CoChat

Agentic AI and Large Language Models for Autonomous IoT Cybersecurity: A Systematic Survey, Taxonomy, and Research Roadmap

Open access

Published 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

Crossref · 10.3390/electronics15122740

Cites

146 of 146 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

27 Valid · 18 Not valid · 61 Not checked · 154 Own results and interpretations

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.

See all claims — free account
  • Probably not a substantial claim. 105
  • It may state two separate claims as one. 131
  • The quote only partly states this claim. 2

Create a free account to see these claims.

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.

5 of 7 · the rest aren’t loaded

2 more findings

Create a free account to see every finding of this check.

See all findings — free account

Citations

Cites

Research works printed but not linked: 6

  1. 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.
  2. 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.
  3. 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.
  4. 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.
  5. Judith, A., Kathrine, G.J.W., Silas, S., and Andrew, J. (2023). Efficient DL-Based Cyber-Attack Detection for IoMT Devices. Eng. Proc., 59.
  6. 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/.

Metadata

Authors
Nageshwaran, V., & Ezekiel, S.
Affiliations
  • University of the Cumberlands (US)
  • Indiana University of Pennsylvania (US)
Corresponding author
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
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.