Can Large Language Models Be Trusted for AI-Based Intrusion Detection in Cybersecurity? An Evidence-Grounded Log Analysis
The DOI matches its Crossref record; 24 of the 24 references it cites were found in a registry or an index; 3 of the 30 sources its claims cite were checked for support.
Abstract
Security Operations Centers (SOCs) rely heavily on system logs for intrusion detection, yet the growing volume and complexity of modern infrastructure logs make manual analysis increasingly impractical. While traditional intrusion detection systems provide automated alerts, they often lack contextual reasoning and interpretable explanations. Recent advances in large language models (LLMs) offer new opportunities for semantic log analysis and natural-language reasoning, motivating their use as automated SOC assistants. However, existing studies largely evaluate LLMs based on detection performance or qualitative usefulness, without rigorously assessing whether their explanations are grounded in actual log evidence. In this paper, we propose a systematic evaluation framework for applying closed-weight, web-accessible LLMs to intrusion detection from raw Linux system logs. Using standardized prompts and fixed log windows, we treat each LLM as a black-box SOC analyst that produces structured intrusion assessments, suspicion scores, and cited evidence. Beyond traditional labels, we introduce evidence faithfulness as a core evaluation metric, measuring whether the evidence referenced by a model is explicitly present in the input logs. We evaluate multiple widely used closed weight online LLMs, including ChatGPT, Gemini, Grok, NotebookLM, Perplexity, and Claude. Our results reveal significant differences in suspicion calibration, alert generation, and evidence grounding across models. While several LLMs consistently flag suspicious activity, unfaithful or partially hallucinated evidence remains common. These findings highlight that reliable explanation grounding is essential for trustworthy LLM-assisted intrusion detection and motivate future research on evaluation-driven prompt and model design.
Overview
CoChat Summary
TL;DR
The paper proposes an evaluation framework that treats six closed-weight, web-accessible LLMs as black-box SOC analysts over raw Linux system logs, adding a strict evidence-faithfulness metric that checks whether cited log lines appear verbatim. It finds large differences in suspicion calibration and grounding, with NotebookLM at 0% evidence faithfulness despite high suspicion scores, while ChatGPT, Gemini, and Claude show high grounding.
Key findings
- Across Table I, mean suspicion scores ranged from 49.50 (ChatGPT) to 93.00 (NotebookLM); Claude had 83.83 and a 66.67% attack rate, while all other models had 0% attack rate.
- Evidence faithfulness was 100% for ChatGPT and Claude, 90% for Gemini, 80% for Grok and Perplexity, and 0% for NotebookLM due to systematic paraphrasing despite confident judgments.
- Average alerts per window were 3.00 for Claude and NotebookLM, 2.80 for Gemini, 2.50 for ChatGPT, and 2.00 for Grok and Perplexity.
- Suspicion scores were model-specific and not directly comparable; strong reasoning did not guarantee forensic traceability, and evidence faithfulness was independent of detection accuracy.
Methods
- Linux system logs from `/var/log/messages` on a server environment with simulated intrusion and anomalous activities were partitioned into non-overlapping, chronologically ordered windows of exactly 200 consecutive log lines.
- Six black-box LLMs were evaluated: ChatGPT (GPT-5.2), Grok 4.1, Gemini 3, NotebookLM (Gemini 3), Claude (Opus 4.5; subset of windows due to context limitations), and Perplexity (Sonar), using a standardized SOC-analyst prompt and strict JSON output schema without fine-tuning or retrieval augmentation.
- Evidence faithfulness was a binary window-level metric requiring all cited evidence lines to appear verbatim in the input log window; other metrics included mean and standard deviation of suspicion scores, alerts per window, intrusion label distribution, and evidence faithfulness rate, with no ground-truth attack labels assumed.
Limitations
- The paper states limitations including the absence of external ground truth, black-box model treatment, and logs from a single environment; Claude was evaluated on a subset of windows due to context-length limitations and excluded from aggregate comparisons where appropriate.
- Evident from the text: the small window counts (6 for Claude, 10 for other models) and single-environment logs limit generalizability, and the lack of ground-truth labels means actual detection accuracy cannot be assessed.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
2 Valid · 1 Not valid · 13 Not checked · 18 Own results and interpretations
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Recent NLP studies have emphasized measuring how well model outputs stay supported by provided data (i.e., faithfulness) [2], [5], but this has not been systematically applied to intrusion detection scenarios.”
What the sources say
- Not checked — no full text of the sourcecites [2] source
- Supportedcites [5] source
The citation is attached to the first clause, about NLP studies measuring whether model outputs are supported by provided sources. Appendix J defines faithfulness as whether a sentence is justified by its source, and Appendix B describes the paper's attributability measure. The subsequent statement about intrusion detection is the citing authors' framing, not a finding attributed to this paper. (LLM, conf=0.90)
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.
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“LLMs are known to sometimes produce convincing yet ungrounded statements or hallucinated details [2], [17].”
What the sources say
- Not checked — no full text of the sourcecites [2] source
- Supportedcites [17] Yang, 2025, Large Language Models for Network Intrusion Detection Systems: Foundations, Implementations, and Future Directions source
The Challenges/Validity passage (p. 7) states that 'Due to the inherent unpredictability of LLMs, threat assessments may be compromised by hallucinations, where models generate inaccurate, misleading, or contextually irrelevant outputs,' which is exactly the content the citing sentence attributes: LLMs sometimes emit ungrounded/hallucinated output. The same paper's LLM background section (p. 2) frames LLMs as generative models trained on vast data, and the validity discussion treats hallucination as a recognized, recurring limitation rather than a hypothetical, so the hedged general statement is carried with equal or greater confidence. The source states the point in its own voice (it cites Huang et al.'s hallucination survey only as footnote 19 for the term), so this is not a borrowed claim; the only nuance not literally spelled out is that the outputs are 'convincing', which is implied by 'misleading' and is wording rather than substance. (LLM, conf=0.85)
boosted- 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.
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“However, existing studies largely evaluate LLMs based on detection performance or qualitative usefulness, without rigorously assessing whether their explanations are grounded in actual log evidence.”
- Also stated in: § I. Introduction · plain
“Current evaluations largely focus on detection accuracy or user satisfaction, with little attention to whether an LLM's stated evidence truly reflects the contents of the logs.”
What the sources say
- Not checked — no full text of the sourcecites [2] source
- Could not verifycites [5] sourceExperimental — accuracy not yet measured
The Related Work section discusses evaluation of citation quality in evidence-based question answering, but does not characterize studies of log analysis. A search for log evidence, detection performance, and qualitative usefulness found no passage supporting the claim about how those studies evaluate LLM explanations. (LLM, conf=0.76)
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.
- Also stated in: § I. Introduction · plain
None of the sources cited for this claim could be matched to a paper in Crossref, so there is nothing to check it against.
In the paper
“ChatGPT [6], Gemini [10], and Claude [7] maintain high grounding by citing exact log excerpts.”
What the sources say
- Not matchedcites [6] ChatGPT
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [10] Gemini
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [7] Claude
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
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.
- Not matched
None of the sources cited for this claim could be matched to a paper in Crossref, so there is nothing to check it against.
In the paper
“Perplexity [8] and Grok [11] show mixed grounding on ambiguous logs.”
What the sources say
- Not matchedcites [8] Perplexity
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [11] Grok
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
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.
- Not matched
4 more claims
Create a free account to see all 34 claims and to verify the 7 references still open.
- It may state two separate claims as one. 15
- Probably not a substantial claim. 14
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 major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ A. Aggregate Model Performance
“NotebookLM [9] and Claude [7] generate consistently high suspicion scores with frequent attack escalations”
§ References
“| NotebookLM[9] | 93.00 | 3.50 | 0.00 | 0.00 | 0.00 | 0.00 | 3.00 | 3.00 | 10 |”
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
§ IV. Experimental Setup
“Each model independently analyzes the same set of log windows without access to prior context, memory, or external tools.”
§ References
“| Claude[7] | 83.83 | 13.50 | 100.00 | 1.00 | 66.67 | 66.67 | 3.00 | 3.00 | 6 |”
Citations
Cites
Survey of Hallucination in Natural Language Generation
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023) · ACM Computing Surveys · doi:10.1145/3571730 · Cited by 5,025
Mi, H., Wang, H., Zhou, Y., Lyu, M. R.-T., & Cai, H. (2013) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2013.21 · Cited by 176
Generative AI in cybersecurity: A comprehensive review of LLM applications and vulnerabilities
Ferrag, M. A., Alwahedi, F., Battah, A., Cherif, B., Mechri, A., Tihanyi, N., Bisztray, T., & Debbah, M. (2025) · Internet of Things and Cyber-Physical Systems · doi:10.1016/j.iotcps.2025.01.001 · Cited by 148 · cited 3×
“Large language models (LLMs) offer a promising avenue to analyze unstructured logs and derive insights in a more flexible manner, potentially correlating diverse events, interpreting log semantics, and generating intuitive explanations [17], [18].”
LILAC: Log Parsing using LLMs with Adaptive Parsing Cache
Jiang, Z., Liu, J., Chen, Z., Li, Y., Huang, J., Huo, Y., He, P., Gu, J., & Lyu, M. R. (2024) · Proceedings of the ACM on Software Engineering · doi:10.1145/3643733 · Cited by 123
ChatGPT for digital forensic investigation: The good, the bad, and the unknown
Scanlon, M., Breitinger, F., Hargreaves, C., Hilgert, J.-N., & Sheppard, J. (2023) · Forensic Science International: Digital Investigation · doi:10.1016/j.fsidi.2023.301609 · Cited by 118
LogPrompt: Prompt Engineering Towards Zero-Shot and Interpretable Log Analysis
Liu, Y., Tao, S., Meng, W., Yao, F., Zhao, X., & Yang, H. (2024) · Proceedings of the 2024 IEEE/ACM 46th International Conference on Software Engineering: Companion Proceedings · doi:10.1145/3639478.3643108 · Cited by 96
Trustworthy LLMs: a Survey and Guideline for Evaluating Large Language Models' Alignment
Liu, Y., Yao, Y., Ton, J.-F., Zhang, X., Guo, R., Cheng, H., Klochkov, Y., Taufiq, M. F., & Li, H. (2023) · arXiv · doi:10.48550/arxiv.2308.05374 · Cited by 79
Ali, T., & Kostakos, P. (2023) · arXiv · doi:10.48550/arxiv.2309.16021 · Cited by 47
PreLog: A Pre-trained Model for Log Analytics
Le, V.-H., & Zhang, H. (2024) · Proceedings of the ACM on Management of Data · doi:10.1145/3654966 · Cited by 46
Towards Explainable Network Intrusion Detection using Large Language Models
Houssel, P. R. B., Singh, P., Layeghy, S., & Portmann, M. (2024) · 2024 IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT) · doi:10.1109/bdcat63179.2024.00021 · Cited by 43
Research works printed but not linked: 1
- Li, IDS-Agent: An LLM agent for explainable intrusion detection in IoT networks
Metadata
- Authors
- Beyzavi, H., Shiraskar, S., Albluwi, A., Albalawi, M., Rizk, D., Rizk, F., Rizk, P., Nehmetallah, G., Rizk, R., & Santosh, K.
- Affiliations
- IBM (United States) (US)
- Catholic University of America (US)
- King Abdulaziz University (SA)
- The University of Texas at Tyler (US)
- Arctus Inc. (Canada) (CA)
- University of South Dakota (US)
- Venue
- 2026 IEEE Conference on Artificial Intelligence (CAI)
- Publisher
- IEEE
- Published
- 2026-05-08
- Pages
- 1805-1810
- OpenAlex
- W7163066643
- Semantic Scholar
- 23545df201b589c90e24c99757e85a1bcd5e969f
- Cited by
- 0
- References
- 14
- FWCI
- 0.00
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- IEEE Conference on Artificial Intelligence (CAI) (publishedVersion)
- Topics
- Physical Sciences › Computer Science › Computer Networks and Communications › Network Security and Intrusion Detection
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Physical Sciences › Computer Science › Computer Networks and Communications › Software System Performance and Reliability
- Keywords
- large language models, intrusion detection, security operations center, system log analysis, hallucinations, LLM evaluation
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 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Can Large Language Models Be Trusted for AI-Based Intrusion Detection in Cybersecurity? An Evidence-Grounded Log Analysis. https://app.cochat.ai/papers/10.1109/cai68641.2026.11536239
- The paper
- Beyzavi, H., Shiraskar, S., Albluwi, A., Albalawi, M., Rizk, D., Rizk, F., Rizk, P., Nehmetallah, G., Rizk, R., & Santosh, K. (2026). Can Large Language Models Be Trusted for AI-Based Intrusion Detection in Cybersecurity? An Evidence-Grounded Log Analysis. 2026 IEEE Conference on Artificial Intelligence (CAI). https://doi.org/10.1109/cai68641.2026.11536239
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.