CY0-001

Free CY0-001 practice questions

10 questions from our CY0-001 practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.

Back to the CY0-001 practice exams

Question 1 of 10, Basic AI Concepts Related to Cybersecurity

Five hospitals want to jointly train a model that detects anomalous access to patient records, but regulations prohibit any hospital from sending raw records to a shared location. Which technique allows them to train a shared model under this constraint?

  1. AQuantization of each hospital's local model before uploading the training records
  2. BTransfer learning from a public model followed by central retraining on pooled data
  3. CFederated learning, where each site trains locally and only model updates are aggregated Correct answer
  4. DData augmentation that creates synthetic copies of every hospital's records

Answer and explanation: Federated learning keeps raw data on each participant's systems and shares only parameter updates for aggregation, satisfying the no-data-movement rule. Central retraining on pooled data moves the records, which is prohibited. Synthetic augmentation does not by itself create a jointly trained model and still implies sharing data derived from records, and quantization only shrinks a model; it does not address where training data lives.

Question 2 of 10, Basic AI Concepts Related to Cybersecurity

An organization needs an on-premises model that classifies internal support tickets by urgency. It must run on modest CPU servers, keep all data internal, and does not need broad general knowledge. Which option BEST fits?

  1. AA reinforcement learning agent rewarded per ticket
  2. BA GAN trained to generate synthetic ticket text
  3. CA small language model fine-tuned on internal tickets Correct answer
  4. DA frontier LLM accessed through a public cloud provider's API

Answer and explanation: A small language model is cheap enough to run on modest hardware on premises, and fine-tuning it on internal tickets gives strong performance on a narrow task while data stays internal. A public cloud frontier LLM sends data outside and is oversized for the task. A GAN generates data rather than classifying it, and reinforcement learning needs an interactive reward environment that a classification task does not require.

Question 3 of 10, Securing AI Systems. Choose 2.

An LLM plug-in accepts a free-form URL parameter from the model and fetches it from inside the corporate network without validation. Testers use the assistant to make the plug-in retrieve internal admin pages. Which weaknesses and controls apply? (Choose two.)

  1. ARaise the model temperature so it generates fewer predictable URLs
  2. BMembership inference; return labels only instead of confidence scores
  3. CWatermark plug-in responses to prove which component produced them
  4. DApply network egress controls so the plug-in reaches only approved hosts Correct answer
  5. EInsecure plug-in design; validate parameters and restrict destinations Correct answer

Answer and explanation: A plug-in that trusts free-form model input is insecure plug-in design; strict parameter validation and allow-listed destinations, enforced by network egress controls, prevent server-side request forgery into internal systems. Membership inference controls address training-data leakage. Temperature changes randomness rather than enforcing security, and watermarking proves provenance but does not block requests.

Question 4 of 10, Securing AI Systems

A competitor appears to have released a model that behaves almost identically to a company's proprietary fraud model. Logs show one API key issued millions of systematically varied queries over several weeks. Which attack is MOST likely, and which control would have reduced it?

  1. AData poisoning; signed training datasets with integrity verification on load
  2. BMembership inference; differential privacy applied during model training
  3. CPrompt injection; a prompt firewall with jailbreak signature detection rules
  4. DModel theft through extraction; per-key query quotas with anomaly alerting Correct answer

Answer and explanation: High-volume, systematically varied queries used to replicate a model's behavior indicate model extraction, and quotas plus alerts on unusual query patterns limit how much an attacker can harvest. Poisoning alters training data rather than copying a model. Membership inference seeks to learn whether a record was in training data, and prompt injection manipulates instructions; neither explains a cloned model.

Question 5 of 10, Securing AI Systems

The 2025 edition of the OWASP Top 10 for LLM Applications introduced a category covering risks in how embeddings are generated, stored, and retrieved in RAG systems, such as cross-tenant leakage. What is this category called?

  1. AImproper output handling
  2. BUnbounded consumption
  3. CSensitive information disclosure
  4. DVector and embedding weaknesses Correct answer

Answer and explanation: Vector and embedding weaknesses addresses risks in RAG embedding pipelines, including unauthorized access and leakage across tenants. Unbounded consumption covers resource exhaustion and cost abuse. Sensitive information disclosure is the broader category for models revealing confidential data in general, and improper output handling covers unsafe use of model output downstream.

Question 6 of 10, Securing AI Systems

A red team needs to map observed behaviors against an ML-based fraud model to documented adversary tactics and techniques, including real-world case studies of attacks on AI systems. Which resource is MOST appropriate?

  1. AMITRE ATLAS knowledge base of adversarial AI tactics Correct answer
  2. BCVE AI Working Group vulnerability disclosures
  3. COWASP Machine Learning Security Top 10 risk list
  4. DMIT AI Risk Repository of categorized AI risks

Answer and explanation: MITRE ATLAS is structured as a matrix of adversary tactics and techniques against AI systems and includes case studies of real attacks, which is exactly what mapping red-team observations requires. The OWASP ML Security Top 10 ranks risk categories but is not a tactics-and-techniques matrix. The MIT AI Risk Repository catalogs broad AI risks, including societal ones, and the CVE AI Working Group focuses on how AI-related vulnerabilities are recorded, not on mapping adversary behavior.

Question 7 of 10, AI-assisted Security. Choose 2.

A threat intelligence team observes adversaries using AI to automatically combine leaked credentials, social media profiles, and exposed cloud assets into prioritized target lists. Which defensive measures MOST directly reduce the value of this activity? (Choose two.)

  1. ARequire longer password expiration periods for all user accounts
  2. BEnforce phishing-resistant MFA on all externally reachable services Correct answer
  3. CContinuously inventory and reduce the externally exposed attack surface Correct answer
  4. DMove all logs to a single storage account to simplify review
  5. ERely on annual awareness training as the only countermeasure

Answer and explanation: Automated data correlation turns leaked credentials and exposed assets into attack paths, so phishing-resistant MFA neutralizes leaked passwords and attack surface management removes exposed assets before they are found. Password expiration settings do not stop reuse of freshly leaked credentials. Consolidating logs aids review but does not reduce what attackers can correlate, and annual training alone does not remove exposed assets or protect leaked accounts.

Question 8 of 10, AI-assisted Security

A SOC wants to use an AI agent to triage phishing reports: extract indicators, check reputation services, and close obvious false positives. Leadership is concerned about the agent wrongly closing real incidents. Which design BEST balances automation and risk?

  1. ADisable the agent's integrations and use it only as a chatbot for answering analyst questions
  2. BHave the agent enrich and recommend, with analysts approving closures until accuracy is proven Correct answer
  3. CAllow the agent to close reports automatically but review a random 1% sample every quarter
  4. DLet the agent close any report scored below a fixed confidence value with no further review

Answer and explanation: Keeping a human in the loop for closure decisions while the agent handles enrichment captures most of the efficiency gain and prevents unreviewed mistakes until the agent's accuracy is validated. Fully automatic closure based on a fixed score or a tiny quarterly sample leaves real incidents exposed for long periods. Removing integrations eliminates the automation benefit that the SOC is seeking.

Question 9 of 10, AI Governance, Risk, and Compliance

Employees in marketing have started pasting customer lists into a free public chatbot to generate campaign copy, without IT approval. Which risk does this describe, and what is the BEST first response?

  1. AShadow AI; publish an acceptable use policy and offer a sanctioned tool Correct answer
  2. BModel skewing; add rate limits to the marketing team's network egress
  3. CExcessive agency; restrict the chatbot's plug-in permissions in settings
  4. DModel drift; retrain the public chatbot on corrected marketing data

Answer and explanation: Unsanctioned use of AI tools by staff is shadow AI, and the effective first response is a clear policy combined with an approved alternative so users have a safe option. The organization cannot retrain a public chatbot, and drift is not the issue. Rate limits do not prevent data leakage, and excessive agency concerns autonomous tool permissions rather than staff behavior.

Question 10 of 10, AI Governance, Risk, and Compliance

A company that sells an AI-based hiring tool to employers in Europe wants to understand its regulatory obligations, including risk classification and conformity requirements for high-risk systems. Which framework should it review?

  1. ANIST AI Risk Management Framework
  2. BISO/IEC 27001
  3. CEU AI Act Correct answer
  4. DOECD AI Principles

Answer and explanation: The EU AI Act is binding regulation that classifies AI systems by risk and places conformity obligations on high-risk uses such as employment decisions. The NIST AI RMF and OECD AI Principles are voluntary guidance without legal obligations. ISO/IEC 27001 is an information security management standard and does not classify AI system risk.