Papers
Topics
Authors
Recent
Search
2000 character limit reached

AgentSOC: A Multi-Layer Agentic AI Framework for Security Operations Automation

Published 22 Apr 2026 in cs.CR, cs.AI, and cs.CL | (2604.20134v1)

Abstract: Security Operations Centers (SOCs) increasingly encounter difficulties in correlating heterogeneous alerts, interpreting multi-stage attack progressions, and selecting safe and effective response actions. This study introduces AgentSOC, a multi-layered agentic AI framework that enhances SOC automation by integrating perception, anticipatory reasoning, and risk-based action planning. The proposed architecture consolidates several layers of abstraction to provide a single operational loop to support normalizing alerts, enriching context, generating hypotheses, validating structural feasibility, and executing policy-compliant responses. Conceptually evaluated within a large enterprise environment, AgentSOC improves triage consistency, anticipates attackers' intentions, and provides recommended containment options that are both operationally feasible and well-balanced between security efficacy and operational impact. The results suggest that hybrid agentic reasoning has the potential to serve as a foundation for developing adaptive, safer SOC automation in large enterprises. Additionally, a minimal Proof-Of-Concept (POC) demonstration using LANL authentication data demonstrated the feasibility of the proposed architecture.

Summary

  • The paper introduces AgentSOC, a multi-layer AI framework that integrates perception, agentic reasoning, and adaptive action to automate and enhance SOC operations.
  • It employs LLM-driven hypothesis generation combined with graph-based structural validation to prioritize risk-constrained, policy-compliant incident responses.
  • Empirical results show sub-second incident processing and significant manual triage reduction, highlighting its potential to improve response speed and operational safety.

AgentSOC: A Multi-Layer Agentic AI Framework for Security Operations Automation

Introduction and Problem Statement

The AgentSOC framework addresses longstanding limitations in Security Operations Centers (SOCs) where automation remains bounded by fragmented toolchains, static rule-based reasoning, and lack of contextual alignment between response recommendations and enterprise constraints. Traditional SIEM/SOAR/XDR platforms automate low-level tasks, but alert fatigue, slow triage, and inconsistent containment responses persist due to excessive false positives and reliance on manual correlation. LLM-driven copilots offer enrichment and suggestion capability but remain advisory, often hallucinating infeasible response actions and lacking grounding in enterprise topology, identity, and privileged access graphs. Despite widespread adoption of knowledge bases such as MITRE ATT&CK, their application remains largely reactive and disconnected from anticipatory, policy-compliant automation.

Architectural Overview

AgentSOC is architected as a multi-layer, agentic AI system with an integrated operational loop encompassing perception, agentic reasoning, and action. The design is characterized by four primary subsystems:

  • Perception Layer: Normalizes heterogeneous alerts, injects contextual enterprise metadata (e.g., asset criticality, privilege levels, topological data), and suppresses noise through clustering and deduplication.
  • Agentic Reasoning Layer: Employs an LLM-based Narrative Counterfactual Engine (NCE) to generate multiple hypothesis attack progressions, a Structural Simulation Engine (SSE) that validates scenario feasibility against real network/identity graphs, and a Risk Scoring and Evaluation Module (RSEM) that computes composite risk metrics for candidate actions.
  • Action and Playbook Layer: Synthesizes adaptive playbooks by translating feasible hypotheses into executable response workflows, integrates policy guardrails for containment options, and connects to response interfaces for (optionally) autonomous execution.
  • Internal Knowledge Store: Consolidates static and dynamic context, including asset inventories, privilege relations, operational state, and executed action outcomes to close the feedback loop.

The framework is fully closed-loop: detection triggers a sense-reason-act cycle, with post-action validation and knowledge updating to ensure adaptive behavior over time.

Mechanistic Innovations

AgentSOC introduces several innovations otherwise absent from existing SOC automation paradigms:

  • LLM-driven Multi-Hypothesis Generation: Rather than pursuing a single rule-matched interpretation, the NCE module uses LLM reasoning aligned with MITRE ATT&CK TTPs to forecast alternative attacker intentions (e.g., credential misuse, lateral movement, privilege escalation) per incident record.
  • Structural Feasibility Validation: The SSE traverses enterprise-specific identity and network graphs, ensuring that only structurally consistent attack paths and policy-compliant mitigations are retained for further analysis. This approach mitigates frequent pitfalls of LLM-driven 'hallucination'.
  • Risk-Constrained Action Ranking: Actions are objectively scored by RSEM via a tunable composite metric that balances technical containment value against business impact and execution risk, enabling context-sensitive, policy-compliant response selection.
  • Closed-Loop Adaptation: Real-time monitoring and feedback from operational outcomes feed into the knowledge store, supporting continuous calibration of response strategies and facilitating human-in-the-loop oversight where required.

Numerical Results and Empirical Observations

The proof-of-concept demonstration, built atop 5,000 events from the LANL authentication dataset and a synthetic 50-node enterprise topology, yielded several notable empirical results:

  • Sub-second End-to-End Latency: The average processing time per incident was 506 ms—of which LLM hypothesis generation (NCE) dominated total latency—demonstrating feasibility for integration into real-world SOC environments where incident cycles are measured in minutes.
  • Reduction in Manual Triage: Perception and contextual enrichment reduced triage time from minutes to milliseconds via automated correlation and deduplication, collapsing multi-alert scenarios into single enriched incident objects.
  • Anticipatory Hypothesis Generation: NCE produced multiple plausible attack trajectories with associated confidence metrics; graph-based SSE validation filtered infeasible or policy-violating hypotheses, such as those lacking structural support or network reachability.
  • Risk-Informed Action Selection: RSEM ranked containment strategies based on containment effectiveness and operational disruption, with policies enforcing that high-impact actions (e.g., session revocation, privilege restriction) are routed for analyst approval, while low-risk automated actions (e.g., monitoring) can be autonomously executed.

Of particular note, host isolation was most frequently recommended for high-confidence suspicious events due to its low operational impact and high containment score. The design maintained auditability and rollback capability via execution interfaces and dry-run modes, critical for enterprise trust and safety.

Comparative Analysis and Practical Implications

In direct comparison to existing SIEM, SOAR, and LLM copilot systems, AgentSOC advances the state of the art in several key dimensions: reliable synthesis of cross-domain alerts, structural grounding of reasoning, and explicit business impact awareness. Unlike rigid playbook-driven automation, AgentSOC’s hybrid architecture supports context-adaptive, explainable decision-making, and selective autonomy. Compared to LLM copilots, AgentSOC’s structural validation and policy guardrails significantly reduce the risk of infeasible or unsafe recommendations.

The practical implications are considerable: by minimizing manual intervention for the majority of routine and moderately complex cases, AgentSOC has direct potential to alleviate SOC workforce shortages, decrease mean time-to-detect (MTTD) and mean time-to-respond (MTTR), and limit adverse operational impact due to indiscriminate or uncoordinated mitigation action.

Limitations and Future Directions

Several limitations are acknowledged:

  • Evaluation used static dataset subsets and limited network size; real-time adaptability at enterprise scale was not fully validated.
  • MITRE ATT&CK mappings and risk heuristics presently do not account for zero-day tactics or advanced adversarial adaptation.
  • Response action execution primarily operated in dry-run simulation, not closed-loop with production controls and live telemetry.
  • LLM processing latency is tolerable in current SOC contexts but may constrain throughput at extreme event volumes unless optimized.
  • Quantification of business impact remains heuristic-driven rather than empirically calibrated.

Future directions include live telemetry integration, expanding to multi-cloud and hybrid IT/OT environments, reinforcement learning-informed risk scoring, deeper integration with proactive threat intelligence feeds, and robust production-grade SOAR connectors with improved rollback and human-in-the-loop escalation support.

Conclusion

AgentSOC presents a systematic, agentic AI framework for SOC automation that holistically unifies enriched perception, counterfactual hypothesis generation, structural feasibility validation, and risk-aware action planning. Empirical evaluation demonstrates sub-second closed-loop operation, substantial reduction in triage time, and policy-compliant, anticipatory containment recommendations that surpass the capabilities of current SIEM/SOAR/XDR and LLM copilot implementations. This framework provides a technically rigorous foundation for adaptive, explainable, and operationally safe autonomous defensive systems in enterprise environments, with significant implications for future real-world SOC deployment.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Collections

Sign up for free to add this paper to one or more collections.