📘 Free CY0-001 Sample Questions
Which of the following job roles in an organizational governance structure develops a model from business use cases?
A
Platform architect
B
. AI risk analyst
C
. Machine learning operations (MLOps) engineer
D
Data scientist
Correct Answer:
D. Data scientist
Explanation:
The correct answer is Data scientist because this role is intrinsically responsible for translating business
challenges and use cases into data-driven solutions through the development of analytical and machine
learning models. A data scientist operates at the convergence of statistics, computer science, and domain
expertise, working to extract meaningful insights and construct predictive or prescriptive models tailored to
specific organizational demands. Their process commences with a deep understanding of the business
objective, which they then translate into a quantifiable problem, followed by the identification and acquisition
of pertinent data sources.
Key responsibilities include meticulous data collection, cleaning, exploratory data analysis, and feature
engineering to meticulously prepare the dataset for model training. Subsequently, they judiciously select
appropriate machine learning algorithms, train models using the processed data, and rigorously evaluate their
performance against predefined metrics. This cyclical process of model development, validation, and iterative
refinement is directly driven by the initial business use case, aiming to resolve specific problems like
predicting customer churn, optimizing logistical operations, or enhancing fraud detection mechanisms.
In stark contrast, the other roles serve distinct functions within the AI/ML ecosystem. A Platform architect
focuses on designing and implementing the foundational infrastructure and services that underpin AI/ML
solutions, ensuring their scalability, security, and seamless integration, rather than the development of the ML
models themselves. An AI risk analyst specializes in identifying, evaluating, and mitigating ethical, bias,
led by a data scientist.
privacy, and security risks associated with AI systems throughout their lifecycle, a role separate from model
creation. A Machine Learning Operations (MLOps) engineer concentrates on the deployment, continuous
monitoring, and ongoing maintenance of ML models in production environments, ensuring their operational
efficiency and robust lifecycle management, which typically occurs after the initial model development phase
Cloud computing concepts are fundamental to modern data science practices, as data scientists extensively
leverage cloud-based platforms and services for their complex tasks. These platforms provide highly scalable
compute resources (e.g., virtual machines, GPUs), robust managed data storage solutions (e.g., object storage,
data lakes), and specialized machine learning services (e.g., managed Jupyter notebooks, AutoML capabilities,
distributed training frameworks). This powerful infrastructure enables data scientists to conduct rapid
experimentation, train sophisticated models on vast datasets, and foster collaborative environments without
the burden of managing underlying hardware. Ultimately, the data scientist is the key player in
conceptualizing and building the data model based on business requirements.
Authoritative Links for Further Research:
IBM - What is a Data Scientist?: https://www.ibm.com/topics/data-scientist
Microsoft Azure - Data Scientist role description: https://learn.microsoft.com/en-us/azure/data-
explorer/data-scientist-landing-page
AWS - Machine Learning Specialist Roles: https://aws.amazon.com/certification/certified-machine-learning-
specialty/
rest within AI systems that involve data disclosure. Which of the following is the most suitable control?
A
Data lineage
B
Rate limits
C
. Encryption
D
Masking
Correct Answer:
C. . Encryption
Explanation:
For a financial institution, securing data at rest within AI systems, particularly concerning data disclosure,
requires robust controls. The most suitable control among the options is encryption because it directly
addresses the confidentiality of stored data.
Encryption transforms data into an unreadable format, rendering it unintelligible to any unauthorized entity
that gains access to the storage medium. This is paramount for "data at rest," which refers to data stored on
persistent storage like databases, file systems, data lakes, or object storage services that often underpin AI
training and inference. If a financial institution's storage infrastructure were to be compromised, the
encrypted data would remain protected, significantly mitigating the risk of unauthorized data disclosure.
Financial institutions operate under stringent regulatory frameworks, such as GDPR, CCPA, PCI DSS, and
various banking regulations, which frequently mandate encryption for sensitive customer and transactional
data. Implementing encryption ensures compliance and helps maintain customer trust by demonstrating a
commitment to data privacy. In cloud computing environments, where AI systems are commonly deployed,
cloud providers offer robust, managed encryption services (e.g., AWS KMS, Azure Key Vault, Google Cloud
KMS) that integrate seamlessly with storage services (e.g., S3, Blob Storage, Cloud Storage). These services
simplify key management and automate much of the encryption process, making it easier to implement at
scale.
Authoritative Links for Further Research:
While other options have their merits, they are less suitable as the primary control for securing data at rest
against disclosure. Data lineage helps track data's origin and transformations, crucial for governance and
auditing, but it doesn't protect the data's confidentiality itself. Rate limits control access frequency to prevent
abuse or denial-of-service attacks, focusing on availability and operational security, not the inherent
confidentiality of stored data. Masking is useful for anonymizing data for specific purposes like testing or
analytics, but it typically applies to copies or views of data and doesn't inherently protect the underlying,
original data at rest from a direct breach if it's stored unmasked. Therefore, encryption stands out as the
foundational security measure for protecting data at rest from unauthorized disclosure in high-stakes
environments like financial institutions leveraging AI.
NIST Special Publication 800-111: Guide to Storage Encryption Technologies for End User Devices: While
focused on end-user devices, its principles for data at rest encryption are universally applicable.
https://nvlpubs.nist.gov/nistpubs/legacy/sp/nistspecialpublication800-111.pdf
Cloud Security Alliance (CSA) Cloud Controls Matrix (CCM): Provides security principles and controls,
including specific guidance on data encryption at rest in cloud environments.
https://cloudsecurityalliance.org/research/artifacts/cloud-controls-matrix/
AWS Key Management Service (KMS) Documentation: An example of how cloud providers implement
A security engineer needs to monitor an AI-based system for runtime operations. The engineer is mostly concerned
about the visibility of internal activity. Which of the following is the most appropriate monitoring solution?
A
Deploying a security information and event management (SIEM) tool
B
Implementing a web application firewall (WAF) with header logging
C
Relying on vendor model controls and monitoring prompt inputs
D
Enabling stack call and debugging level traces at the function level
Correct Answer:
D. Enabling stack call and debugging level traces at the function level
Explanation:
The security engineer's primary concern is obtaining visibility into the internal activity of an AI-based system
during runtime operations. This implies a need to understand the granular processes, computations, and
decision-making happening within the AI model and its supporting application logic, beyond just its inputs and
outputs.
Option D: Enabling stack call and debugging level traces at the function level is the most appropriate
solution because it directly addresses this requirement for deep internal visibility.
1. Granular Insight: Stack call traces provide a precise sequence of function executions, revealing the
exact path code takes during runtime. Debugging level traces capture detailed information about
emerge internally.
system arrived at a particular output.
variables, intermediate computations, control flow, and state changes within each function. For an AI
system, this translates to understanding how input data is processed through various layers, how
features are extracted, how models make predictions, and where potential anomalies or biases might
2. AI Specificity: AI systems, particularly complex deep learning models, often behave as "black
boxes." Detailed function-level traces are crucial for explainable AI (XAI) and for debugging
logical errors that a high-level log might miss.
unexpected behaviors, performance issues, or security vulnerabilities like data poisoning or model
inversion attacks that manifest internally. It allows engineers to pinpoint exactly where and why an AI
3. Runtime Observability: This method provides real-time, in-depth monitoring of the system's
operational mechanics, which is vital for detecting subtle shifts in behavior, resource consumption, or
4. Cloud Integration: In cloud environments, logging and monitoring services (e.g., AWS CloudWatch
Logs, Google Cloud Logging, Azure Monitor Logs) are designed to ingest, store, and analyze high
volumes of detailed log data, including these granular traces. Tools like OpenTelemetry enable
distributed tracing across microservices, providing end-to-end visibility of internal calls even in
complex, cloud-native AI architectures. This level of detail helps in correlating issues across services
and identifying bottlenecks or security events originating deep within the system.
Why other options are less appropriate:
A. Deploying a security information and event management (SIEM) tool: While essential for aggregating
logs and correlating security events across an enterprise, a SIEM consumes logs. It does not generate the
deep internal, function-level traces required to understand the AI system's internal workings. If the underlying
application doesn't provide such granular logs, the SIEM won't have them to analyze for internal activity.
B. Implementing a web application firewall (WAF) with header logging: A WAF operates at the perimeter,
protecting the application from external web-based threats. While header logging provides valuable
information about incoming requests and outgoing responses, it offers no visibility whatsoever into the
internal processing, computational steps, or decision-making logic happening inside the AI application.
C. Relying on vendor model controls and monitoring prompt inputs: Vendor model controls are often black-
box solutions that provide guardrails but limited insight into the internal mechanisms. Monitoring prompt
inputs is crucial for detecting prompt injection or adversarial inputs, but it only captures the external interface
with the AI system. It fails to provide visibility into the internal steps the AI model takes to process those
prompts and generate outputs, which is the engineer's explicit concern.
Therefore, enabling stack call and debugging level traces offers the most comprehensive and direct approach
to gaining the desired visibility into an AI system's internal runtime activity.
Authoritative Links for Further Research:
1. OpenTelemetry - Distributed Tracing: A vendor-neutral open standard for instrumenting, generating,
collecting, and exporting telemetry data (traces, metrics, logs) to help analyze software performance
and behavior. It's fundamental for deep internal visibility in cloud-native applications.
OpenTelemetry Official Website
2. Microsoft Azure Monitor Application Insights - Distributed Tracing: An example of a cloud-native
service providing application performance monitoring and distributed tracing capabilities, which
would be used to collect and visualize such function-level traces.
What is distributed tracing?
3. Google Cloud Operations Suite (formerly Stackdriver) - Logging: Demonstrates how cloud providers
offer robust logging capabilities capable of ingesting high-volume, detailed application logs
including debug and trace levels.
Cloud Logging Overview
4. AWS CloudWatch Logs - Centralized Logging: Similar to Google Cloud, AWS provides services for
collecting, monitoring, and storing logs from applications running on AWS, supporting the ingestion
of detailed debug and trace logs.
What is Amazon CloudWatch Logs?
5. Explainable AI (XAI) Concepts: While not directly about tracing, XAI methodologies often rely on
understanding internal model behavior, for which detailed debugging information is foundational.
Explainable AI (XAI) for Dummies (IBM provides good introductory material)
Which of the following should an auditor reference when reviewing a company's human resources Al systems for
legal non-compliance?
A
Organization for Economic Cooperation and Development (OECD) standard
B
National Institute of Standards and Technology (NIST) Al Risk Management Framework 9RMF)
C
European Union (EU) AI Act
D
International Organization for Standardization (ISO)
Correct Answer:
C. European Union (EU) AI Act
Explanation:
The correct answer is C. European Union (EU) AI Act.
The EU AI Act is the most appropriate reference for an auditor reviewing a company's human resources (HR)
AI systems for legal non-compliance because it is a comprehensive, legally binding regulation specifically
designed to govern artificial intelligence systems within the European Union. Unlike the other options, which
are frameworks, guidelines, or voluntary standards, the EU AI Act carries the force of law, including
significant penalties for non-compliance.
HR AI systems, such as those used for recruitment, performance evaluation, or worker monitoring, are
explicitly categorized as "high-risk" under the EU AI Act. This designation is due to their potential to
significantly impact individuals' fundamental rights, including the right to non-discrimination, privacy, human
dignity, and fair working conditions. Consequently, systems deemed high-risk are subject to stringent
requirements concerning data governance, transparency, human oversight, robustness, accuracy,
cybersecurity, and conformity assessment.
For an auditor, this means directly referencing the Act's detailed provisions to assess whether the company's
HR AI systems meet these mandatory legal obligations. The audit would involve examining documentation,
system design, data quality, impact assessments, and operational procedures against the specific articles of
the Act. For instance, an auditor would check if the company has conducted a fundamental rights impact
assessment and implemented appropriate risk mitigation measures as required.
maintaining compliance.
Authoritative Links for Further Research:
framework-ai
In a cloud computing context, where many HR AI systems are deployed as Software as a Service (SaaS) or AI
as a Service (AIaaS), the EU AI Act places responsibilities on both the AI system providers and the
organizations deploying them. An auditor would need to scrutinize contractual agreements with cloud
providers to understand the shared responsibility model for compliance. This includes ensuring that the cloud
environment supports the necessary data governance, security, and auditability requirements mandated by
the Act, particularly regarding data residency, access controls, and the ability to conduct independent audits
of the AI models. The company using the cloud-based HR AI system is ultimately responsible for ensuring its
use complies with the Act, regardless of where the system is hosted or who developed it. Therefore, an
auditor would verify the company's due diligence in selecting cloud AI vendors and its internal processes for
European Commission - The EU AI Act: https://digital-strategy.ec.europa.eu/en/policies/regulatory-
Council of the EU - AI Act Final Text: https://www.consilium.europa.eu/en/press/press-
releases/2023/12/08/artificial-intelligence-council-and-parliament-reach-a-provisional-agreement-on-the-ai-
act/ (This often links to the latest agreed-upon text once formally adopted and published in the Official
Journal.)
Official Journal of the European Union (search for AI Act once published): https://eur-
lex.europa.eu/homepage.html
An airline corporation wants to implement a chatbot application using a large language model (LLM) so its
customers:
Can ask question and receive answers about flight details.
Have the option to upload files.
Which of the following security controls should the airline use to protect against malicious input and unauthorized
use beyond the service-level agreement? (Choose two.)
A
Prompt guardrails
B
Role-based access controls
C
Firewall rules
D
Model token quotas
Correct Answer:
A. Prompt guardrails
Explanation:
To effectively protect an airline's chatbot application from malicious input and unauthorized use beyond
service-level agreements (SLAs), prompt guardrails and model token quotas are the most direct and crucial
security controls.
Prompt guardrails are essential for protecting against malicious input, such as prompt injection or
jailbreaking attempts. These guardrails define specific rules and policies that dictate the chatbot's
acceptable behavior, topics, and output. For an airline chatbot, guardrails would ensure the model stays
focused on flight details, prevents it from revealing sensitive internal corporate information, and rejects any
prompts attempting to manipulate its core function or elicit harmful responses. This application-layer security
directly mitigates the risks of attackers coercing the LLM into unauthorized actions or generating
inappropriate content, aligning with the principle of least privilege for the chatbot's functionality.
Model token quotas, on the other hand, are vital for protecting against unauthorized use beyond the service-
level agreement. Tokens are the fundamental units of text processed by LLMs, and setting quotas limits the
amount of processing an individual user or application can consume within a given timeframe. This control
directly addresses potential denial-of-service (DoS) attacks by preventing malicious actors from deliberately
flooding the chatbot with excessive requests to exhaust computing resources, incur exorbitant costs, or
degrade service for legitimate customers. By enforcing these usage limits, token quotas ensure fair usage,
maintain service availability, and act as a critical cost management mechanism in a cloud environment,
directly preventing the unauthorized use that could violate an SLA.
choices for the outlined security concerns.
Authoritative Links for Further Research:
https://llm.owasp.org/
While other options like Role-Based Access Controls (RBAC) and firewall rules are important for overall
security, they don't directly address the specified threats in the same manner. RBAC primarily governs
internal access to the LLM infrastructure and data, not the inputs from external customers to the public
chatbot. Firewall rules operate at the network layer and cannot interpret the semantic content of malicious
prompts or enforce LLM-specific usage limits. Therefore, prompt guardrails counter malicious input
effectively, and model token quotas manage unauthorized resource consumption, making them the superior
Prompt Guardrails/LLM Security: OWASP Top 10 for Large Language Model Applications:
Model Token Quotas/API Management: AWS Documentation on Throttling and Quotas (General API Gateway
concept, applicable to LLM APIs): https://docs.aws.amazon.com/apigateway/latest/developerguide/api-
gateway-request-throttling.html (While specific to API Gateway, the concept of managing requests and
tokens is universal for LLM service consumption)
General LLM Security Best Practices: Microsoft Azure AI Security: https://learn.microsoft.com/en-
us/azure/ai-services/concepts/ai-security
A security operations center (SOC) has a very high volume of logs and alerts. The manager proposes the
implementation of machine learning (ML) system to help with triage. Which of the following tasks is most suitable?
A
Applying filters on specific alerts
B
. Automatically patching vulnerable systems
C
Identifying and classifying alerts
D
Summarizing the content of alerts
Correct Answer:
C. Identifying and classifying alerts
Explanation:
A security operations center (SOC) facing an overwhelming volume of logs and alerts requires intelligent
automation to efficiently manage triage. Of the options provided, Identifying and classifying alerts is the
most suitable task for a machine learning (ML) system.
ML excels at pattern recognition and anomaly detection across vast datasets, making it ideal for sifting
through security logs. By training on historical data, an ML model can learn to distinguish between benign
noise, false positives, and genuine threats. This capability allows the system to automatically categorize
alerts based on severity, threat type, or affected assets, effectively prioritizing critical incidents. This
automation significantly reduces alert fatigue for security analysts, enabling them to focus their expertise on
high-fidelity alerts that require immediate investigation and response. Ultimately, ML enhances the efficiency
and accuracy of the initial triage process, which is fundamental to a well-functioning SOC.
In contrast, other options are less suitable as the primary ML task for initial triage:
Applying filters on specific alerts is often accomplished through predefined rules, regular expressions, or
basic correlation logic within a Security Information and Event Management (SIEM) system. While ML can
refine filter efficacy by identifying evolving patterns, the core task of "applying filters" is generally rules-
based rather than a direct, primary ML application for high-volume triage.
fundamental step in managing alert volume.
For further research:
Automatically patching vulnerable systems involves high-risk actions that typically require human oversight
due to potential operational impacts and unintended consequences. While Security Orchestration,
Automation, and Response (SOAR) platforms might leverage ML for recommendations, fully autonomous
patching by an ML system without human review is not advisable as a core triage function.
Summarizing the content of alerts using Natural Language Processing (NLP) is valuable, but its utility for
triage is secondary to classification. Before summarization becomes impactful, the system or analyst first
needs to determine which alerts are important enough to warrant a detailed summary. Classification is a more
From a cloud computing perspective, implementing ML for alert classification is highly practical. Cloud-based
SIEM solutions (e.g., Azure Sentinel, AWS Security Hub) can aggregate massive log volumes cost-effectively.
Managed ML services (e.g., AWS SageMaker, Azure Machine Learning, Google Cloud AI Platform) provide
scalable infrastructure and tools for building, training, and deploying ML models without managing underlying
servers. The scalability of cloud platforms is crucial for processing the fluctuating and high volume of security
logs required to train robust ML models and perform real-time inference.
NIST IR 8276, "Artificial Intelligence in Cybersecurity and Protecting AI Systems: Challenges and
Opportunities": https://nvlpubs.nist.gov/nistpubs/ir/8276/NIST.IR.8276.pdf
Microsoft Azure, "Applying machine learning to cybersecurity": https://www.microsoft.com/en-
us/security/blog/2023/04/18/applying-machine-learning-to-cybersecurity/
IBM, "How AI and machine learning enhance cybersecurity": https://www.ibm.com/topics/ai-cybersecurity
An organization recently created a custom model that integrates with a language model (LLM). The developer
notices that the application programming interface (API) costs have increased. Which of the following is the best
control to reduce cost?
A
Implementing prompt templates
B
Increasing central processing unit (CPU) and memory
C
Reducing the model size
D
. Adjusting token limits
Correct Answer:
D. . Adjusting token limits
Explanation:
Increased API costs for an LLM integration are almost universally tied to token consumption, which
represents the fundamental billing unit for most language model providers. Tokens are segments of words or
characters that the LLM processes for both input (prompts) and output (completions). The more tokens an
application sends to and receives from the LLM API, the higher the incurred cost.
Adjusting token limits is the most direct and effective control to reduce these expenses. This parameter, often
called max_tokens or similar, explicitly sets the maximum number of tokens the LLM can generate in response
to a prompt. By implementing a lower token limit, developers can prevent the model from generating overly
verbose or unnecessarily long responses, which directly curtails the output token count and, consequently,
the associated API cost. This acts as a hard cap on expenditure per response, providing immediate and
predictable cost savings.
In contrast, other options are less impactful for direct API cost reduction. Implementing prompt templates (A)
can standardize inputs, potentially leading to more efficient prompts, but it doesn't inherently impose a limit
on the output length or guarantee fewer overall tokens. While well-designed templates might subtly influence
output brevity, it's not a direct cost control mechanism like an explicit token limit. Increasing central
processing unit (CPU) and memory (B) pertains to the computational resources of the application or hosting
environment, not the billing structure of a third-party LLM API. These resources affect the application's
performance or the cost of self-hosting a model, not the per-token charges from an external provider.
Reducing the model size (C) refers to switching to a smaller, less complex LLM. While smaller models
generally have lower per-token costs, this is an architectural choice, not a continuous control over usage
within an already selected model. Moreover, for an organization integrating with an LLM, reducing "the model
size" isn't a runtime parameter; they typically choose from available model sizes (e.g., gpt-3.5-turbo instead of
gpt-4). Adjusting token limits, however, provides granular control over consumption regardless of the chosen
model.
Therefore, by setting strict token limits, an organization directly manages the maximum financial outlay for
each LLM interaction, making it the best control for mitigating rising API costs.
Authoritative Links for Further Research:
OpenAI Pricing: Provides detailed information on token-based pricing for their
models.https://openai.com/pricing
Google Cloud Vertex AI Generative AI Pricing (Gemini, PaLM): Explains cost structures based on tokens for
Google's LLMs.https://cloud.google.com/vertex-ai/generative-ai/pricing
Anthropic Claude API Pricing: Details their token-based pricing for various Claude
models.https://www.anthropic.com/api/docs/pricing
A security administrator needs to improve an Al model. During an initial investigation, the administrator notices
that two successive login features are recorded every day, and then a successful login occurs after a specific time
interval. All the successful login attempts have been during office hours.
Which of the following techniques should the administrator use to improve the Al model's security?
A
Access management
B
Pattern recognition
C
Signature matching
D
Vulnerability analysis
Correct Answer:
B. Pattern recognition
Explanation:
To improve the AI model's security given the described scenario, the administrator should use pattern
recognition. The investigation reveals a consistent sequence: two successive login "features" (which could
imply failed attempts, specific credential types, or other login-related data) daily, followed by a successful
login after a distinct time interval, always occurring within office hours. This sequence is a classic example of
a recurring behavior, or a "pattern."
Pattern recognition is an AI technique used to identify regularities, trends, or specific structures within data.
In this context, the AI model needs to learn to identify and understand this specific sequence of events. By
recognizing this pattern, the AI can establish a baseline for what constitutes "normal" or expected behavior
related to these specific logins. For instance, it might indicate a legitimate automated process, a user with a
complex login flow, or even a sophisticated attack sequence.
Simply put, the AI must first recognize this unique pattern before it can effectively classify it as benign or
malicious, or to flag deviations from it. Without pattern recognition, the AI would treat each login event in
isolation or struggle to connect the "two successive login features" with the subsequent successful attempt
and its timing. This technique allows the model to differentiate between expected operational patterns and
genuine anomalies that might signal a security threat.
In a cloud computing environment, where vast amounts of telemetry data are generated from identity and
access management (IAM) systems, virtual machines, serverless functions, and other services, pattern
recognition is crucial. Cloud security platforms (like SIEM, UEBA, and CSPM solutions) leverage machine
learning to ingest, normalize, and analyze these logs at scale. They build profiles of user and entity behavior
by recognizing patterns over time and across distributed cloud resources, enabling the detection of advanced
persistent threats, insider threats, and sophisticated attacks that span multiple cloud services.
observing and understanding the login sequence.
Authoritative Links for Further Research:
with-ueba
Access management focuses on enforcing policies for who can do what; it doesn't describe the technique for
Signature matching relies on comparing events against known malicious patterns (signatures) but is less
effective against novel or evolving threats, or in understanding unique legitimate patterns. The described
scenario is a behavior, not necessarily a known malicious signature.
Vulnerability analysis identifies weaknesses in systems; while the observed pattern might point to a
vulnerability, vulnerability analysis is not the technique for identifying and understanding the pattern itself.
Therefore, pattern recognition is the foundational technique required for the AI model to learn, adapt, and
improve its ability to secure the environment by making sense of the observed login sequence.
AWS Machine Learning in Security: Provides an overview of how machine learning, including pattern
recognition, is applied to security services for threat detection and anomaly identification in cloud
environments. https://aws.amazon.com/security/machine-learning/
Microsoft Azure Sentinel (UEBA) - Identify threats with UEBA: Explains how User and Entity Behavior
Analytics (UEBA) in Azure Sentinel uses machine learning and pattern recognition to detect anomalous
behavior and threats across a cloud estate. https://learn.microsoft.com/en-us/azure/sentinel/identify-threats-
NIST Artificial Intelligence (AI) and Cybersecurity: Provides foundational information and frameworks for
integrating AI into cybersecurity, which implicitly relies on techniques like pattern recognition for effective
threat detection and response. https://www.nist.gov/artificial-intelligence-and-cybersecurity
Which of the following is the most concerning risk for a company that allows corporate end users to use public-
facing large language models (LLMs)?
A
Inaccuracies due to hallucinations
B
Out-of-date acceptable use policies
C
Data security regulatory violations
D
Malicious code generation
Correct Answer:
C. Data security regulatory violations
Explanation:
For a company allowing corporate end users to interact with public-facing Large Language Models (LLMs),
"Data security regulatory violations" stands out as the most concerning risk due to its potential for severe
legal, financial, and reputational ramifications. When employees input sensitive company data—such as
personally identifiable information (PII), protected health information (PHI), intellectual property, trade
secrets, or confidential business strategies—into a public LLM, that data is processed, stored, and potentially
used for model training by the LLM provider. This process effectively transfers custody and control of the
sensitive data to a third party, often without the necessary contractual safeguards or visibility required by
regulations.
This practice directly conflicts with numerous stringent data protection frameworks worldwide, including the
General Data Protection Regulation (GDPR), California Consumer Privacy Act (CCPA), Health Insurance
Portability and Accountability Act (HIPAA), and various industry-specific regulations like NYDFS
Cybersecurity Regulation. These regulations impose strict requirements on how organizations collect,
process, store, transmit, and secure personal and sensitive data, often demanding specific controls over data
residency, data sovereignty, and third-party data handling agreements.
business.
A breach or unauthorized disclosure of data via a public LLM can trigger significant penalties. For instance,
GDPR violations can result in fines up to €20 million or 4% of annual global turnover, whichever is higher,
while HIPAA violations can lead to millions in fines per incident. Beyond financial penalties, companies face
costly legal battles, class-action lawsuits, mandatory public breach notifications, and intensive regulatory
scrutiny.
From a cloud computing perspective, using public LLMs with sensitive data introduces an uncontrolled third-
party risk that bypasses established enterprise data governance and security architectures. Organizations
lose direct oversight and often violate the principles of data residency and sovereignty crucial for compliance,
as they cannot guarantee where or how their data is stored or processed by the LLM provider. This
fundamentally undermines the organization's ability to demonstrate compliance and protect data effectively.
While other risks like hallucinations or malicious code generation are serious, they typically lead to
operational inefficiencies, reputational damage, or specific security incidents; data security regulatory
violations, however, carry direct, legally mandated, and often catastrophic consequences that can cripple a
Authoritative Links for Further Research:
GDPR Official Text: https://gdpr-info.eu/
California Consumer Privacy Act (CCPA): https://oag.ca.gov/privacy/ccpa
NIST AI Risk Management Framework (AI RMF): https://www.nist.gov/artificial-intelligence/ai-risk-
management-framework (Provides context on managing AI risks, including privacy implications)
Which of the following requires developers to harden infrastructure to protect AI systems?
A
Intake processes
B
. Acceptable use policies
C
Development guidelines
D
Configuration standards
Correct Answer:
D. Configuration standards
Explanation:
The correct answer is D. Configuration standards.
Configuration standards are essential technical documents that define the secure baseline for how systems,
networks, applications, and infrastructure components must be configured. These standards specify
mandatory security settings, hardening measures, and operational procedures to mitigate vulnerabilities. For
AI systems, which often rely on complex infrastructure, including cloud computing resources, data storage,
and specialized compute (e.g., GPUs), robust configuration standards are paramount.
Developers are required to adhere to these standards to ensure the underlying infrastructure hosting AI
models, training data, and inference engines is hardened against various threats. This includes mandates for
secure operating system settings, network access controls, encryption protocols for data at rest and in
transit, patch management, privileged access management, and logging configurations. Without these
prescriptive guidelines, infrastructure could be deployed with default, insecure settings, leaving AI systems
vulnerable to unauthorized access, data breaches, model poisoning, or denial-of-service attacks.
Authoritative Links for Further Research:
In a cloud computing environment, configuration standards directly address the customer's responsibility
under the shared responsibility model. While cloud providers secure the "security of the cloud," customers are
responsible for "security in the cloud," which includes configuring their deployed resources securely.
Adherence to configuration standards ensures that developers provision and maintain compute instances,
data lakes, machine learning platforms, and other services with a strong security posture. These standards
are often implemented through Infrastructure as Code (IaC) to ensure consistency and prevent configuration
drift, thereby maintaining continuous hardening of the AI infrastructure. They establish a non-negotiable
security baseline, making them a direct requirement for developers to protect AI systems.
1. NIST Special Publication 800-53, Revision 5: Security and Privacy Controls for Information
Systems and Organizations: This publication provides a catalog of security and privacy controls for
all U.S. federal information systems, which heavily influences industry best practices, including
configuration management.
https://csrc.nist.gov/publications/sp/800-53/rev-5/final
2. CIS Benchmarks (Center for Internet Security): These are globally recognized and accepted best-
practice configuration guides for securely hardening systems. They cover a wide range of
technologies, including operating systems, cloud platforms, and network devices.
https://www.cisecurity.org/cis-benchmarks/
3. AWS Well-Architected Framework - Security Pillar: While specific to AWS, the principles outlined in
the security pillar, particularly around configuration management and operational excellence, are
universally applicable to securing cloud infrastructure.
https://docs.aws.amazon.com/wellarchitected/latest/security-pillar/security-pillar.html
Questions: 1-10 out of 126
Continue Full Practice..
GET ALL 126 QUESTIONS