Skip to content

IRSDA: An Agent-Orchestrated Framework for Enterprise Intrusion Response

🕒 Published (v1): 2025-11-24 19:21 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

IRSDA is a five-tier, multi-agent framework for autonomous enterprise intrusion response that couples a MAPE-K control loop with a graph-based knowledge store (IRSKG), a cybersecurity-fine-tuned LLM (IRSLLM), and a Google ADK-powered agentic orchestration layer. It translates raw IDS alerts into ROE-compliant containment actions while surfacing natural-language explanations to security analysts. The design is demonstrated on a microservices e-commerce platform (Google Online Boutique) under a simulated front-end compromise.

Problem

Traditional intrusion response relies on static playbooks and manual analyst workflows that cannot match the speed or distributed nature of modern multi-stage attacks. IDS and IRS components are siloed, leaving a gap between automated detection and automated, policy-compliant response. Existing LLM-assisted security tools also suffer from hallucination and lack enterprise-grounded knowledge.

Method

IRSDA is structured as an \(n\)-tier client-server + multi-agent system (MAS):

  • Tier I–II: A chatbot client (IRSDAC) and session-managing server (IRSDAS) handle analyst queries.
  • Tier III: An Agentic Orchestration layer (IRSDAAO) hosts an Agentic Brain (AB) that decomposes tasks and routes them to partition-specific subagents. Implemented with Google ADK; three subagents cover execution, synthesis, and monitoring.
  • Tier IV: Partition-level IRS agents operate locally on each system component type \(C_i\), firing GNN/RL computational models and querying the knowledge layer.
  • Tier V:
  • IRSKG — a Neo4j property graph \(G=(V,E)\) storing configurations (partitioned as \(M=\{C_1,\ldots,C_n\}\)), system logs, Rules of Engagement (ROE \(R=\{R_i\}\)), and ML training data; vector indices enable semantic retrieval.
  • IRSLLM — CyBERT fine-tuned via LoRA on enterprise IRSKG corpora; restricted to QA and summarization tasks; prompt templates are whitelisted against injection.

Graph-RAG (GRAG) grounds IRSLLM outputs: for a query \(Q_i\), relevant chunks from IRSKG are retrieved, injected into a prompt template \(T_i\) (multi-shot, zero-shot, or chain-of-thought), and the resulting responses \(A_i\) are ensembled and summarized into a tabbed answer set \(A\).

ROE are encoded as graph triples \(R_i = \{V_a(R_i), E(R_i), V_b(R_i)\}\) that both constrain automated actions and serve as retrieval targets during incident response.

Key Contributions

  • IRSDA framework: end-to-end integration of intrusion detection and automated response via iterative agentic orchestration, MAPE-K loop, and SA-ACS principles.
  • IRSKG schema: a formally specified property-graph knowledge base (vertices, edges, labels, properties) unifying configurations, logs, ROE, and ML training data with vector indices for GRAG.
  • IRSLLM: domain-adapted LLM (CyBERTuned → LoRA fine-tune on enterprise data) with whitelisted prompt templates to constrain scope and mitigate hallucination.
  • GRAG pipeline: graph-retrieval-augmented generation grounded in live enterprise telemetry and dynamic ROE.
  • Case study: end-to-end walkthrough on Google Online Boutique (11 microservices), from front-end compromise detection to ROE-compliant email-service containment with analyst-readable explanation.

Results

  • No quantitative benchmarks (latency, accuracy, F1, containment time) are reported in the provided text.
  • The paper is a system/architecture paper evaluated through a single qualitative case study on the Online Boutique microservices application.
  • The case study demonstrates successful containment of a simulated front-end → email-service abuse scenario with traceable, ROE-compliant outputs and natural-language analyst explanations; specific metrics are not stated in the excerpt.

Limitations

  • No quantitative evaluation against baselines (other IRS frameworks, rule-based systems, or prior LLM-based security assistants).
  • Case study covers a single, controlled scenario on a simulated system; generalizability to production enterprise environments is unvalidated.
  • Scalability of the graph-RAG pipeline and the multi-agent orchestration under high-throughput, real-time alert volumes is not assessed.
  • IRSLLM fine-tuning quality and hallucination rates are not measured empirically; ROE coverage is scenario-specific.
  • Three-subagent orchestration is described as a prototype; extensibility claims are architectural, not empirically tested.
  • Dependency on Google ADK creates a platform coupling despite claims of platform-agnosticism.

Relevance to Harnesses / Meta-Harnesses

IRSDA is a domain-specific meta-harness that mirrors the canonical harness pattern — a central orchestrator (Agentic Brain) dispatches tasks to specialized sub-agents, each of which consults a shared knowledge store and a fine-tuned model, then returns results for synthesis — applied here to the cybersecurity incident-response domain. The MAPE-K loop functions as the harness's control cycle, making IRSDA a concrete instantiation of how harness architecture (monitor → analyze → plan → execute) can be operationalized with LLMs and graph-RAG. The whitelisted prompt template registry (\(T\)) and the structured ensemble/summarization pipeline (\(A_i \to U_i \to A\)) are directly analogous to prompt-management and output-aggregation mechanisms common in general-purpose harnesses. For researchers tracking harness design, IRSDA offers a worked example of partition-aware agent routing, policy-as-graph ROE encoding, and trust-limiting mechanisms (prompt whitelisting, GRAG grounding) at the orchestration layer.