Key Takeaway: AI security operations solutions are essential for reducing the fatigue of security analysts caused by a surge in alerts and false positives, and for addressing the shortage of analytical personnel. AI-driven analysis comprehensively understands the context of alerts to minimize false positives, accurately prioritizes threats through LLM triage, and significantly enhances the speed of threat containment with SOAR-based automated response.
As the security threat landscape becomes increasingly complex and sophisticated, it is challenging to effectively respond to all threats using only traditional rule-based security operations. Distinguishing between actual threats and false positives amidst the continuous deluge of alerts from numerous systems places a significant burden on security operations teams. Consequently, security analysts experience fatigue due to alert overload and false positive processing, and the scarcity of skilled analytical personnel further exacerbates these issues. To address these challenges, the adoption of AI security operations solutions is becoming crucial. AI security operations go beyond mere threat detection; they are a core technology that enhances the overall efficiency of security operations by streamlining the alert triage process and supporting automated responses.
Hierarchical Structure of AI Security Operations Systems
| Layer | Role | Key Technologies |
|---|---|---|
| Collection & Normalization | Collecting and standardizing security data from various sources (logs, traffic, etc.) | Log Collector, NDR sensor, EDR agent, SIEM |
| Rule-based Detection | Identifying known attack patterns and threats | Sigma rule engine, MITRE ATT&CK mapping |
| Correlation & Behavior Analysis | Analyzing correlations and detecting abnormal behavior | Multi-stage attack correlation analysis, UEBA (User and Entity Behavior Analytics) |
| LLM Triage | Analyzing alert context, determining severity, summarizing, and prioritizing | LLM (Large Language Model) |
| Automated Response | Automatically controlling threats according to predefined procedures | SOAR Playbook |
| Human Review (HITL) | Validating AI judgments and making final decisions on complex threats | Security Analyst (Human-In-The-Loop) |
Principles of Minimizing False Positives through LLM Triage
Rule-based detection systems can generate numerous alerts, a significant portion of which often turn out to be false positives. LLMs play a crucial role in efficiently triaging these alerts and reducing false positives. Beyond simple alert messages, LLMs comprehensively analyze various structured and unstructured data, including asset information related to the alert, user behavior history, past threat history, and latest threat intelligence, to gain a deep understanding of the threat context.
Based on the collected contextual information, LLMs determine the actual severity of an alert and assess the likelihood of it being a false positive. For instance, if a specific administrator logs into an uncommonly used system over the weekend and executes a script, the LLM analyzes this to determine whether it is a normal workflow or an anomalous activity. LLMs also clearly record their reasoning, supporting security analysts in reviewing and understanding their judgments. If an LLM finds it difficult to clearly determine if an alert is a threat or if the uncertainty is high, the alert is immediately escalated to a human analyst for a final decision.
Virtual Example: False Positive Handling by LLM Triage at Each Step
The following is a hypothetical example of an LLM processing an alert for an unusual login attempt.
- Alert Generation: A login success alert is generated for an account not typically used on a specific server.
- LLM Context Collection: The LLM collects the following information:
- Asset Information: The server is a development environment and is used for a specific project.
- User Information: The account belongs to developer 'Kim Cheol-su,' who typically manages systems.
- Historical Data: 'Kim Cheol-su' rarely logs into this specific server. However, there are a few instances of remote access to other development servers during similar timeframes over the past month.
- Threat Intelligence: The IP address used for this login is not present in public threat intelligence (IOC) lists.
- LLM Analysis and Judgment: The LLM comprehensively analyzes the collected context.
- The fact that the account owner is a developer and the server is a development environment increases the likelihood of legitimate activity.
- Although login history for this server is infrequent, a similar pattern of access to other development servers suggests a possible new work pattern.
- The absence of the IP in threat intelligence as malicious further supports the possibility of a false positive.
- The LLM classifies this alert as a 'low-severity potential false positive' and records 'possibility of a new developer work pattern' as the basis for its judgment.
- Human Review and Feedback: If the LLM's judgment is deemed uncertain, the alert is escalated to a human analyst. The analyst reviews the contextual information and reasoning provided by the LLM and directly verifies with developer 'Kim Cheol-su.' Upon confirming that 'Kim Cheol-su' recently needed to access that server for a new project, the analyst finalizes the alert as a 'false positive' and provides feedback to the LLM as training data to tune it to recognize similar patterns as false positives in the future.
Through this process, LLMs continuously learn and evolve, effectively reducing the false positive rate of security operations systems and contributing to increased focus on actual threats.
Determining the Level of Automated Response: How Much to Delegate?
In AI security operations systems, SOAR-based automated response is a core capability for rapidly containing threats. However, applying the same level of automation to all threats is not appropriate. The level of automated response must be carefully determined, considering the severity of the threat, its potential impact, and the likelihood of false positives.
| Response Type | Description | Examples |
|---|---|---|
| Automatic Execution | Executes responses immediately without human intervention when predefined conditions are met. Applied to severe threats where the risk of false positives is low and immediate action is required. |
|
| Execution After Approval | The AI system detects threats and proposes response actions, but actual execution proceeds only after final approval by a human analyst. Suitable for responses where there is a possibility of false positives or where the impact on systems could be significant. |
|
| Human-Exclusive | The AI system detects threats and provides contextual information, but response execution is carried out solely by human analyst judgment and manual operation. Applied in cases requiring complex decisions or where an incorrect response could lead to catastrophic results. |
|
These flexible response policies maximize automation efficiency while effectively managing potential risks that may arise.
Key Considerations for Adopting AI Security Operations
To successfully adopt and operate an AI security operations solution, several key precautions must be carefully considered.
1. Ensuring Transparency of AI Decision-Making
AI, especially LLM-based systems, utilize complex internal logic during threat detection and classification. This can lead to a lack of transparency in AI decisions, making it difficult for security analysts to accurately understand threats and make final judgments. Therefore, an AI system must clearly log the 'basis for its judgment' when classifying an alert or proposing a specific response, and be able to present it to analysts in a visualized format. This transparency is essential for analyzing the root cause of false positives and improving AI models.
2. LLM Security and Prompt Injection Defense
When utilizing LLMs for security operations, the security vulnerabilities of the LLM model itself must be considered. Attacks such as 'prompt injection,' in particular, can cause LLMs to malfunction or be used to exfiltrate sensitive internal information. AI security operations solutions must implement robust defense mechanisms against such LLM-based attacks. For instance, they should include strict validation of user input prompts, anomaly detection, and guardrail functions to verify the LLM's output.
3. Compliance with Data Processing and Storage Location Regulations
Security operations systems collect and analyze sensitive log data from organizations. Therefore, data processing and storage locations must be strictly managed in accordance with data privacy regulations (e.g., GDPR, domestic Personal Information Protection Act, etc.) and internal corporate policies. When adopting cloud-based AI security operations solutions, it is crucial to thoroughly verify where the data is stored and processed (which regional data centers), as well as the data encryption and access control policies in place. Compliance is a critical prerequisite for the adoption of AI security operations.
By meticulously reviewing and preparing for these considerations, organizations can maximize the effectiveness of AI security operations systems and minimize potential risks.
AI Security Operations Structure of KYRA MDR
KYRA MDR is SeekersLab's AI-powered MDR (Managed Detection & Response) platform, providing intelligent security operations services. KYRA MDR integrates SIEM, NDR, EASM, and SOAR into a single platform to deliver 24/7 security monitoring.
KYRA MDR's AI security operations function with the following structure:
- Data Collection: Various security data is collected through Log Collectors that gather firewall, server, and cloud logs via syslog or files, Windows EDR agents, NDR sensors that collect flow metadata from TAP/SPAN mirrored traffic, and External Attack Surface Management (EASM).
- Detection: Threats are detected based on the Sigma rule engine (over 3,200 rules) and MITRE ATT&CK mapping (14 tactics, over 200 techniques). Additionally, behavior analysis and multi-stage attack correlation analysis identify complex threats, and threat intelligence IOC (Indicator of Compromise) enrichment enhances detection accuracy.
- AI Analysis (LLM Triage): LLM-based alert classification (triage) and analytical summaries are provided to reduce the workload of security analysts. Rules are continuously tuned through false positive learning, and critical judgments maintain a Human-In-The-Loop (HITL) structure reviewed by human analysts to compensate for AI limitations.
- Response: Automated response capabilities are provided based on SOAR playbooks. This includes executing blocking actions such as host isolation, supporting efficient investigation workflows, and assisting in the generation of executive reports.
- Compliance: An ISMS-P compliance dashboard is provided, supporting major security framework templates such as ISO 27001, SOC 2, and PCI-DSS, along with audit logs.
Through this AI-driven integrated security operations structure, KYRA MDR strengthens an organization's security threat response capabilities.
FAQ: AI Security Operations Solutions
Q1: How do AI security operations solutions differ from traditional SIEMs?
A1: AI security operations solutions encompass the log collection and rule-based detection capabilities of traditional SIEM (Security Information and Event Management) systems, while integrating advanced LLM (Large Language Model)-based analysis and SOAR (Security Orchestration, Automation and Response)-based automated response functionalities. This approach reduces false positives, enhances alert triage efficiency, and significantly improves threat response speed.
Q2: What is the core principle by which AI reduces false positives?
A2: AI, particularly LLM-based analysis, goes beyond merely checking for rule matches. It comprehensively understands and interprets the context in which an alert was generated (including assets, users, historical data, and threat intelligence). This enables it to distinguish between actions that appear similar to legitimate business activities but are malicious, or to identify activities that seem abnormal but are in fact legitimate, thereby effectively reducing false positives.
Q3: Is it safe for AI to automatically respond to all threats?
A3: It is not appropriate for AI to fully automatically respond to all threats. AI security operations systems can flexibly set response levels, such as 'Automatic Execution,' 'Execution After Approval,' and 'Human-Exclusive,' depending on the severity and potential impact of the threat. For complex or high-impact threats, safety is ensured through a Human-In-The-Loop (HITL) approach, which involves final judgment and review by human analysts.
Q4: How should data privacy issues be managed when adopting AI security operations?
A4: As AI security operations systems process sensitive log data, compliance with regulations regarding data processing and storage locations (e.g., ISMS-P, etc.) is critically important. Prior to solution adoption, organizations must thoroughly verify data encryption, access control, and data sovereignty policies, and clearly ascertain which regional data centers store and process the data.
Q5: Could LLM-based AI security operations systems be vulnerable to prompt injection attacks?
A5: Systems utilizing LLMs must incorporate defense mechanisms against attacks such as prompt injection. AI security operations solutions protect LLM models from these attacks and prevent misuse through their own security features, including validation of user input prompts, anomaly detection, and LLM output verification.
Related Articles
- What is MDR? Differences from Security Operations and Criteria for Choosing MDR Services
- EDR, XDR, MDR Differences: A Comparison of Security Operations Approaches
- How Small and Medium-sized Enterprises Can Begin Security Operations and Cost Structure
- What is an NDR Security Solution? Differences Between NDR, EDR, and SIEM
- Seekurity SOAR
- Seekurity SIEM
Get Started with Security Operations using KYRA MDR
If organizations wish to strengthen their security capabilities through AI-driven intelligent security operations and automated response, they are encouraged to experience KYRA MDR. Explore KYRA MDR's detection and alerting features directly with a free plan and establish an optimal security operations strategy tailored to the organizational environment.







