AI-Enabled Security and Future Developments

Introduction

Cybersecurity is entering a period in which defensive systems increasingly use machine learning, generative artificial intelligence, autonomous decision-making, deception, and post-quantum cryptography. These technologies can improve the speed and scale of security operations, but they also introduce new attack surfaces, sources of uncertainty, and governance challenges. A security system that relies on artificial intelligence is not automatically more secure than a conventional system. Its effectiveness depends on data quality, model robustness, operational integration, human oversight, and the ability to validate decisions under adversarial conditions.

Traditional security controls are largely rule-based. Firewalls, intrusion-prevention systems, antivirus engines, and security information and event management platforms compare observed activity with signatures, policies, or known indicators of compromise. These controls remain essential, but they are challenged by polymorphic malware, living-off-the-land techniques, supply-chain compromise, insider threats, encrypted traffic, and attacks that exploit previously unknown vulnerabilities. Machine learning can identify statistical patterns associated with malicious behavior, while generative AI can assist analysts in investigating and explaining events. At the same time, attackers can use the same technologies to generate convincing phishing messages, discover vulnerabilities, evade detection, and automate influence operations.

The central issue is therefore not whether AI will be used in cybersecurity, but how it should be designed, evaluated, governed, and combined with conventional controls. This chapter examines AI-enabled detection and response, generative AI in security operations, adversarial AI, automation and orchestration, quantum threats, autonomous defense, deception technologies, and future research directions.

Learning Objectives

After completing this chapter, a student should be able to:

1. Explain how machine-learning methods support cybersecurity detection and response.

2. Critically evaluate supervised, unsupervised, semi-supervised, and reinforcement-learning approaches.

3. Design a responsible architecture for generative AI in security operations.

4. Analyse adversarial attacks against AI systems and appropriate countermeasures.

5. Distinguish security automation, orchestration, and autonomous defense.

6. Assess quantum computing threats and the strategic importance of post-quantum cryptography.

7. Explain deception technologies and their role in detection and threat intelligence.

8. Evaluate operational, ethical, regulatory, and research challenges associated with AI-enabled security.

1. Machine Learning for Security Detection

1.1 Detection as a decision problem

A security detector observes evidence—such as network flows, process executions, authentication events, DNS requests, or cloud API calls—and produces a decision or prioritisation. The detector may classify an event as benign or malicious, assign a risk score, or recommend an investigation. In practice, detection is rarely a purely technical classification task. It is a decision problem involving incomplete information, changing environments, asymmetric costs, and limited analyst capacity.

A false positive occurs when benign activity is classified as malicious. A false negative occurs when malicious activity is missed. The relative cost of these errors depends on the asset and context. A false positive involving a low-value workstation may consume analyst time; a false negative involving a domain controller may permit a major breach. Consequently, accuracy alone is insufficient for evaluating a security model.

For binary classification, precision and recall are commonly used. If TP denotes true positives, FP false positives, and FN false negatives, then:

Precision = TP / (TP + FP)

Recall = TP / (TP + FN)

Precision measures the proportion of alerts labelled malicious that are actually malicious. Recall measures the proportion of malicious events detected. The F1 score is the harmonic mean of precision and recall:

F1 = 2(Precision × Recall) / (Precision + Recall)

These measures assume clearly defined labels and a meaningful event unit. In security, those assumptions are difficult. Labels may be incomplete, attack campaigns may generate many correlated events, and a model can obtain a high accuracy by predicting the majority benign class. Evaluation should therefore include false positives per analyst shift, detection delay, attack-path coverage, incident cost, and performance under distribution change.

1.2 Learning paradigms

Supervised learning uses labelled examples of benign and malicious activity. Classification algorithms include logistic regression, decision trees, random forests, gradient-boosted trees, support vector machines, and neural networks. Supervised methods can perform well when labels are representative and current. However, security labels are often noisy, expensive to obtain, and biased toward previously discovered attack techniques.

Unsupervised learning searches for structure without requiring labels. Clustering, density estimation, and dimensionality-reduction methods can identify unusual hosts, accounts, or communication patterns. An anomaly is not necessarily an attack: a legitimate software deployment may be statistically unusual. Unsupervised systems are therefore useful for discovery and triage, but they require contextual interpretation.

Semi-supervised and self-supervised learning use a small amount of labelled data and a large amount of unlabelled telemetry. Self-supervised methods learn representations by predicting properties of the data itself, such as the next event in a sequence or whether two observations belong to the same activity. These approaches can reduce labelling requirements, but representation quality and downstream validation remain important.

Reinforcement learning models an agent that selects actions and receives rewards or penalties. It has potential in adaptive access control, honeypot placement, and response planning. Deployment is difficult because exploratory actions can disrupt production systems. A safe reinforcement-learning design must constrain actions, use simulation or digital twins where possible, and require approval for high-impact changes.

1.3 Feature engineering and representation

Security features may be extracted from network, endpoint, identity, application, and cloud data. Examples include connection frequency, destination reputation, parent-child process relationships, authentication failure rates, privilege changes, executable provenance, file entropy, and deviations from a user’s normal access pattern. Features should be selected with awareness of privacy, operational meaning, and evasion risk.

A useful feature is not simply correlated with historical attacks; it should remain informative when attackers adapt. For example, a fixed list of suspicious file names may be easy to evade, whereas a relationship between an office application, a script interpreter, and an unusual outbound connection may represent a more durable behavioural signal. Graph representations are particularly useful for modelling relationships among identities, processes, hosts, domains, files, and cloud resources.

1.4 Concept drift and adversarial adaptation

Security data changes over time. Organisations deploy new software, users change working patterns, cloud architectures evolve, and attackers modify their techniques. This phenomenon is called concept drift. A model trained on last year’s data may become unreliable even if its implementation has not changed.

Drift management can include periodic retraining, continuous monitoring of input distributions, champion–challenger models, and human review of uncertain cases. Retraining must not be automatic without controls: poisoned or incorrectly labelled data can cause the model to learn harmful behaviour. Model versioning, reproducible pipelines, data lineage, and rollback mechanisms are essential.

2. From Detection to Response

2.1 Risk-based prioritisation

Machine learning is most useful when it improves a complete operational workflow rather than merely producing another alert. A practical risk score can combine evidence concerning the confidence of a detection, the criticality of the affected asset, the privileges of the account, the stage of the attack, and the potential business impact. Such a score should be interpretable enough for analysts to challenge it.

A model may identify a suspicious PowerShell process, but response depends on whether it occurred on a laboratory workstation, a payment server, or a domain controller. Contextual enrichment—asset inventories, identity data, vulnerability information, threat intelligence, and business ownership—is therefore as important as the classifier itself.

2.2 Detection and response metrics

Useful operational metrics include mean time to detect, mean time to contain, alert volume, alert-to-incident conversion rate, investigation effort, and recurrence rate. Metrics must be interpreted carefully. Reducing mean time to respond by automatically isolating every suspicious endpoint may improve speed while damaging availability and increasing false containment. Balanced evaluation should include security outcomes, service reliability, analyst workload, and user impact.

2.3 Human–AI collaboration

Human analysts contribute contextual reasoning, organisational knowledge, ethical judgement, and responsibility for exceptional decisions. AI systems contribute scale, consistency, rapid correlation, and the ability to examine large event volumes. The most reliable arrangement is usually a calibrated partnership rather than complete replacement of analysts.

Human oversight should be proportionate to impact. Low-risk actions, such as adding an indicator to a temporary investigation list, may be automated. High-risk actions, such as disabling a privileged account or shutting down a production service, should require strong evidence, approval, or carefully defined emergency procedures. Every automated action should produce an audit record explaining the trigger, evidence, policy, and outcome.

3. Generative AI in Security Operations

Generative AI systems produce text, code, queries, summaries, or other content from prompts and contextual data. In a security operations centre, possible applications include alert summarisation, natural-language querying of telemetry, incident timelines, detection-rule drafting, malware-analysis assistance, playbook generation, and translation of technical findings for management.

3.1 Retrieval-augmented security assistance

A language model should not be treated as an authoritative source of current organisational facts. A retrieval-augmented architecture first retrieves relevant content from approved sources—such as asset inventories, incident records, policies, and threat-intelligence repositories—and then supplies that material to the model as context. The response should cite the retrieved evidence and distinguish facts from hypotheses.

A secure design normally includes:

• identity-aware access to data;
• separation of tenants and incident cases;
• protection against prompt injection in logs, documents, and web content;
• output filtering and validation;
• controls preventing sensitive data from being sent to unauthorised providers;
• complete logging of prompts, retrieved sources, responses, and actions;
• human approval for consequential changes.

3.2 Benefits and failure modes

Generative AI can reduce the time required to interpret large volumes of technical information. It can help junior analysts understand unfamiliar commands and help experienced analysts search across heterogeneous evidence. However, language models can hallucinate—producing plausible but unsupported statements. They may also misinterpret temporal relationships, confuse similarly named assets, leak confidential information, or generate unsafe commands.

The correct operational pattern is therefore “assist, verify, execute under policy,” not “generate and trust.” Responses should be evaluated against source evidence, structured schemas, deterministic tools, and known test cases. Where possible, a model should call specialised tools for querying, reputation checks, sandboxing, or ticket creation rather than inventing results.

3.3 Generative AI for attackers

Attackers can use generative systems to create convincing multilingual phishing content, tailor social-engineering messages, generate malware variations, automate reconnaissance, and produce fraudulent audio or video. The defensive response includes stronger identity verification, phishing-resistant authentication, email and collaboration-platform controls, behavioural analytics, secure software development, and employee training that addresses synthetic content. Detection based solely on writing style is unreliable because generated content can be edited and legitimate writing can appear machine-generated.

4. Adversarial AI and Machine-Learning Security

Adversarial AI concerns attacks against AI systems, attacks enabled by AI, and strategic interactions in which both attacker and defender adapt. The machine-learning pipeline includes data collection, labelling, training, deployment, monitoring, and response. Each stage can be attacked.

4.1 Major attack classes

Data poisoning inserts or manipulates training data so that the model learns incorrect associations. A targeted poisoning attack may create a backdoor that activates for a particular pattern.

Evasion modifies an input at inference time to avoid detection. Examples include changing malware characteristics, altering network timing, or crafting an input that causes misclassification.

Model extraction attempts to reproduce a model by querying it repeatedly. The extracted model may reveal intellectual property or help an attacker design evasion inputs.

Membership inference attempts to determine whether a particular record appeared in training data. This can expose sensitive information, especially when models overfit.

Prompt injection and indirect instruction attacks target generative systems by placing malicious instructions in retrieved documents, emails, web pages, or logs. The model may then disclose information or invoke tools inappropriately.

4.2 Defensive principles

Defending AI requires security controls around both the model and its environment. Important practices include trusted data provenance, access-controlled datasets, anomaly checks on training data, robust validation, adversarial testing, rate limiting, output constraints, model isolation, least-privilege tool access, and monitoring for unusual queries or model behaviour.

Robustness must be tested under realistic transformations rather than only clean benchmark data. A model should be evaluated on time-separated data, new attack families, different network segments, and operational noise. Explainability tools can help analysts understand model behaviour, but explanations are not guarantees of causal correctness. A visually persuasive explanation may still be misleading.

5. Automation, Orchestration, and Autonomous Defense

Security automation is the execution of a defined task without manual intervention. Orchestration coordinates multiple tools and processes, commonly through security orchestration, automation, and response platforms. Autonomous defense goes further: a system senses conditions, selects actions, executes them, and adapts with limited human intervention.

The distinction is important. A playbook that enriches an indicator and opens a ticket is automation. A workflow that queries endpoint telemetry, checks asset criticality, updates a firewall, and records the case is orchestration. A system that independently changes defensive policy in response to an evolving attack is autonomous defense.

5.1 Designing safe response actions

Automation should begin with reversible, low-impact actions. Examples include collecting volatile evidence, increasing logging, requesting multifactor authentication, or placing an endpoint in a restricted network segment. Irreversible or disruptive actions require higher confidence and stronger approval. Policy should specify preconditions, limits, timeouts, rollback procedures, and escalation paths.

A mature architecture uses a closed-loop cycle:

1. Sense: collect and normalise telemetry.

2. Interpret: correlate evidence and estimate confidence.

3. Decide: select an action under policy constraints.

4. Act: execute through authenticated interfaces.

5. Verify: assess whether the action achieved its purpose.

6. Learn: record outcomes for improvement and audit.

The loop should fail safely. If telemetry is unavailable or confidence is low, the system should preserve evidence and escalate rather than take an unbounded action.

5.2 Risks of excessive autonomy

Autonomous systems may amplify errors at machine speed. A faulty detector can disable many legitimate accounts; a compromised orchestration platform can become a high-privilege attack mechanism; and a feedback loop can repeatedly reinforce an incorrect decision. Segregation of duties, strong authentication, change control, independent monitoring, and emergency shutdown mechanisms are therefore necessary.

6. Quantum Threats and Post-Quantum Security

Quantum computing creates a strategic challenge for public-key cryptography. Shor’s algorithm, if implemented on a sufficiently capable fault-tolerant quantum computer, would threaten widely used integer-factorisation and discrete-logarithm systems, including RSA and elliptic-curve cryptography. Grover’s algorithm provides a quadratic speedup for exhaustive search, affecting symmetric-key and hash security more gradually.

The practical danger exists before a cryptographically relevant quantum computer is available. An adversary may collect encrypted traffic today and attempt to decrypt it in the future, an approach often called “harvest now, decrypt later.” Information with a long confidentiality lifetime—such as diplomatic, health, industrial, or national-security data—requires early planning.

Post-quantum cryptography uses classical algorithms designed to resist known quantum attacks. Migration is not achieved by replacing one library alone. Organisations must inventory cryptographic dependencies, identify long-lived secrets, examine certificates and protocols, test performance and interoperability, plan hybrid deployments, and establish crypto-agility. Hybrid approaches may combine a conventional algorithm with a post-quantum algorithm during transition, subject to standards and implementation guidance.

Quantum-resistant migration also requires attention to signatures, key exchange, firmware updates, embedded devices, archives, and third-party services. Large public keys or signatures can affect bandwidth, storage, handshake latency, and constrained devices. Risk-based prioritisation is essential.

7. Deception Technologies

Deception technologies deliberately create controlled false resources or signals to detect, delay, and study attackers. Honeypots emulate vulnerable systems; honey credentials appear attractive but are monitored; decoy documents contain canary values; and deception grids distribute realistic but isolated assets throughout an environment.

Deception is valuable because legitimate users should rarely interact with carefully designed decoys. An interaction can therefore have a higher signal-to-noise ratio than a general anomaly alert. Deception also reveals attacker tools, movement patterns, and objectives. However, decoys must be isolated, monitored, maintained, and legally governed. A poorly configured honeypot can become a platform for attacking others or can expose sensitive internal information.

Effective deception is believable but bounded. It should imitate relevant technologies without reproducing unnecessary production privileges. Alerts should include context: which identity accessed the decoy, how it was reached, what commands were issued, and whether the activity suggests reconnaissance or lateral movement.

8. Illustrative Case: AI-Assisted Defense in a Financial Institution

Background

This illustrative case concerns a regional financial institution operating online banking, payment services, and a hybrid cloud environment. Its security operations centre receives approximately 100,000 daily events from endpoint, identity, network, and cloud platforms. Analysts complain that high alert volume delays investigation of account takeover and ransomware incidents.

Problem

The institution deploys a machine-learning model to identify unusual authentication and process activity. Initial testing shows good results on historical data, but performance declines after a major cloud migration. A generative assistant is then proposed to summarise alerts and recommend containment actions.

Analysis

The security team discovers three problems. First, historical labels overrepresent malware incidents and underrepresent identity abuse. Second, the model treats cloud service accounts as if they behaved like human users. Third, the generative assistant can access unclassified incident data but lacks reliable asset ownership and business-impact context. A pilot also reveals that malicious instructions embedded in an imported threat report can influence the assistant’s response.

Approach and solution

The institution adopts a layered architecture. Separate behavioural models are developed for human identities, service identities, endpoints, and cloud workloads. Evaluation uses time-separated data and measures precision, recall, alert burden, detection delay, and performance by attack scenario. Asset criticality is integrated into prioritisation rather than hidden inside an opaque score.

The generative assistant is limited to retrieval from approved sources. It produces evidence-linked summaries and structured recommendations but cannot directly disable accounts or modify firewall rules. Low-risk actions, such as collecting endpoint information, are automated. Account suspension requires analyst approval, except under a documented emergency policy. All tool calls are authenticated, logged, and checked against an allow-list.

Honey credentials are placed in isolated test environments and monitored. A post-quantum migration inventory identifies external-facing certificates and long-retention encrypted archives as early priorities.

Outcome

The pilot reduces repetitive investigation work and improves consistency of incident timelines. It does not eliminate false positives, but analysts spend more time on high-confidence cases. The institution also identifies model drift after a later software deployment and rolls back to a validated version. The most important outcome is governance: the organisation treats AI as a controlled security capability rather than an autonomous authority.

Learning points

1. Model performance must be evaluated under changing operational conditions.

2. Context and asset identity are essential for meaningful prioritisation.

3. Generative systems require evidence retrieval, tool restrictions, and human approval.

4. Safe automation is graduated according to impact and reversibility.

5. Migration planning for quantum threats should begin with cryptographic inventory and data lifetime.

9. Governance, Ethics, and Regulation

AI-enabled security involves sensitive personal and organisational data. Monitoring employee behaviour may be justified for security, but it can become disproportionate surveillance. Organisations should define purpose, minimise data collection, restrict access, establish retention periods, and provide accountability for consequential decisions.

Important governance questions include: Who owns the model? Who approves training data? Who is responsible when an automated response causes an outage? Can an analyst challenge a model decision? How are suppliers assessed? How are affected individuals notified when automated security controls deny access?

Security teams should align AI deployments with applicable privacy, cybersecurity, critical-infrastructure, sectoral, and emerging AI regulations. Compliance is not a substitute for engineering assurance. Documentation should cover intended use, limitations, data sources, performance by population or environment, known failure modes, monitoring, incident response, and decommissioning.

10. Future Research Directions

Several research directions are likely to shape advanced security practice:

• Privacy-preserving learning: federated learning, secure aggregation, and privacy-enhancing computation may support collaboration without centralising raw data, although they introduce communication, poisoning, and governance challenges.
• Causal security analytics: causal models may distinguish genuine attack drivers from correlations that merely reflect organisational routines.
• Foundation models for cyber operations: large models trained on code, logs, protocols, and attack knowledge may improve reasoning, but evaluation, provenance, and hallucination control remain open problems.
• Formal assurance of autonomous response: policy languages, verification, runtime enforcement, and safety cases could constrain high-impact automation.
• Graph and temporal reasoning: attack paths and campaigns are naturally relational and sequential, motivating graph neural networks and temporal models.
• Human-centred security operations: research is needed on calibrated trust, cognitive workload, explanation quality, and how analysts detect model failure.
• AI supply-chain security: organisations must secure datasets, pretrained models, plugins, model-serving infrastructure, and update channels.
• Quantum-safe migration engineering: practical interoperability testing and crypto-agility are as important as algorithm selection.

A significant research challenge is evaluation. Static benchmark datasets do not adequately represent adaptive attackers, changing environments, incomplete labels, or operational consequences. Future studies should combine controlled experiments, red-team exercises, longitudinal deployment evidence, and transparent reporting of failures.

Best-Practice Checklist

An organisation implementing AI-enabled security should:

1. Define the security problem and decision consequences before selecting a model.

2. Establish trustworthy data provenance, labels, access control, and retention.

3. Evaluate using realistic, time-separated, and adversarial test data.

4. Measure operational outcomes, not accuracy alone.

5. Monitor drift, uncertainty, false positives, and subgroup or environment differences.

6. Use retrieval and citations for generative assistance.

7. Restrict model access to tools and data by least privilege.

8. Require approval for high-impact or irreversible actions.

9. Log, test, version, and audit every model and playbook.

10. Maintain rollback, incident-response, and emergency shutdown procedures.

11. Include deception and conventional controls as complementary layers.

12. Begin cryptographic inventory and post-quantum migration planning early.

13. Train analysts to challenge AI outputs rather than accept them automatically.

14. Review privacy, ethics, supplier, and regulatory implications throughout the lifecycle.

Chapter Summary

AI can improve cybersecurity by identifying behavioural patterns, correlating large volumes of evidence, prioritising risk, assisting analysts, and automating carefully bounded response actions. Supervised, unsupervised, semi-supervised, and reinforcement-learning methods have different data requirements and operational risks. Precision, recall, and F1 are useful but must be complemented by measures such as detection delay, analyst workload, and business impact.

Generative AI is valuable for retrieval-based investigation assistance, summarisation, and workflow support, but hallucination, prompt injection, data leakage, and unsafe tool use require strong controls. Adversarial AI attacks the data, model, interface, and operational pipeline through poisoning, evasion, extraction, inference, and indirect instructions. Automation and orchestration should be graduated according to confidence, reversibility, and impact; autonomous defense requires policy constraints, verification, auditability, and human accountability.

Quantum computing threatens current public-key systems and motivates cryptographic inventory, crypto-agility, and post-quantum migration. Deception technologies provide controlled, high-value signals when decoys are realistic, isolated, and monitored. Future progress depends not only on more capable models but also on trustworthy data, robust evaluation, formal assurance, privacy protection, human-centred design, and responsible governance.

Key Terms

Machine learning: Computational methods that learn patterns from data to support prediction or decision-making.

Concept drift: Change over time in the relationship between inputs and the target behaviour a model must recognise.

Precision: The proportion of positive predictions that are true positives.

Recall: The proportion of actual positive cases detected by a system.

False positive: Benign activity incorrectly classified as malicious.

False negative: Malicious activity that a system fails to detect.

Generative AI: AI that produces new content such as text, code, summaries, or images.

Retrieval-augmented generation: Generation grounded in information retrieved from approved external sources.

Prompt injection: Malicious instructions designed to manipulate a generative model through direct or indirect input.

Adversarial AI: The study and practice of attacking or defending AI systems in security contexts.

Data poisoning: Manipulation of training data to impair or control model behaviour.

Evasion attack: Modification of an input to cause a deployed model to misclassify it.

Security automation: Execution of a defined security task without manual intervention.

Orchestration: Coordination of multiple security tools, processes, and actions.

Autonomous defense: Adaptive sensing, decision, and response with limited human intervention.

Post-quantum cryptography: Cryptographic methods designed to resist attacks from classical and quantum computers.

Crypto-agility: The ability to replace cryptographic algorithms and components with limited disruption.

Honeypot: A monitored decoy system designed to attract or detect unauthorised activity.

Model drift: Degradation caused by changes in data, environment, or attacker behaviour.

Human-in-the-loop: A design in which a person reviews or approves selected system decisions.

Review Questions

Short Answer Questions

1. Distinguish between precision and recall in security detection.

2. What is concept drift, and why does it affect deployed security models?

3. Define retrieval-augmented generation.

4. State two risks associated with autonomous defense.

5. Why is post-quantum migration required before large-scale quantum computers exist?

Long Answer Questions

1. Compare supervised, unsupervised, semi-supervised, and reinforcement learning for cybersecurity applications.

2. Explain how a generative AI assistant should be integrated into a security operations centre.

3. Discuss the principal adversarial attacks against machine-learning security systems and corresponding controls.

4. Analyse the differences among automation, orchestration, and autonomous defense.

5. Evaluate the role of deception technologies in a layered security architecture.

Analytical / Application Questions

1. A detector produces 80 true positives, 20 false positives, and 40 false negatives. Calculate its precision, recall, and F1 score, and interpret the operational implications.

2. Design a graduated response policy for an AI system that detects suspicious activity on a privileged administrator account. Identify actions that may be automated and actions requiring human approval.

3. A university stores encrypted research data that must remain confidential for twenty years. Develop a prioritised plan for assessing quantum risk and beginning post-quantum migration.

Multiple Choice Questions with Answers

Question 1

Which metric measures the proportion of detected positive alerts that are genuinely malicious?

A. Recall

B. Precision

C. Availability

D. Latency

Correct Answer: B. Precision

Question 2

Which learning approach is most directly dependent on labelled benign and malicious examples?

A. Supervised learning

B. Unsupervised learning

C. Reinforcement learning

D. Clustering without labels

Correct Answer: A. Supervised learning

Question 3

What is the primary purpose of retrieval-augmented generation in security operations?

A. To remove the need for access control

B. To ground generated responses in approved evidence

C. To guarantee that a model cannot hallucinate

D. To replace all security telemetry

Correct Answer: B. To ground generated responses in approved evidence

Question 4

Which attack manipulates training data to influence a model’s behaviour?

A. Evasion

B. Model extraction

C. Data poisoning

D. Denial of service

Correct Answer: C. Data poisoning

Question 5

Which action is generally most appropriate for early-stage security automation?

A. Destroying suspected evidence

B. Permanently disabling all suspicious accounts

C. Collecting additional evidence

D. Deleting production systems

Correct Answer: C. Collecting additional evidence

Question 6

Why is concept drift significant in cybersecurity machine learning?

A. Security data and attacker behaviour change over time

B. It guarantees higher recall

C. It eliminates the need for retraining

D. It affects only hardware performance

Correct Answer: A. Security data and attacker behaviour change over time

Question 7

Which cryptographic systems are principally threatened by Shor’s algorithm?

A. Public-key systems based on factoring or discrete logarithms

B. All password hashes immediately

C. Only physical access-control systems

D. Random-number generators with no cryptographic use

Correct Answer: A. Public-key systems based on factoring or discrete logarithms

Question 8

What is a key property of a well-designed honeypot?

A. It provides unrestricted production privileges

B. It is isolated and monitored

C. It must contain real customer data

D. It prevents all attacks from occurring

Correct Answer: B. It is isolated and monitored

Question 9

Which control most directly limits the damage a generative model could cause through tool use?

A. Least-privilege tool access

B. Increasing response length

C. Removing audit logs

D. Disabling source retrieval

Correct Answer: A. Least-privilege tool access

Question 10

A system has 80 true positives and 20 false positives. What is its precision?

A. 20%

B. 40%

C. 80%

D. 100%

Correct Answer: C. 80%