission. Encryption would not have prevented the unauthorized modification.

References:

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 3.1 (Securing AI Systems – Access Control and Least Privilege) and Domain 3.2 (Implementing AI Security Controls – Privileged Access Management).

NIST SP 800-53 (AC-6): Least Privilege—mandates restricting access rights to the minimum necessary.

Explanation:

Rate limiting is the most effective control to prevent a denial-of-service (DoS) attack via brute force on an AI model API. A brute force DoS attack floods the API with an overwhelming number of requests, exhausting compute resources, degrading performance, and rendering the model unavailable to legitimate users. Rate limiting restricts the number of API calls a single user, IP address, or API key can make within a specified timeframe (e.g., 100 requests per minute). By enforcing these thresholds, the security team ensures that even if an attacker sends massive volumes of requests, the API gateway or load balancer will reject excess traffic before it reaches the AI model. This preserves availability for legitimate hospital users and prevents cost spikes associated with excessive inference usage.

Why the other options are incorrect:

A. Tokenization:
Tokenization is a data protection technique that replaces sensitive data (e.g., PII, PHI) with non-sensitive placeholders (tokens). It protects data privacy and reduces compliance scope but does not mitigate DoS attacks or control request volume.

B. Model guardrails:
Guardrails enforce safety and policy rules on model inputs and outputs—such as blocking harmful content, detecting PII, or preventing prompt injection. They do not limit the number of requests, making them ineffective against volumetric DoS attacks.

D. Prompt firewall:
A prompt firewall filters or sanitizes individual prompts to block malicious inputs, jailbreaks, or injection attempts. It operates at the content level, not at the traffic volume level, and would not prevent thousands of legitimate-looking requests from overwhelming the system.

References (CY0-001):

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 3.1 (Securing AI Systems – API Security and Rate Limiting) and Domain 3.2 (Implementing AI Security Controls – Denial-of-Service Prevention).

OWASP API Security Top 10 (API4:2023 – Unrestricted Resource Consumption): Explicitly identifies rate limiting as the primary defense against resource exhaustion attacks.

Explanation:

OWASP is the most appropriate resource for setting up guardrails for an LLM application, specifically the OWASP Top 10 for Large Language Models. This framework provides comprehensive, industry-recognized guidance on the unique security risks facing LLM applications—including prompt injection, insecure output handling, model denial of service, sensitive information disclosure, supply chain vulnerabilities, and excessive agency. OWASP offers actionable recommendations for implementing guardrails, such as input validation, output sanitization, content filtering, and privilege enforcement. By leveraging OWASP's guidelines, organizations can systematically address known threats and establish robust guardrails that satisfy compliance requirements while ensuring secure and responsible LLM deployment.


Why the other options are incorrect:

A. Retrieval-augmented generation (RAG):
RAG is an architectural pattern that enhances LLM responses by retrieving relevant external data from a vector database. While RAG improves accuracy and reduces hallucinations, it is not a security resource and does not provide guardrail guidance. RAG itself requires guardrails, which OWASP would help define.

C. LLM libraries:
LLM libraries (e.g., Hugging Face Transformers, LangChain, LlamaIndex) provide development frameworks for building and deploying models. They offer tools for implementation but do not provide security frameworks, best practices, or compliance guidance for guardrails. They are implementation tools, not security resources.

D. Security incident and event management (SIEM):
SIEM is an operational tool for collecting, analyzing, and alerting on security logs and events. It monitors and detects threats after they occur but does not provide proactive guidance on building guardrails or meeting compliance requirements.

References:

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 3.2 (Implementing AI Security Controls – Guardrails) and Domain 2.0 (AI Security Risks and Threats).

OWASP Top 10 for LLMs (v1.1): The definitive resource for LLM-specific security risks and countermeasures.

Explanation:

The NIST AI RMF (AI Risk Management Framework) is the most appropriate resource for managing potential security issues related to model training. The NIST AI RMF provides comprehensive, actionable guidance specifically designed for AI systems across the entire lifecycle—including the training phase. It addresses key security risks such as data poisoning (malicious manipulation of training data), data leakage, bias introduction, model integrity, and adversarial robustness. The framework offers a structured approach to identify, assess, and mitigate risks throughout the AI lifecycle, with dedicated sections on securing training pipelines, validating data provenance, and monitoring for drift. Its AI-specific focus makes it directly applicable to model training security challenges.

Why the other options are incorrect:

B. International Organization for Standardization (ISO) 27001:
ISO 27001 is an information security management standard for general IT systems. While it provides a foundation for security governance, it does not address AI-specific threats such as data poisoning, model extraction, or adversarial attacks during training. It lacks the technical depth required for model training security.

C. Organization for Economic Co-operation and Development (OECD):
The OECD provides high-level principles for trustworthy AI (e.g., transparency, accountability, human-centered values). These are broad governance guidelines, not operational security frameworks for managing technical risks during model training.

D. General Data Protection Regulation (GDPR):
GDPR is a privacy and data protection regulation focused on handling personal data. While it imposes requirements on data used for training (e.g., consent, minimization, right to deletion), it does not address model security threats like poisoning, extraction, or adversarial robustness. It is a compliance regulation, not a security risk management framework.

References

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 1.3 (AI Lifecycle – Training Security) and Domain 3.0 (Securing AI Systems – Risk Management).

NIST AI RMF (2023): Explicitly addresses security risks across the AI lifecycle, including training data integrity and model robustness.

Explanation

Implementing MLOps directly improves the observability and auditing of an AI system by providing a structured, automated framework for monitoring, logging, and managing the AI lifecycle. MLOps establishes comprehensive telemetry across the entire pipeline—including data versioning, model versioning, training metrics, inference logs, drift detection, and performance tracking. Centralized logging and monitoring enable security analysts and auditors to trace every model interaction, identify anomalies, reconstruct events, and produce audit trails for compliance. MLOps also automates model validation and deployment, ensuring that changes are documented and approved. This systematic approach transforms AI operations from ad hoc processes into governed, repeatable workflows—significantly enhancing visibility and accountability.

Why the other options are incorrect:

A. Redeploying the model:
Redeploying a model simply pushes a new version into production. It does not improve observability or provide audit capabilities—it merely updates the model. Without MLOps, redeployment lacks version tracking and change documentation.

B. Using manual detection:
Manual detection—such as human review of alerts or logs—is inefficient, error-prone, and unscalable for AI systems that generate millions of events. It does not provide systematic audit trails or continuous observability. Manual processes are reactive, not proactive.

D. Using anomaly detections:
Anomaly detection is a specific monitoring technique that identifies outliers in model behavior, inputs, or outputs. While it contributes to observability, it is only one component of a broader observability framework. Anomaly detection alone does not provide comprehensive auditing, version control, or lifecycle management—all of which MLOps delivers.

References:

CompTIA SecAI+ CY0-001 Exam Objectives:Domain 1.3 (AI Lifecycle – MLOps and Model Monitoring) and Domain 4.2 (AI for Security Operations – Logging and Auditing).

NIST AI RMF: Recommends MLOps for continuous monitoring and auditability throughout the AI lifecycle.

Explanation:

Automated penetration testing is the technique the security team is most likely using. The description matches an AI-driven penetration testing tool that iteratively assesses the current security state, uses AI to select attack vectors, and adapts its approach based on previous outcomes—essentially simulating an intelligent adversary. This is a hallmark of autonomous or AI-assisted penetration testing, where the AI models decision-making processes, prioritizes high-value targets, and recommends or executes exploitation steps without continuous human intervention. By leveraging AI, the team can scale reconnaissance and exploitation efforts, cover more ground faster, and uncover complex attack chains that manual testers might miss.

Why the other options are incorrect:

A. Manual signature matching:
Manual signature matching involves comparing network traffic, files, or system behavior against a database of known indicators of compromise (IoCs) or attack signatures. It is a reactive, rule-based detection method and does not involve iterative attack vector selection or adaptive decision-making.

B. Code quality testing:
Code quality testing (e.g., static analysis, linters, or unit testing) evaluates software for bugs, maintainability, and security vulnerabilities in the source code. It does not "try to bypass organizational boundaries" or simulate adversarial behavior—it is a development-phase quality assurance activity.

C. Fraud detection:
Fraud detection uses AI/ML to identify anomalous transactions, user behaviors, or financial activities indicative of fraud. It is a defensive monitoring capability, not an offensive technique to breach organizational boundaries or discover attack paths.

References:

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 2.1 (AI-Powered Cyberattacks – AI for Penetration Testing) and Domain 4.1 (AI for Security Operations – Automated Red Teaming).

Explanation:

Response validation is the best control for mitigating the issue shown in the prompt. The LLM was asked for a simple list of application names containing "Content Management Solution," but the actual output included sensitive and irrelevant internal data—specifically, Application ID: 12345, Employee ID: 24563, Inherent Risk Rating: 1, Company Name: Marketing 123, Comments: Not effective. This indicates the model either hallucinated or inadvertently retrieved and exposed internal data that should never have been included in the response. Response validation involves applying a validation layer or schema enforcement on the model's output before it is returned to the user. This can include format enforcement (e.g., ensuring the output matches a predefined JSON schema), content filtering (e.g., stripping PII, employee IDs, or internal risk scores), and data leakage prevention (e.g., scanning for regex patterns like Employee ID: \d+). By validating the response against expected structure and content rules, the organization ensures that only safe, authorized, and appropriately formatted data is returned to the end user.

Why the other options are incorrect:

A. Model training:
Model training is the phase where the model learns from data. While improving training data quality could reduce hallucinations, it does not prevent the model from inadvertently including sensitive data in outputs. Retraining is not a practical real-time control for this already-deployed model. Additionally, training cannot enforce output structure or prevent leakage of internal IDs that the model may have memorized.

C. Access controls:
Access controls restrict who can query the model. In this scenario, the architect is authorized to use the LLM, so access controls would not have prevented the sensitive data from being returned. The issue is with what the model outputs, not who is asking.

D. Integrity monitoring:
Integrity monitoring tracks changes to system files, configurations, or model weights to detect tampering. It does not inspect or validate the content of model responses at inference time and would not prevent data leakage.

References:

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 3.2 (Implementing AI Security Controls – Output Validation and Data Leakage Prevention) and Domain 2.2 (Threats to AI Systems – Sensitive Information Disclosure).

Explanation:

ISO/IEC 42001 is the international standard specifically designed for Artificial Intelligence Management Systems (AIMS) . Published in December 2023, it provides a comprehensive framework for organizations to establish, implement, maintain, and continually improve an AI management system. ISO 42001 addresses the entire AI lifecycle—covering development, deployment, monitoring, and governance—with a strong emphasis on responsible AI practices including fairness, transparency, accountability, robustness, and data quality. It also includes requirements for risk assessment, impact analysis, and third-party assurance, making it the ideal standard for certification to demonstrate responsible AI practices to auditors, regulators, and business partners. Organizations seeking third-party validation of their AI governance and responsible AI maturity should pursue ISO 42001 certification.

Why the other options are incorrect:

A. ISO 20000:
ISO 20000 is the standard for IT Service Management (ITSM) . It focuses on service delivery, incident management, change management, and service-level agreements. It does not address AI-specific risks, ethics, or responsible AI practices.

B. ISO 27001:
ISO 27001 is the standard for Information Security Management Systems (ISMS) . While it provides a robust framework for protecting data and managing security risks, it is general-purpose and does not specifically address AI governance, algorithmic bias, explainability, or responsible AI principles. It complements ISO 42001 but is not the correct choice for responsible AI third-party assurance.

C. ISO 27701:
ISO 27701 is a privacy extension to ISO 27001, focusing on Privacy Information Management (PIMS). It addresses personal data protection, consent management, and GDPR compliance. While privacy is a component of responsible AI, ISO 27701 does not cover the broader scope of AI governance, fairness, transparency, or algorithm accountability.

References:

CompTIA SecAI+ CY0-001 Exam Objectives: Domain 1.4 (AI Ethics and Responsible AI) and Domain 3.0 (Securing AI Systems – Compliance and Standards).

ISO/IEC 42001:2023:The first global standard for AI management systems, explicitly designed for responsible AI governance.