AutoBnB: Multi-Agent Incident Response with Large Language Models
Abstract
Incident response (IR) is a critical aspect of cyberse-curity, requiring rapid decision-making and coordinated efforts to address cyberattacks effectively. Leveraging large language models (LLMs) as intelligent agents offers a novel approach to enhancing collaboration and efficiency in IR scenarios. This paper explores the application of LLM-based multi-agent col-laboration using the Backdoors & Breaches (B&B) framework, a tabletop game designed for cybersecurity training. We simulate realistic IR dynamics through various team structures, including centralized, decentralized, and hybrid configurations. By ana-lyzing agent interactions and performance across these setups, we provide insights into optimizing multi-agent collaboration for incident response. Our findings highlight the potential of LLMs to enhance decision-making, improve adaptability, and streamline IR processes, paving the way for more effective and coordinated responses to cyber threats.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
IRCopilot is a third example that uses Planner, Generator, Reflector, and Analyst agents to simulate a full range of incident response processes and documentation.
“IRCopilot is a third example that uses Planner, Generator, Reflector, and Analyst agents to simulate a full range of incident response processes and documentation [34].”
From The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines (2026) · cites this paper as [34]
The multi-agent paradigm allows multi-step multi-task complexity (i.e., associating multiple log sources) to be passed between agents, leading to more reliable reasoning.
“It allows the multi-step multi-task complexity (i.e., associating multiple log sources) to be passed between agents, leading to more reliable reasoning [33], [34].”
From The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines (2026) · cites this paper as [34]
Gen-3 has its very own class of failure modes. Agents may have competing hypotheses, resulting in diverging paths to investigate.
“but Gen-3 has its very own class of failure modes. Agents may have competing hypotheses, resulting in diverging paths to investigate [33], [34].”
From The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines (2026) · cites this paper as [34]
Systems such as Microsoft Security Copilot and the IRCopilot prototype exhibit elements of Gen-5, with discrete components for detection, investigation, containment, and remediation, each with clear roles.
“Systems like Microsoft Security Copilot or academic prototypes like IRCopilot exhibit elements of this idea, with discrete components for detection, investigation, containment, and remediation, each with clear roles [34].”
From The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines (2026) · cites this paper as [34]
IRCopilot reports up to about 150% more tasks completed than a traditional single monolithic LLM.
“IRCopilot reports up to ~150% more tasks completed compared to a traditional single monolithic LLM, which emphasizes what is possible with semi-automated orchestration with constrained autonomy [34].”
From The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines (2026) · cites this paper as [34]
Gen-3 architectures remain a valuable foundation for end-to-end SOC workflows as they distribute the cognitive load and enable multi-tiered quality checks.
“However, despite these challenges, Gen-3 architectures remain a valuable foundation for end-toend SOC workflows as they distribute the cognitive load and enable multi-tiered quality checks [33], [34].”
From The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines (2026) · cites this paper as [34]
Cites
Still being checked.
Cited by
Sorted by citations
AI-Augmented SOC: A Survey of LLMs and Agents for Security Automation
Srinivas, S., Kirk, B., Zendejas, J., Espino, M., Boskovich, M., Bari, A., Dajani, K., & Alzahrani, N. (2025) · Journal of Cybersecurity and Privacy · doi:10.3390/jcp5040095 · Cited by 39 · see citing works
A Multi-Agent System for Cybersecurity Threat Detection and Correlation Using Large Language Models
Hmimou, Y., Tabaa, M., Khiat, A., & Hidila, Z. (2025) · IEEE Access · doi:10.1109/access.2025.3602681 · Cited by 25 · see citing works
Vinay, V. (2026) · 2026 IEEE 5th International Conference on AI in Cybersecurity (ICAIC) · doi:10.1109/icaic67076.2026.11395809 · Cited by 15 · see citing works
Leveraging large language models to enhance multi-agent risk assessment in supply chain networks
Quan, Y., Liu, Z., Bénaben, F., & Montreuil, B. (2026) · International Journal of Production Research · doi:10.1080/00207543.2026.2619562 · Cited by 9 · see citing works
Self-citationAutoBnB-RAG: Enhancing Multi-Agent Incident Response with Retrieval-Augmented Generation
Liu, Z., & Anwar, A. (2025) · IEEE ... International Conference on Data Mining workshops · doi:10.1109/icdmw69685.2025.00141 · Cited by 7 · see citing works
Self-citationNing, B., He, K., Zong, X., & Sun, Y. (2025) · International Conference Automatics and Informatics (ICAI) · doi:10.1109/icai67591.2025.11325015 · Cited by 2 · see citing works
Tiwari, R., & Nene, M. J. (2026) · IEEE International Conference for Convergence in Computing Technology (I3CTCON) · doi:10.1109/i3ctcon68242.2026.11507245
A systematic evaluation of large language models for autonomous cyber defense
Jacques, T. (2026) · Applied Intelligence · doi:10.1007/s10489-026-07386-9
AI-Powered Unified Intrusion Detection and Autonomous Response Architecture for Smart Grids
Cheng, B., Hsu, P.-H., & Hung, Y. H. A. T.-L. (2026) · IntechOpen eBooks · doi:10.5772/intechopen.1015375
Generative AI and Cybersecurity Education, Training, and Awareness: A Systematic Mapping Study
Arpacı, İ., Khan, N. F., Qayyum, S., & Балоглу, М. (2026) · IEEE Access · doi:10.1109/access.2026.3738195
Toward Trustworthy AI Helpdesk: A Multi-Agent Architecture for Cybersecurity Incident Pre-Triage
Zharfa, A., Girinoto, Hadiprakoso, R. B., Qomariasih, N., & Musonip, M. L. R. (2026) · International Conference on Computer, Control, Informatics and its Applications · doi:10.1109/ic3ina72033.2026.11714886
- Authors
- Liu, Z.
- Affiliations
- Georgia Institute of Technology (US)
- Corresponding author
- Liu, Zefang
- ORCID
- Liu, Zefang
- Venue
- 2025 13th International Symposium on Digital Forensics and Security (ISDFS)
- Publisher
- IEEE
- Published
- 2025-04-24
- Pages
- 1-6
- OpenAlex
- W4411207733
- Cited by
- 11
- References
- 0
- Citations by year
- 2026: 7 · 2025: 4
- FWCI
- 12.58
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- International Symposium on Digital Forensics and Security (ISDFS) (publishedVersion)
- Topics
- Physical Sciences › Computer Science › Information Systems › Information and Cyber Security
- Physical Sciences › Computer Science › Computer Networks and Communications › Network Security and Intrusion Detection
- Physical Sciences › Computer Science › Information Systems › Digital and Cyber Forensics
- Keywords
- incident response, multi-agent systems, large language models, LLM agents, tabletop games, cybersecurity training, agent cooperation, cyberattack response
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
