Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap
Open accessPublished version of a preprint
The DOI matches its Crossref record; 146 of the 146 references it cites were found in a registry or an index; 43 of the 159 sources its claims cite were checked for support.
Abstract
This survey reviews agentic AI and LLM-driven approaches for autonomous IoT cybersecurity and identifies trustworthy multi-agent collaboration and resource-hardened edge agents as high-priority directions.
Overview
CoChat Summary
TL;DR
This PRISMA-style systematic survey synthesizes 153 peer-reviewed studies (2020–2026) on LLM- and agentic-AI-driven IoT cybersecurity, organizing them along a four-pillar taxonomy of agent architecture, reasoning strategy, action scope, and deployment topology. It concludes that the field is dominated by detection-only, single-dataset demonstrations, and proposes a 2026 roadmap centered on federated agentic learning, verifiable reasoning, trustworthy multi-agent collaboration, and resource-hardened edge agents.
Key findings
- Of the 41 studies coded as primary IoT-IDS work, 38 (~93%) operate in detection-only mode; the corpus contains no system conforming to MCP/ANP/A2A agent-interoperability protocols, and only 5 of 153 studies (~3%) report any compound multi-step failure mode (error propagation, context poisoning, tool-use amplification).
- Cross-dataset transfer degrades accuracy by 10–20 percentage points across at least nine primary studies, and only 2 of 13 agentic vulnerability-discovery studies report a verified false-discovery rate.
- Reported frontiers (Table 2): edge/on-device SLMs at 70–400 ms p50 with ~99% on Edge-IIoTset; fog/gateway 7–13B models at 300–900 ms with ~98% on CICIoT2023; cloud 70B+ models at 1–5 s end-to-end with ~98% prediction accuracy. Prompt-injection susceptibility is largely unmeasured at the edge against a CyberSecEval baseline of 26–41%.
- The companion kit's deterministic stub scored F1 = 0.83 on synthetic Edge-IIoTset-style traffic but only 0.06 accuracy on real Edge-IIoTset flows — reported as a negative control quantifying the synthetic-to-real gap.
Methods
- PRISMA 2020 protocol over IEEE Xplore, ACM Digital Library, and MDPI journals, January 2020–May 2026 (final search 29 May 2026); 487 candidate records → 63 duplicates removed → 424 title/abstract screened → 218 excluded as out of scope → 206 full-text screened → 53 excluded → 153 included studies.
- Corpus distribution: 58 IEEE, 39 ACM DL, 56 MDPI; by year, 9 (2020), 14 (2021), 18 (2022), 24 (2023), 36 (2024), 38 (2025), 14 (2026 to cut-off).
- Each study coded along seven dimensions (identifier, IoT layer/topology, agent architecture, reasoning strategy, cybersecurity focus, benchmarks, limitations); first-author screening with senior-author re-screening of a 15% random sample (Cohen's κ = 0.84, 90.4% agreement; seven adjusted inclusion decisions).
Limitations
- Stated: (1) corpus restricted to three venues, omitting USENIX Security, IEEE S&P, ACM CCS, NDSS and Elsevier/Springer journals; (2) screening primarily by one reviewer with only a 15% spot-check rather than two independent reviewers; (3) paywalled papers extracted from abstract/metadata only; (4) temporal upper bound of May 2026; (5) English-only corpus; (6) the companion reproducibility kit is v0.3.0 scaffolding, never validated in production, with only a deterministic stub run on synthetic traffic.
- Evident from the text: 56 of 153 included studies (37%) come from MDPI, the publisher of the manuscript itself — a disclosed but unresolved venue-concentration bias; the paper also notes that the top-tier-venue narrative scan was selected by expert judgement rather than reproducible screening and does not contribute to the 153-paper count.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
27 Valid · 18 Not valid · 61 Not checked · 154 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“SOAR and threathunting surveys [27,28] focus on workflow automation rather than autonomous reasoning.”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 4. Related Surveys and Positioning
What the sources say
- Disputedcites [27] Srinivas, S., Kirk, B., Zendejas, J., and Alzahrani, N. (2025). AI-Augmented SOC: A Survey of LLMs and Agents for Security Automation. J. Cybersecur. Priv., 5. sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) describes this paper as a survey of LLM and AI-agent applications across SOC functions, not specifically a SOAR or threat-hunting survey. Its Capability Maturity Model section explicitly examines progression from manual to fully autonomous operations, including independent complex analysis. That contradicts the characterization that it focuses on workflow automation rather than autonomous reasoning. (LLM, conf=0.87)
Checked by a reasoning model — may be wrong.
- Disputedcites [28] Kurnia, R., Brata, Z.A., Nelistiani, G.A., Heo, S., and Kim, H. (2025). Toward Robust Security Orchestration and Automated Response in SOC with a Hyper-Automation Approach Using Agentic AI. Information, 16. sourceExperimental — accuracy not yet measured
Our copy of 10.3390/info16050365 is not a SOAR/threat-hunting survey but a framework proposal (Abstract; Key Contributions) whose central argument runs counter to the claim: Section 2.2 states that traditional no-code/low-code SOAR is limited by static playbooks whereas 'contemporary SOAR solutions leverage LLMs to dynamically generate code... these AI-powered systems integrate contextual awareness and reasoning to respond in real time', and Sections 2.3/2.5.2 develop agentic AI and LLM 'multi-step reasoning' for autonomous incident response. The paper does discuss workflow/playbook automation (Section 2.2.4; Introduction on playbook challenges; References 6-12 on threat-hunting automation and playbook generation), which touches the 'workflow automation' half, but it explicitly foregrounds autonomous reasoning rather than omitting it. Thus the characterization 'focus on workflow automation rather than autonomous reasoning' is contradicted by our copy. (LLM, conf=0.72)
Checked by a reasoning model — may be wrong.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The 5G/6G transition compounds the problem by introducing network slicing-logical partitions whose isolation guarantees become themselves the target of slice-hopping attacks”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 2.1. The IoT Security Landscape
What the sources say
- Supportedcites [38] Allaw, Z., Zein, O., and Ahmad, A.-M. (2025). Cross-Layer Security for 5G/6G Network Slices: An SDN, NFV, and AI-Based Hybrid Framework. Sensors, 25. source
The Introduction describes network slicing as a 5G/6G feature that creates isolated virtual networks but raises concerns about cross-slice attacks. Section 2.2 adds that poor slice isolation can permit unauthorized access and cross-contamination between services, supporting the substance of the cited sentence. (LLM, conf=0.88)
- Partly supportedcites [39] Dias, J., Pinto, P., Santos, R., and Malta, S. (2025). 5G Network Slicing: Security Challenges, Attack Vectors, and Mitigation Approaches. Sensors, 25. sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) describes network slicing as independent virtual networks on shared infrastructure, and the Discussion (page unavailable) reports cross-slice access and lateral movement enabled by poor isolation. These passages support the substance of attacks crossing slice boundaries, but the paper addresses 5G slicing; it does not establish the claimed 5G/6G transition as the cause of compounded IoT security risks or identify those attacks by the term “slice-hopping.” (LLM, conf=0.82)
- Partly supportedcites [48] Botez, R., Zinca, D., and Dobrota, V. (2025). Redefining 6G Network Slicing: AI-Driven Solutions for Future Use Cases. Electronics, 14. sourceExperimental — accuracy not yet measured
Our copy squarely supports the definitional half of the sentence: the Introduction (Sec. 1) states that network slicing is 'already fundamental to 5G but envisioned to play an even more critical role in 6G' and creates 'virtualized, application-specific network segments on shared infrastructure' for the proliferating IoT, and Sec. 1.1 defines a slice as 'an independent, logically isolated virtual network that operates on shared physical infrastructure'. The attributed security point is absent: no passage in our copy mentions slice-hopping or any attack that targets slice isolation; on the contrary, the Introduction frames isolation as protective ('slicing also bolsters security by isolating sensitive data flows ... mitigating risks of data breaches or unauthorized access'). Searches for 'slice hopping attack', 'cross-slice', 'adversary moving between slices', and 'IoT security challenges/threats' returned only slicing-performance and requirement material (Sec. 1.1, Sec. 2 Related Works, Tables 3-5). So the source supports the transition-and-partition framing but not the claim that isolation guarantees become an attack target. (LLM, conf=0.70)
- Partly supportedcites [49] Karahan, S.N., Güllü, M., Karhan, D., Çimen, S., Osmanca, M.S., and Barışçı, N. (2025). Realistic Performance Assessment of ML Algorithms for 6G Network Slicing: A Dual-Methodology Approach with Explainable AI. Electronics, 14. sourceExperimental — accuracy not yet measured
Our copy supports the framing half of the claim: the Introduction (p.1) and §2.1 describe network slicing as 'virtualized end-to-end logical network slices on shared physical infrastructure' established in 5G and evolving into the 6G paradigm, and §2 Related Work and Background lists 'ensuring slice isolation... protecting data security and privacy' among the key challenges of the 5G/6G vertical scenarios, with §2.1 (summarising Cunha et al. [23]) naming 'slice isolation, inter-slice handover' as significant security challenges. However, the specific element attributed here — that these isolation guarantees become the target of 'slice-hopping attacks' — is nowhere stated: searches for 'slice-hopping attack', 'hopping between slices', and 'attacker bypass isolation' returned only semantic neighbours (slice classification performance, handover algorithms, SHAP tables), never an attack class exploiting inter-slice isolation. The paper is also a slice-classification performance study, so the isolation-threat material is background reported from other works rather than a finding of this paper. (LLM, conf=0.68)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“And they raise governance questions about accountability when autonomous decisions cause harm [27,28].”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 2.3. The Agentic AI Paradigm
What the sources say
- Supportedcites [27] Srinivas, S., Kirk, B., Zendejas, J., and Alzahrani, N. (2025). AI-Augmented SOC: A Survey of LLMs and Agents for Security Automation. J. Cybersecur. Priv., 5. source
Section 7 (Threats to Validity; page unavailable) identifies accountability for misaligned decisions and the ethical implications of semi-autonomous escalation as poorly explored. Section 4 (Capability Maturity Model; page unavailable) says greater autonomy shifts human roles toward governance and requires stronger ethical standards. Together, these support the claim as a paraphrase of a governance concern, not a claim that harm has occurred. (LLM, conf=0.87)
- Partly supportedcites [28] Kurnia, R., Brata, Z.A., Nelistiani, G.A., Heo, S., and Kim, H. (2025). Toward Robust Security Orchestration and Automated Response in SOC with a Hyper-Automation Approach Using Agentic AI. Information, 16. sourceExperimental — accuracy not yet measured
Our copy does touch the governance/accountability point, but only in the literature-review passage of Section 2.5.2 (LLMs and Agentic AI in Cybersecurity Domain), where it says the ethical implications of deploying autonomous AI systems in cybersecurity are explored by Radanliev et al., 'emphasizing the need for transparency and accountability' — i.e. the paper reports someone else's treatment rather than establishing the point itself. Nothing in our copy ties accountability questions to autonomous decisions that cause harm: the nearest risk discussion (Section 5, Discussion) is confined to misclassification/false positives, training-data quality, interpretability and the need for human intervention to validate critical steps, and Section 4.5 describes the agent deferring to human intervention rather than causing harm. The 'when autonomous decisions cause harm' qualifier in the citing sentence is therefore unsupported by this source. (LLM, conf=0.72)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Agents inherit and amplify the weaknesses of the underlying LLM-hallucination, prompt sensitivity, and contextual drift-and propagate them across longer decision chains [20,21].”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 2.3. The Agentic AI Paradigm
What the sources say
- Partly supportedcites [20] Jaffal, N.O., Alkhanafseh, M., and Mohaisen, D. (2025). Large Language Models in Cybersecurity: A Survey of Applications, Vulnerabilities, and Defense Techniques. AI, 6. sourceExperimental — accuracy not yet measured
Section 5 identifies hallucinations as an LLM risk, and §5.3 describes how prompts can manipulate LLM outputs. Section 4.5.2 also notes contextual limitations. The paper does not establish that agents amplify these weaknesses or propagate them across longer decision chains; no page numbers are available in our copy. (LLM, conf=0.82)
- Partly supportedcites [21] Atlam, H.F. (2025). LLMs in Cyber Security: Bridging Practice and Education. Big Data Cogn. Comput., 9. sourceExperimental — accuracy not yet measured
Section 6.1, “Model Reliability and Accuracy” (page not provided), reports LLM hallucinations and reliability that varies with prompt structure and context. The paper also describes difficulty with multi-step reasoning, but the passages do not establish that agents amplify these weaknesses or propagate them across longer decision chains; they do not specifically establish contextual drift in agents. (LLM, conf=0.84)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“IoV security has been reviewed by Maoudi et al. [102] and Junejo et al. [103] without an LLM agent lens.”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 4. Related Surveys and Positioning
What the sources say
- Not checked — no full text of the sourcecites [102] Das, 2026, Towards the Safety of Intelligent Transportation: A Survey on the Security Challenges and Mitigations in Internet of Vehicles (IoV), ACM Trans. Sens. Netw. source
- Partly supportedcites [103] Alalwany, E., and Mahgoub, I. (2024). Security and Trust Management in the Internet of Vehicles (IoV): Challenges and Machine Learning Solutions. Sensors, 24. sourceExperimental — accuracy not yet measured
The Introduction and Section 4 (Related Surveys; page not provided) show that this paper reviews IoV security and discusses earlier IoV-security surveys, with no LLM-agent lens evident in the cited passages. But this source is by Alalwany and Mahgoub, and the passages do not establish that the named Maoudi and Junejo papers are the reviews described in the claim. (LLM, conf=0.82)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
66 more claims · 8 with their cited sources checked
Create a free account to see all 260 claims and to verify the 44 references still open.
- Probably not a substantial claim. 105
- It may state two separate claims as one. 131
- The quote only partly states this claim. 2
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 5.4.3. Cloud-Centric
“Cloud topology removes resource constraints but introduces unbounded latency (hundreds of milliseconds end-to-end)”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > Model
“Cloud-centric 70B+ params, frontier API models 1-5 s end-to-end”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > Sanity-check validation run.
“No number from Appendix A is used anywhere in the body of this survey”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > Sanity-check validation run.
“the stub contains no language model, and its figures (binary attack-vs.-benignF 1 = 0.83) validate harness functionality only”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 5.3.2. Response Orchestration
“CyberAlly [27] is reported to have cut MTTR from 8 h to 90 min”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > Agentic AI does not eliminate the
“| CyberAlly [24] | Multi-step ReAct with analyst hand-off | SOC ticketing + EDR API | Vendor-reported MTTR delta |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 4. Related Surveys and Positioning
“Yao et al. [20] survey LLM applications across eight cybersecurity domains”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > References
“20. Jaffal, N.O.; Alkhanafseh, M.; Mohaisen, D. Large Language Models in Cybersecurity: A Survey of Applications, Vulnerabilities, and Defense Techniques.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
“the 5% prompt-injection ceiling is a tenth of the 26-41% CyberSecEval baseline for production LLMs”
§ Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap > 9.4. Resource-Hardened Edge Agents
Reported 0.1 · recomputed 0.192308
2 more findings
Create a free account to see every finding of this check.
Citations
Cites
Communication-Efficient Learning of Deep Networks from Decentralized Data
McMahan, H. B., Moore, E., Ramage, D., Hampson, S., & Arcas, B. A. Y. (2016) · arXiv · doi:10.48550/arxiv.1602.05629 · Cited by 14,215
Chain-Of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q. V., & Zhou, D. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1800 · Cited by 4,505
CICIoT2023: A real-time dataset and benchmark for large-scale attacks in IoT environment
Neto, E. C. P., Dadkhah, S., Ferreira, R., Zohourian, A., Lu, R., & Ghorbani, A. A. (2023) · doi:10.20944/preprints202305.0443.v1 · Cited by 1,235
CICIoT2023: A Real-Time Dataset and Benchmark for Large-Scale Attacks in IoT Environment
Neto, E. C. P., Dadkhah, S., Ferreira, R., Zohourian, A., Lu, R., & Ghorbani, A. A. (2023) · Sensors · doi:10.3390/s23135941 · Cited by 1,194
Machine Learning DDoS Detection for Consumer Internet of Things Devices
Doshi, R., Apthorpe, N., & Feamster, N. (2018) · 2018 IEEE Security and Privacy Workshops (SPW) · doi:10.1109/spw.2018.00013 · Cited by 680
Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., & Fritz, M. (2023) · Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security · doi:10.1145/3605764.3623985 · Cited by 667 · cited 3×
“The earlier direct-prompt-injection concept was named by Perez and Ribeiro in a NeurIPS 2022 ML Safety Workshop Best Paper, but as that work has not appeared in a formal peer-reviewed proceedings, we cite the peer-reviewed Greshake follow-up [94] as the foundational reference.”
InfluentialWhen LLMs meet cybersecurity: a systematic literature review
Zhang, J., Bu, H., Wen, H., Liu, Y., Fei, H., Xi, R., Li, L., Yang, Y., Zhu, H., & Meng, D. (2025) · Cybersecurity · doi:10.1186/s42400-025-00361-w · Cited by 383 · cited 2×
“The closest excluded-venue comparators—the Springer and Elsevier systematic reviews discussed in Section 3.8 [99,100]—are likewise partial with respect to this intersection: both treat LLMs-for-security broadly, but neither centers agentic control loops, IoT deployment topology, or a companion…”
InfluentialBotnet Detection in the Internet of Things using Deep Learning Approaches
McDermott, C. D., Majdani, F., & Petrovski, A. V. (2018) · 2018 International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn.2018.8489489 · Cited by 350
Zhou, X., Liang, W., Li, W., Yan, K., Shimizu, S., & Wang, K. I.-K. (2022) · IEEE Internet of Things Journal · doi:10.1109/jiot.2021.3130434 · Cited by 325 · cited 2×
“The first is the conventional model from non-LLM intrusion detection: adversaries craft traffic perturbations that evade ML classifiers [142–144].”
Artificial intelligence risk management framework :
National Institute of Standards and Technology (US) (2024) · NIST Trustworthy and Responsible AI · doi:10.6028/nist.ai.600-1 · Cited by 274
Research works printed but not linked: 6
- Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K.R., and Cao, Y. (2023, January 1–5). ReAct: Synergizing Reasoning and Acting in Language Models. Proceedings of the Eleventh International Conference on Learning Representations (ICLR 2023), Kigali, Rwanda. Available online: https://openreview.net/forum?id=WE_vluYUL-X.
- European Parliament and Council of the European Union (2026, May 29). Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act). Official Journal of the European Union, L Series. Available online: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689.
- Deng, G., Liu, Y., Mayoral-Vilches, V., Liu, P., Li, Y., Xu, Y., Zhang, T., Liu, Y., Pinzger, M., and Rass, S. (2024). PentestGPT: Evaluating and Harnessing Large Language Models for Automated Penetration Testing. Proceedings of the 33rd USENIX Security Symposium (USENIX Security ’24), Philadelphia, PA, USA, 14–16 August 2024, USENIX Association. Available online: https://www.usenix.org/conference/usenixsecurity24/presentation/deng.
- Arp, D., Quiring, E., Pendlebury, F., Warnecke, A., Pierazzi, F., Wressnegger, C., Cavallaro, L., and Rieck, K. (2022). Dos and Don’ts of Machine Learning in Computer Security. Proceedings of the 31st USENIX Security Symposium (USENIX Security ’22), Boston, MA, USA, 10–12 August 2022, USENIX Association. Available online: https://www.usenix.org/conference/usenixsecurity22/presentation/arp.
- Judith, A., Kathrine, G.J.W., Silas, S., and Andrew, J. (2023). Efficient DL-Based Cyber-Attack Detection for IoMT Devices. Eng. Proc., 59.
- Kumar, P.V., Pittala, R.B., Jashwanth, P., Vedamsh, R., Chandrika, D.S., and Reddy, Y.V. (2025). Automated Threat Detection and Response Using SOAR and Endpoint Detection and Response. Proceedings of the 2025 OITS International Conference on Information Technology (OCIT), IEEE. Available online: https://ieeexplore.ieee.org/document/11400041/.
Cited by
Nageshwaran, V., & Ezekiel, S. (2026) · doi:10.20944/preprints202606.0142.v1 · Cited by 4
Self-citationAutonomous LLM Agent for Real-Time DDoS Attack Classification and Mitigation Rule Generation
Phani, K. V. S., Karthik, P. N., Mahammad, F. S., Subramanyam, M. V., & Srikanth, B. (2026) · International Conference on Intelligent and Sustainable AI Systems (ICOSAAS) · doi:10.1109/icosaas68663.2026.11648534 · Cited by 1
Garike, V. (2026) · Research Square · doi:10.21203/rs.3.rs-10582034/v1
Metadata
- Authors
- Nageshwaran, V., & Ezekiel, S.
- Affiliations
- University of the Cumberlands (US)
- Indiana University of Pennsylvania (US)
- Corresponding author
- Nageshwaran, Vinoth
- ORCID
- Nageshwaran, Vinoth
- Venue
- Electronics
- Publisher
- MDPI AG
- Published
- 2026-06-22
- Volume
- 15
- Issue
- 12
- Pages
- 2740
- ISSN
- 2079-9292 (ISSN-L)
- OpenAlex
- W7165680161
- Cited by
- 4
- References
- 128
- Citations by year
- 2026: 4
- FWCI
- 14.96
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Electronics (publishedVersion, cc-by)
- Preprints.org (acceptedVersion, cc-by)
- Topics
- Physical Sciences › Computer Science › Computer Networks and Communications › Network Security and Intrusion Detection
- Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Keywords
- agentic AI, large language models, Internet of Things security, multi-agent systems, chain-of-thought reasoning, React, tool use, threat detection and response, vulnerability detection, deception, edge computing, fog computing, cloud deployment, hallucinations, prompt injection, explainability, privacy, federated learning, research roadmap, taxonomy (general classification scheme)
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
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
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.