Technical Summary: AgenticVM – Agentic AI for Adaptive Software Vulnerability Management

1. Problem Statement

Modern software vulnerability management is increasingly strained by the scale and complexity of contemporary systems. Security Operations Centers (SOCs) face a critical bottleneck where vulnerability discovery consistently outpaces remediation efforts. This is driven by:

• High Alert Volumes: Scanners generate massive amounts of raw findings, often leading to persistent backlogs.
• Fragmented Toolchains: Disconnected tools and manual triage processes create inefficiencies.
• Incomplete Context: Analysts often must prioritize under uncertainty due to missing or delayed official severity scores (CVSS) and incomplete vulnerability descriptions.
• Limitations of Current Approaches: Traditional rule-based systems are brittle against evolving attacker behaviors, while classical machine learning requires extensive feature engineering and retraining. Existing Large Language Model (LLM) applications often optimize for isolated subtasks rather than full operational integration, lacking the governance and observability required for real-world deployment.

2. Methodology

The paper introduces AgenticVM, a modular, multi-agent framework designed to automate the entire vulnerability management lifecycle. The methodology is structured into five phases: literature review, data acquisition, system architecture design, implementation, and evaluation.

A. System Architecture

AgenticVM employs a multi-agent orchestration layer (using LangGraph) that coordinates specialized agents, each handling distinct tasks with appropriate techniques (rule-based, learned ML, or LLM-driven):

1. Detection Agent [Hybrid]: Ingests heterogeneous sources (e.g., Trivy, Snyk, Grype). It uses deterministic rule-based parsers for standard scanner outputs and LLM-based parsing for unstructured logs to ensure high recall.
2. Assessment Agent [Rule-based]: Acts as a logical verification gate to reduce false positives by cross-referencing identified vulnerabilities against project context and local asset metadata. It prioritizes transparency and repeatability over generative flexibility.
3. Prediction Agent [Learned ML]: Utilizes a BERT-based transformer model to infer missing CVSS v3.1 metrics (Attack Vector, Complexity, Privileges Required, User Interaction, Scope, and Impact metrics) from vulnerability descriptions. It includes a confidence-gating policy where low-confidence predictions are flagged for manual review.
4. Integration Agent [Rule-based]: Consolidates outputs from previous agents, enforcing strict schema validation (via Pydantic-style records) to ensure data consistency and completeness.
5. Prioritisation Agent [Rule-based]: Applies deterministic CVSS-based tiering (e.g., thresholding at 7.0) to filter and map vulnerabilities into priority bands. It incorporates signals like CISA Known Exploited Vulnerabilities (KEV) as enrichment rather than opaque multipliers.
6. Recommendation Agent [LLM-driven]: Synthesizes remediation advice using a retrieval-grounded generation strategy, combining vulnerability context with authoritative sources (NVD, CISA KEV). It supports analyst decision-making but requires human approval for destructive actions.

B. Data and Evaluation

• Datasets: The framework was trained and evaluated using 169,883 unique CVE records from the National Vulnerability Database (NVD), European Union Vulnerability Database (EUVD), and CISA KEV catalog.
• Scenarios: Experiments were conducted on real-world open-source applications (Online Boutique, Train Ticket, Beer Shop) using raw scanner outputs.
• Metrics: Evaluation included standard classification metrics (Accuracy, Precision, Recall, F1) for the prediction model and operational metrics for the agent workflow: Task Completion Time (TCT), Cost Efficiency Rate (CER), Token Consumption Rate (TCR), and Cooperation Performance (Cperf).

3. Key Contributions

The paper makes three primary contributions:

1. AgenticVM Framework: A modular multi-agent system that integrates LLMs with conventional security tooling to automate the full lifecycle from detection to remediation recommendation.
2. Hybrid Pipeline: A novel combination of rule-based processing, a BERT-based CVSS prediction module, and specialized LLM agents. This approach enables reliable severity estimation and alert noise reduction without compromising risk coverage.
3. Empirical Validation and Engineering Insights: The study demonstrates significant improvements in alert reduction and prioritization efficiency. It also distills practical design insights regarding agent decomposition, tool–LLM integration boundaries, failure containment, and human-in-the-loop governance.

4. Experimental Results

• CVSS Prediction Performance: The BERT-small model achieved an overall accuracy of 89.3% across the eight CVSS base metrics. It showed particularly strong performance in predicting Attack Vector, Attack Complexity, User Interaction, and Scope.
• Alert Reduction: AgenticVM successfully compressed raw scanner outputs into compact, actionable queues.
  • In the Train Ticket scenario, raw detections were reduced from 3,983 findings to 82 high-priority items (a 97.9% reduction).
  • In the Online Boutique scenario, 32 detections were reduced to 6 prioritized items (81.3% reduction).
• Baseline Comparisons: Compared to a manual workflow (0% reduction) and a single-agent LLM workflow (modest reduction, e.g., 28.5% in Train Ticket), AgenticVM demonstrated superior noise reduction and provenance control.
• Ablation Studies: Removing the Prediction Agent led to under-prioritization of vulnerabilities with missing CVSS data. Removing the Assessment Agent increased false positives. Removing the Prioritisation Agent increased downstream effort and execution time.
• Stability: Repeated runs showed robust stability with minor variations typical of probabilistic LLM workflows.

5. Significance and Claims

The paper positions AgenticVM not as a fully autonomous remediation engine, but as an analyst-amplification framework. Its significance lies in:

• Operational Feasibility: It demonstrates that agentic AI can be deployed in real-world security operations to reduce manual workload while maintaining reliable performance.
• Risk Visibility: The system reduces alert fatigue without compromising risk visibility, allowing teams to focus on high-priority items.
• Governance and Trust: By decomposing tasks into specialized agents, the framework improves fault isolation, observability, and auditability. It explicitly incorporates human-in-the-loop governance, ensuring that destructive actions require human approval.
• Design Insights: The work provides practical guidance on how to balance deterministic logic (for validation and prioritization) with LLM capabilities (for parsing and recommendation) to mitigate hallucination risks and ensure reproducibility.

The authors conclude that while the system significantly improves prioritization efficiency, future work is needed to connect recommendations to sandboxed testing, regression checks, and more robust human-centered evaluations of trust calibration.