Automated Triage of Cybersecurity Alerts and Threat Intelligence

106 concepts98 papers2 proposals

Security Operations Centers (SOCs) face a fundamental scalability crisis: the volume of security data grows exponentially, but human cognitive capacity remains fixed. Modern enterprise networks generate millions of daily events, forcing defenders to navigate a dual deluge of information. Internally, automated anomaly and rule-based detectors flood dashboards with low-level alerts, a high percentage of which are false positives. Externally, an overwhelming stream of Cyber Threat Intelligence (CTI)—such as malware reports, security bulletins, and open-source intelligence (OSINT)—is published daily in unstructured, human-readable text.

When human analysts are forced to manually investigate thousands of isolated alerts or parse dense technical blogs to update defense rules, they experience cognitive overload and alert fatigue. This operational exhaustion creates a dangerous window of opportunity: critical, multi-stage attacks can easily blend into the background noise, remaining undetected for months.

To solve this problem, a major research direction focuses on automated triage. By leveraging graph theory, natural language processing (NLP), large language models (LLMs), and game theory, this paradigm aims to compress raw data into high-level, actionable insights, shifting the analyst's role from manual data processing to strategic decision-making.


Compressing Alert Streams via Contextual Correlation

The primary driver of alert fatigue is the atomicity of traditional detection systems. A single network event—such as a port scan, a failed login, or a modified system file—triggers an isolated alarm. However, sophisticated cyber attacks are rarely single events; they are multi-stage campaigns.

Instead of forcing analysts to manually connect these dots, automated triage frameworks use alert correlation to reconstruct the underlying attack graph. The intuition is simple: by grouping alerts that share temporal, spatial, or behavioral attributes, thousands of low-level notifications can be compressed into a handful of high-level security narratives.

[Isolated Alert: Port Scan] ---\
[Isolated Alert: Failed Login] ---> [Correlation Engine] ---> [Single Actionable Attack Narrative]
[Isolated Alert: File Modified] ---/

Researchers have developed several algorithmic approaches to achieve this compression:

  • Graph and Tree-Based Aggregation: Sadoddin et al. demonstrated that massive streams of intrusion alerts can be correlated in real-time by combining graph-based aggregation with a Frequent Structured Pattern Tree (FSP_Tree). This structure automatically identifies recurring multi-step attack patterns across large networks. Similarly, Miao et al. utilized multi-similarity factor aggregation and suffix tree modeling to eliminate alert redundancy and isolate genuine threat sequences.
  • Sequence Clustering: Rather than relying on rigid rules, Spahn introduced a semi-supervised framework (DEEPCASE) that analyzes the context of security events. By clustering similar sequences of events together, the system allows analysts to review entire behavioral patterns at once, reducing manual workloads by over 90%.
  • Filtering Benign Shifts: A major source of false positives is the inability of detectors to distinguish between malicious activity and harmless system changes (like software updates). To address this, Feyereisl et al. proposed a bio-inspired approach modeled after the human immune system's Toll-Like Receptors (TLRs), integrating "danger sensors" into self-organizing maps to provide the context needed to separate benign behavioral shifts from actual threats.

By filtering out redundant noise and structural bottlenecks—sometimes using mechanisms like Token Bucket Filters to prevent log-flooding attacks from crashing the security infrastructure itself (Tedesco et al.)—these systems ensure that human attention is reserved for complex, multi-stage threats.


Operationalizing Unstructured Threat Intelligence

While internal alerts represent what is happening inside a network, external threat intelligence tells defenders what could happen. However, valuable threat data is frequently locked away in unstructured text formats like blog posts, vendor reports, and security bulletins (Bahrami et al., Farhang et al.). Manually reading these documents to extract Indicators of Compromise (IOCs)—such as malicious URLs, IP addresses, or file hashes—is a slow, error-prone process that cannot scale.

Automated triage addresses this by using advanced NLP techniques to transform raw text into machine-readable, structured knowledge.

Sequence Labeling and Semantic Extraction

Early automated systems relied on hand-crafted rules, but modern approaches use neural sequence labeling to extract semantic entities directly from messy text. For example, Long et al. developed an end-to-end model combining multi-head self-attention with contextual features to automatically extract IOCs without requiring deep domain expertise. To capture more complex relationships, Ma et al. combined deep learning (BiLSTM-CNN-CRF) with traditional linguistic features to extract specific malware actions and entities from dense technical reports.

Knowledge Graphs and Semantic Ranking

Extracting isolated indicators is only the first step; understanding the relationships between threats requires mapping them to standardized security frameworks.

[Unstructured Report Text] 
       │
       ▼ (LLM / Sequence Labeling)
[Threat Triples: Attacker -> Uses -> Malware] 
       │
       ▼ (Semantic Ranking)
[Structured Knowledge Graph / MITRE ATT&CK Mapping]

Lekssays et al. shifted the problem from traditional text classification to a semantic ranking approach, significantly improving the accuracy of mapping unstructured text to specific MITRE ATT&CK techniques.

To build fully queryable repositories of threat data, Fieblinger et al. demonstrated that fine-tuned, open-source LLMs can effectively extract threat "triples" (subject-predicate-object relations) from blog posts to synthesize structured Knowledge Graphs. This automated synthesis allows defenders to ingest informal threat intelligence from developer forums and social media in real-time, bypassing slow, centralized databases (Liu et al.).

Bridging Intelligence to Action

The ultimate goal of extracting threat intelligence is to use it for active defense. Frameworks like THREATRAPTOR, developed by Gao et al., close the loop by using an NLP pipeline to parse unstructured threat reports and automatically synthesize precise search queries. These queries can be executed directly against massive system logs to hunt for threats, turning a hours-long manual translation process into an automated, near-instantaneous operation.


Strategic Triage and Event-Centric Explainability

Even with advanced correlation and extraction, some alerts will always require human verification. The final pillar of automated triage focuses on optimizing this human-in-the-loop interaction by answering two questions: Which alerts should the analyst check first? and Why did the system flag them?

Game-Theoretic Resource Allocation

When the volume of alerts exceeds human capacity, defenders cannot simply review alerts in chronological order. Strategic adversaries can exploit this limitation by generating a high volume of false positives to intentionally hide their malicious actions in the noise.

To counter this, researchers apply game theory to model alert triage as a strategic interaction. Schlenker extended Threat Screening Games (TSGs) to optimize how limited human investigation time is allocated under uncertainty and adversarial deception. Similarly, Yan et al. modeled database auditing as a Stackelberg game, designing optimization algorithms that prioritize alerts to maximize both the detection of active threats and the deterrence of future attacks.

Shifting to Event-Centric Narratives

Traditional machine learning models often present "feature-level" explanations (e.g., flagging an alert because a specific packet length or mathematical weight crossed a threshold). These generic, internal explanations often overwhelm analysts, contributing further to operational fatigue (Pant et al.).

To build trust and accelerate triage, recent architectures focus on event-centric explainability. Instead of showing raw data features, these systems generate human-readable justifications grounded in verifiable evidence:

  • Graph Explanations: For security models built on Graph Neural Networks (GNNs), He et al. developed the ILLUMINATI framework, which identifies the exact subgraphs, edges, and node attributes driving a prediction, giving analysts a visual, evidence-based justification for why a vulnerability was flagged.
  • Multi-Agent Investigations: The authors of (EC)2 (arXiv 2026) introduced a multi-agent LLM framework that performs automated, hypothesis-driven forensic investigations. Instead of forcing an analyst to figure out why an anomaly occurred, the agents investigate the network evidence and present a synthesized narrative explaining whether the event represents a misconfiguration or a multi-stage attack.
  • Reasoning-Enabled Triage: By combining Chain-of-Thought (CoT) reasoning with confidence calibration, the authors of a recent 30B parameter cybersecurity model (arXiv 2026) demonstrated that reasoning-focused language models can achieve massive improvements in benign alert recall, effectively acting as an automated first-tier analyst.

By combining strategic prioritization with clear, narrative-driven explanations, automated triage frameworks transform raw, overwhelming security data into a manageable stream of high-fidelity, actionable insights.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_alert_fatigue_432a8b48,
  title = {alert fatigue},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/alert-fatigue/rdr_432a8b48df1a4317b36a6c6cd42862e1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_alert_fatigue_432a8b48_v1,
  title = {alert fatigue},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/alert-fatigue/rdr_432a8b48df1a4317b36a6c6cd42862e1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake