A is correct: Peer-group behavioral analysis is the correct answer. UEBA systems use machine learning to group users by shared attributes such as department, role, and typical access patterns, then compare each user's activity against their peer group. When an accounting user accesses source code repositories that only software engineers use, peer-group analysis flags this as a significant deviation from how similar users behave, making it a strong indicator of potential insider threat or credential compromise.

B is incorrect: Network intrusion signature matching relies on comparing network traffic against databases of known attack signatures. This technique detects previously identified attack patterns but cannot identify behavioral deviations such as a user accessing resources that their peer group does not typically use. The scenario describes anomalous user behavior relative to peers, not a known network-based attack signature.

C is incorrect: Scheduled endpoint vulnerability scanning identifies security weaknesses in systems and applications, such as missing patches or misconfigurations. It does not monitor or analyze user behavior patterns or compare a user's activity against their peer group. Vulnerability scanning addresses system security posture rather than behavioral anomaly detection for insider threats.

D is incorrect: Static access control list evaluation determines whether a user has the permissions to access a resource but does not evaluate whether the access is behaviorally anomalous compared to peer-group norms. If the accounting user has been granted legitimate access to the repository, ACL evaluation alone would not flag this activity, whereas peer-group analysis identifies it as unusual for users in that role.

A is incorrect: The SOAR playbook does not replace the SIEM's detection capabilities. The SIEM generates the initial alert that triggers the playbook, and the SOAR enriches it with additional data from other tools. The AI severity scoring supplements detection by adding context-aware triage, but the SIEM remains the primary detection and correlation platform within the overall security architecture.

B is incorrect: While SOAR playbooks automate many response steps, they do not eliminate the need for human analysts. High-impact decisions and novel threats still require human judgment and oversight. The playbook described enriches and triages alerts to support analyst decision-making rather than replacing analysts entirely. Best practices recommend maintaining human oversight for critical response actions.

C is incorrect: SOAR platforms are not designed to replace SIEM systems for centralized log storage and analysis. SOAR orchestrates actions and automates workflows across existing tools, while the SIEM handles data collection, correlation, and detection. The playbook described queries the SIEM and other tools for enrichment rather than centralizing their data stores.

D is correct: Reducing mean time to respond through automated enrichment and triage is correct. The playbook gathers contextual data from multiple sources and uses AI to assign severity, which automates the labor-intensive enrichment and triage steps analysts would otherwise perform manually. SOAR playbooks that automate multi-source data gathering and AI-driven severity scoring significantly accelerate incident response while providing richer context for analyst decision making.

A is incorrect: Inspecting inbound network traffic for known exploit payload signatures focuses on network-layer detection rather than endpoint behavior analysis. While useful for detecting certain network-based attacks using known exploits, this approach does not address the core challenge of identifying malware that continuously mutates its code structure on the endpoint to evade file-based and host-based detection. Reference: https://www.microsoft.com/en-us/security/business/security-101/what-is-edr-endpoint-detection-response

B is correct: Monitoring runtime behaviors and system call patterns is the most effective EDR capability for detecting polymorphic malware that restructures its code on each execution. Since the malware changes its binary structure to evade signature-based detection, static analysis methods will fail. EDR solutions using behavioral analysis can identify the malware by observing consistent malicious activities such as suspicious API call sequences, process injection, and abnormal system interactions that remain constant despite code changes. Reference: https://attack.mitre.org/techniques/T1027/014/

C is incorrect: Scanning executable files for static byte-sequence signature matches is another form of signature-based detection that polymorphic malware specifically defeats. By altering its code structure to produce different byte sequences each time it executes, the malware ensures that no single static signature can consistently match across its variants. This is the core evasion technique of polymorphic malware. Reference: https://attack.mitre.org/techniques/T1027/014/

D is incorrect: Comparing file hashes against known malware indicator databases is a signature-based detection method. Because polymorphic malware changes its binary code structure with each execution, each variant produces a unique hash value. This means hash-based comparisons cannot reliably identify new variants that have never been cataloged, making this approach fundamentally ineffective against the described threat. Reference: https://attack.mitre.org/techniques/T1027/014/

A is correct: TLS handshake fingerprinting and connection metadata analysis is correct. AI-powered NDR platforms detect C2 communications in encrypted traffic by analyzing TLS Client Hello characteristics such as JA3 and JA4 fingerprints, certificate attributes, session timing, and connection metadata. This approach identifies malicious patterns based on how a client establishes communication rather than what is communicated inside the encrypted tunnel.

B is incorrect: Deep packet payload inspection with content pattern matching requires access to unencrypted packet contents to compare against known patterns. Since the traffic is TLS-encrypted and the organization cannot perform active decryption, the NDR platform cannot directly inspect payload data. Content pattern matching is effective for cleartext traffic but cannot be applied to encrypted communications without a decryption mechanism.

C is incorrect: Automated certificate revocation list generation is a certificate lifecycle management function rather than a network threat detection technique. While validating certificates against revocation lists can identify use of compromised certificates, generating and distributing CRLs does not enable an NDR platform to detect patterns of malicious C2 communications hidden within encrypted network sessions.

D is incorrect: Extracting symmetric keys from TLS session negotiation constitutes a form of decryption, which contradicts the stated requirement. Modern TLS implementations that support perfect forward secrecy prevent passive key extraction from observed handshakes. This approach is impractical against current TLS versions and would not allow the organization to detect threats without active decryption capabilities.

A is correct: Machine learning models that forecast emerging threats and dynamically update detection rules represent the proactive defense approach. Proactive defense uses AI to analyze threat trends, predict likely attack vectors, and preemptively adjust security controls before threats materialize. This shifts the organization from waiting for attacks to occur and then responding to anticipating and preparing for threats in advance using AI-driven predictive capabilities.

B is incorrect: Scheduled weekly imports of threat indicators represent a manual, periodic update process that remains fundamentally reactive. This approach relies on known indicators that have already been published, meaning threats must be discovered and documented elsewhere before the organization can defend against them. Proactive defense requires AI-driven capabilities that predict and adapt to emerging threats continuously and automatically.

C is incorrect: Generating executive summaries of past security incidents is a reporting and communication function, not a proactive defense capability. While summarizing historical incidents can inform strategic planning, it does not directly shift the organization to a proactive defense posture. Proactive defense requires forward-looking AI capabilities that predict emerging threats and automatically adjust security controls in anticipation.

D is incorrect: Automated logging for post-incident forensic analysis is a reactive capability, not a proactive one. While comprehensive logging is important for investigating incidents after they occur, it does not help the organization anticipate or prevent threats before they materialize. Proactive defense requires predictive capabilities that adjust defenses in advance rather than tools focused solely on after-the-fact investigation.

A is incorrect: Log aggregation is the process of collecting, centralizing, and indexing data from multiple sources for storage and later analysis. While a TIP does aggregate indicators from multiple feeds, the scenario emphasizes AI-driven addition of contextual metadata such as geolocation and threat actor attribution to each indicator. Log aggregation focuses on data collection and storage, not on enhancing the contextual quality of threat intelligence through enrichment.

B is incorrect: Automated signature generation involves creating detection patterns or rules that security tools use to identify known threats in network traffic. While AI can assist with rule creation, the scenario specifically describes adding contextual information like geolocation and attribution to raw indicators. Enrichment enhances the quality and context of threat intelligence, whereas signature generation focuses on creating detection logic for network security devices.

C is correct: Automated indicator enrichment is the correct answer. AI-powered indicator enrichment uses machine learning and automated data analytics to add contextual metadata to raw indicators of compromise (IOCs). By appending geolocation, threat actor attribution, and confidence ratings, AI transforms low-value raw data into enriched, actionable intelligence that analysts can quickly assess. This is a core AI use case in threat intelligence platforms, enabling faster triage and more informed decision-making without requiring manual lookups for each indicator.

D is incorrect: Response orchestration refers to the automated coordination and execution of incident response actions across multiple security tools, such as isolating compromised hosts or blocking malicious IPs. While response orchestration is an important AI-driven security capability, it focuses on taking action against identified threats rather than adding contextual metadata to raw indicators. The scenario describes enriching intelligence data with context, not automating response workflows.

A is incorrect: Setting privileged: true gives the container full access to host resources, effectively disabling all container isolation mechanisms. This configuration combined with a writable root filesystem creates maximum exposure to privilege escalation and host compromise. For AI training workloads, privileged mode is unnecessary and introduces severe security risks that directly contradict the goal of reducing privilege escalation. Reference: https://kubernetes.io/docs/tasks/configure-pod-container/security-context/

B is correct: Setting runAsNonRoot: true and allowPrivilegeEscalation: false is the best configuration for reducing privilege escalation risk. runAsNonRoot ensures the container process does not run as the root user, while allowPrivilegeEscalation: false prevents child processes from gaining more privileges than the parent process. These settings significantly limit the attack surface while allowing AI training containers to operate with necessary user-level permissions. Reference: https://kubernetes.io/docs/tasks/configure-pod-container/security-context/

C is incorrect: Setting hostNetwork: true exposes the container to the host's network namespace, and hostPID: true allows the container to see and interact with all processes on the host. These settings break pod isolation boundaries and introduce critical privilege escalation vectors. AI training workloads should operate within isolated pod network and PID namespaces to maintain security. Reference: https://kubernetes.io/docs/tasks/configure-pod-container/security-context/

D is incorrect: Setting runAsUser: 0 runs the container as the root user, and adding NET_ADMIN allows the container to modify network configurations on the host. Both settings increase the attack surface and directly enable privilege escalation paths. AI training workloads do not require root access or network administration capabilities, making this configuration both insecure and unnecessary. Reference: https://kubernetes.io/docs/tasks/configure-pod-container/security-context/

A is incorrect: While zero-day exploits targeting antivirus engines do exist, this describes a vulnerability exploitation attack against specific software rather than the inherent limitation of signature-based detection against polymorphic malware. The scenario describes repeated, consistent failures despite regular signature updates, which indicates a fundamental architectural limitation of the detection approach rather than an engine-level exploit. Reference: https://attack.mitre.org/techniques/T1027/014/

B is correct: AI-enhanced polymorphic malware uses mutation engines to automatically alter its code with every execution, generating new unique binary patterns each time. Since signature-based detection relies on matching known file patterns in a database, it fundamentally cannot keep pace with malware that produces a different code structure upon each iteration. As documented by MITRE ATT&CK, polymorphic code 'is mutated into a different version of itself that achieves the same purpose or objective as the original,' which is the core reason signature matching fails against this class of threat. Reference: https://attack.mitre.org/techniques/T1027/014/

C is incorrect: While some malware can interfere with security software updates, this describes an update disruption attack rather than the fundamental reason signature-based detection fails against polymorphic malware. Even with the most current and complete signature database, AI-enhanced polymorphic malware that generates unique code variants with each execution would still evade detection because no pre-existing matching signature can account for the newly generated variant. Reference: https://attack.mitre.org/techniques/T1027/014/

D is incorrect: Encrypting system calls relates to evasion of behavioral monitoring, not signature-based detection. Signature-based detection primarily matches known file patterns and code sequences rather than runtime system calls. Additionally, system calls are processed by the operating system kernel and cannot be encrypted from the perspective of endpoint security monitoring tools. The scenario specifically identifies signature-based detection as the failing mechanism. Reference: https://attack.mitre.org/techniques/T1027/014/

A is incorrect: Data poisoning attacks target the training phase by injecting malicious data to corrupt the model's learned behavior. The scenario describes activity occurring at the inference layer against a deployed model through API queries, not manipulation of training data. Data poisoning cannot be conducted through inference-time queries to a production endpoint, making this an incorrect characterization of the observed behavior. Reference: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2025.pdf

B is correct: Membership inference attack is correct. Membership inference attacks exploit differences in a model's confidence scores between data records that were part of the training set (members) and those that were not (non-members). By systematically comparing prediction confidence for known records against unknown records, the attacker is attempting to determine whether specific individuals' data was used to train the model. The runtime monitoring system correctly flagged this pattern because the query behavior, which involves repeated confidence score comparisons, is a hallmark of membership inference techniques. Reference: https://owasp.org/www-project-machine-learning-security-top-10/docs/ML04_2023-Membership_Inference_Attack

C is incorrect: Model inversion attacks aim to reconstruct or infer sensitive attributes of training data by reverse-engineering the model's outputs, typically by submitting crafted inputs and analyzing detailed responses. However, the scenario specifically describes comparing confidence scores for known versus unknown records, which is characteristic of membership inference rather than reconstruction of training data features. Model inversion focuses on recovering data values, not confirming membership. Reference: https://owasp.org/www-project-machine-learning-security-top-10/docs/ML03_2023-Model_Inversion_Attack

D is incorrect: Adversarial evasion attacks involve crafting inputs that cause the model to misclassify or produce incorrect predictions while appearing normal to human observers. The scenario describes querying behavior focused on comparing confidence scores between known and unknown records, not on manipulating inputs to evade detection. Evasion attacks target model accuracy, whereas the described behavior targets training data privacy. Reference: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2023.pdf

A is incorrect: AI-driven NDR solutions do not automatically decrypt all TLS sessions. While some NDR platforms can integrate with TLS decryption infrastructure, automatic universal decryption is not a standard NDR capability and raises significant privacy and compliance concerns. The primary advantage of AI-driven NDR is behavioral analysis for novel threat detection, not comprehensive traffic decryption functionality.

B is incorrect: AI-driven NDR solutions do not replace network sensors with endpoint agents. NDR specifically relies on network-level telemetry including packet capture, flow data, and traffic mirroring from network sensors and taps deployed across the infrastructure. Endpoint detection and response uses agent-based collection, but NDR and EDR serve complementary roles with distinct data sources.

C is correct: Detecting novel threats by establishing behavioral baselines is the primary advantage of AI-driven NDR over traditional signature-based IDS. AI-driven NDR solutions use machine learning to establish a baseline of normal network behavior and then detect anomalies that deviate from this baseline. This enables detection of previously unknown threats, insider threats, and zero-day attacks that lack existing signatures, which a traditional IDS would miss entirely.

D is incorrect: AI-driven NDR solutions do not eliminate all false positives. While behavioral AI analysis can reduce false positive rates compared to rigid signature matching, machine learning models can still generate alerts requiring expert interpretation and validation. No detection system can fully eliminate false positives. The primary advantage of AI-driven NDR is novel threat detection capability, not false positive elimination.