Beyond the Syntax: Do Security Experts Trust LLMs for NIDS Rule Engineering?

arXiv:2607.05916 · cs.CR · Submitted 2026-07-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "Beyond the Syntax".

Elias: Here's the gist of the paper. As network threats evolve, manual NIDS rule engineering has become a critical operational bottleneck This paper presents a human-centered investigation into LLM-based NIDS rule generation Our evaluation reveals a syntax-semantics paradox: although LLMs generate syntactically correct rules, experts find them only partially deployable due to low specificity and logic hallucinations in 12% of cases While the system received a favorable SUS score of 67, practitioners remain skeptical of its autonomous capabilities, viewing LLMs as support tools for drafting and verification rather than independent generators Finally, our statistical analysis indicates that while large-scale models ((≥ 70B) consistently produce syntactically valid rules, small models (≤ 4B) are largely ineffective for IDS rule generation System Formalization and Methodology The research employs a four-phase methodology to investigate LLM-based NIDS rule generation First, the authors conduct a comprehensive literature review of LLM-NIDS intersections, formalizing prevalent system architectures and identifying common inputs Second, they validate the system components and inputs through a pre-study with three domain experts Third, they generate a set of IDS rules using the formalized LLM-based agent system which components are grounded on the outcome of the second phase Finally, they execute a rigorous two-fold evaluation of the generated rule, including a statistical analysis of the generated rules and a user study with security practitioners The system formalization involves defining an interface, workflow, and behavior to ensure a consistent and reproducible basis for evaluation Expert Trust and Semantic Quality Analysis The core finding centers on the "syntax-semantics paradox," where LLMs produce syntactically correct rules but experts deem them only partially deployable due to low specificity and logic hallucinations in 12% of cases Statistical analysis shows that while large-scale models (≥70B) demonstrate high syntactic proficiency, success rates up to 90%, they still require a correction loop for 54.85% of valid rules in the CI1 scenario Furthermore, expert evaluation indicates that only 37.5% of the generated rules received positive assessments, primarily due to insufficient specificity and heavy reliance on content matching Most generated rules rely on pure content matching, with specific fields such as DISTANCE and WITHIN rarely used in the rule bodies Usability and Interaction Patterns The study assesses the usability of the LLM-based system through a user study involving 10 domain experts The system achieved a mean SUS score of 67, satisfying the requirement to consider the system well usable (above 67) Experts rated the interactive generator and PoC finder as highly valuable components for rule automation However, experts rejected full autonomy in favor of a human-in-the-loop workflow grounded in verifiable vulnerability data Experts perceive the system as a high-utility force multiplier for rule drafting and verification, but reject full autonomy in favor of a human-in-the-loop workflow grounded in verifiable vulnerability data Limitations and Future Directions The analysis highlights that while large models are superior for syntactic validity, they struggle to produce rules with the necessary specificity required by SOC experts Hallucinations increase when PoC information is included, correlating with a 10.10% increase in hallucinations (rising from 2.81% to 12.91%) particularly in engines struggling with large context handling Participants tend to stress the need for carefully designing human-driven systems rather than end-to-end automatic rule engineering perspectives Experts seem aware of the limits of generative AI tools for IDS rule engineering, with most using LLMs only sporadically to check syntax validity or provide further insights on already existing rules Ultimately, this work serves as a foundational assessment of LLMs in the rule engineering domain Related Work and System Comparison Existing research explores various LLM-based frameworks, including RuleMaster+

35: and RulePilot

51: , which use PoC data to generate rules The comparative analysis of recent approaches shows that explicit syntax validation and correction mechanisms remain limited, with only a subset of methods incorporating dedicated parsers or grammar checks tailored to IDS frameworks The paper compares different LLM engines using both an aggregation of statistical metrics and an LLM-as-a-judge setting to identify the best performing agent The results show that DeepSeek-R1 70B emerges as the best performing model for both ranking, its relatively lighter nature – w.r.t.

Paper summary: Nadia: So wrapping up this discussion on "Beyond the Syntax: Do Security Experts Trust LLMs for NIDS Rule Engineering?" the paper really digs into why those automated rules aren't always ready to go straight into production <ref:2607.05916#pg1>.

Elias: The authors are showing that while large language models can produce syntactically correct rules, they still struggle with the specificity required by security experts, with a twelve percent failure rate in terms of deployability because of logic hallucinations <ref:2607.05916#pg2>.

Priya: What this means for the world is that we shouldn't treat these LLMs as fully autonomous rule generators right now; they are best seen as powerful tools that help draft and verify rules, not replace the deep expertise of a security analyst <ref:2607.05916#pg3>.

Nadia: Right. The authors highlight that the research focused on formalizing the system to reduce confusion and strengthen what they found about human judgment, interaction patterns, and usability <ref:2607.05916#pg3>.

Elias: They also compared different LLM engines using both statistical metrics and an LLM as a judge setting, finding that DeepSeek-R1 70B performed the best for ranking in this study Related Work and System Comparison.

Priya: Ultimately, the paper suggests that we need to focus on creating human-driven systems where the AI assists with drafting based on verifiable data rather than trying to automate everything end to end <ref:2607.05916#pg3>.

Conclusion: Nadia: So, we're wrapping up our look at "Beyond the Syntax: Do Security Experts Trust LLMs for NIDS Rule Engineering?"

Elias: It boils down to this: LLMs can write rules that look right on paper, but security experts aren't always ready to trust those rules for real-world deployment.

Nadia: That's the core of it. The authors found a syntax-semantics paradox where the AI gets the grammar correct, but it’s often too vague or hallucinates logic that makes them useless in practice.

Priya: From what I see, this means we can't just hand over rule generation to an autonomous system and expect perfect results. The data shows that even big models struggle with specificity.

Elias: Exactly. They found that only about thirty-seven percent of the generated rules were actually good enough for experts to approve, mainly because they relied too much on simple content matching instead of deeper logic.

Nadia: And it's not just the big models that are failing; small models, those under four billion parameters, are largely ineffective for this kind of rule engineering task.

Priya: That’s a huge limitation to keep in mind when we think about scaling these tools across different security teams. The authors stress that the current setup is more useful as a drafting aid than a replacement for an analyst.

Elias: They conclude that the best path forward isn't full autonomy, but rather a human-in-the-loop workflow where the AI helps draft and verify based on known vulnerability data.

Nadia: So, while the technology is getting smarter syntactically, it still needs a lot more help to understand the nuance of what makes a rule actually work in an IDS.

Elias: It really shows that we need to focus on building systems grounded in verifiable facts instead of just letting them guess the logic.

Priya: That leads us right into how these specific results affect our day-to-day work with network defenses.

Sorbonne University · Delft University of Technology

cs.CR

Submitted: 2026-07-07

Updated: 2026-07-07

DOI: 10.1145/3847352.3848108

Code: https://github.com/ollama/ollama

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: As network threats evolve, manual NIDS rule engineering has become a critical operational bottleneck This paper presents a human-centered investigation into LLM-based NIDS rule generation Our…

Key concepts

Syntax-Semantics Paradox
This paradox describes a situation where a system produces output that is technically correct in its structure (syntax) but fails to achieve the desired functional outcome (semantics). In this study, LLMs generate rules that follow the correct format of an IDS rule, but these rules are often too vague or logically flawed for actual deployment by security professionals.
Large Language Models (LLMs) for NIDS Rule Generation
This refers to using advanced AI models, like those with 70B parameters or larger, to automatically write detection rules for network security systems. The study tested whether these models could replace manual rule creation by security analysts in the operational environment.
Syntax vs. Specificity
Syntax refers to the grammatical correctness of a rule—whether it follows the required structure of an IDS language. Specificity refers to how precise and detailed a rule is in identifying threats. The paper found that while LLMs excel at syntax, they often lack the necessary specificity needed for practical security deployment.
Human-in-the-Loop Workflow
This is a proposed system where AI tools assist human experts rather than taking over completely. Experts use the LLM to draft or verify rules, but a human must review and approve them. The study suggests this collaborative approach is preferred over full automation for rule engineering.

Terminology

Summary

As network threats evolve, manual NIDS rule engineering has become a critical operational bottleneck This paper presents a human-centered investigation into LLM-based NIDS rule generation Our evaluation reveals a syntax-semantics paradox: although LLMs generate syntactically correct rules, experts find them only partially deployable due to low specificity and logic hallucinations in 12% of cases While the system received a favorable SUS score of 67, practitioners remain skeptical of its autonomous capabilities, viewing LLMs as support tools for drafting and verification rather than independent generators Finally, our statistical analysis indicates that while large-scale models ((≥ 70B) consistently produce syntactically valid rules, small models (≤ 4B) are largely ineffective for IDS rule generation

System Formalization and Methodology

The research employs a four-phase methodology to investigate LLM-based NIDS rule generation First, the authors conduct a comprehensive literature review of LLM-NIDS intersections, formalizing prevalent system architectures and identifying common inputs Second, they validate the system components and inputs through a pre-study with three domain experts Third, they generate a set of IDS rules using the formalized LLM-based agent system which components are grounded on the outcome of the second phase Finally, they execute a rigorous two-fold evaluation of the generated rule, including a statistical analysis of the generated rules and a user study with security practitioners The system formalization involves defining an interface, workflow, and behavior to ensure a consistent and reproducible basis for evaluation

Expert Trust and Semantic Quality Analysis

The core finding centers on the syntax-semantics paradox, where LLMs produce syntactically correct rules but experts deem them only partially deployable due to low specificity and logic hallucinations in 12% of cases Statistical analysis shows that while large-scale models (≥70B) demonstrate high syntactic proficiency, success rates up to 90%, they still require a correction loop for 54.85% of valid rules in the CI1 scenario Furthermore, expert evaluation indicates that only 37. p. 0

Where the paper says it

While the system received a favorable SUS score of 67, practitioners remain skeptical of its autonomous capabilities, viewing LLMs as support tools for drafting and verification rather than independent generators. Finally, our statistical analysis indicates that while large-scale models ((≥ 70B) consistently produce syntactically valid rules, small models (≤ 4B) are largely ineffective for IDS rule generation.

5% of the generated rules received positive assessments, primarily due to insufficient specificity and heavy reliance on content matching Most generated rules rely on pure content matching, with specific fields such as DISTANCE and WITHIN rarely used in the rule bodies

Usability and Interaction Patterns

The study assesses the usability of the LLM-based system through a user study involving 10 domain experts The system achieved a mean SUS score of 67, satisfying the requirement to consider the system well usable (above 67) Experts rated the interactive generator and PoC finder as highly valuable components for rule automation However, experts rejected full autonomy in favor of a human-in-the-loop workflow grounded in verifiable vulnerability data Experts perceive the system as a high-utility force multiplier for rule drafting and verification, but reject full autonomy in favor of a human-in-the-loop workflow grounded in verifiable vulnerability data

Limitations and Future Directions

The analysis highlights that while large models are superior for syntactic validity, they struggle to produce rules with the necessary specificity required by SOC experts Hallucinations increase when PoC information is included, correlating with a 10.10% increase in hallucinations (rising from 2. p. 0

Where the paper says it

This paper presents a human-centered investigation into LLM-based NIDS rule engineering, formalizing a grounded generation framework and evaluating it through a user study with 10 domain experts. Our evaluation reveals a syntax-semantics paradox: although LLMs generate syntactically correct rules, experts find them only partially deployable due to low specificity and logic hallucinations in 12% of cases.

81% to 12.91%) particularly in engines struggling with large context handling Participants tend to stress the need for carefully designing human-driven systems rather than end-to-end automatic rule engineering perspectives Experts seem aware of the limits of generative AI tools for IDS rule engineering, with most using LLMs only sporadically to check syntax validity or provide further insights on already existing rules Ultimately, this work serves as a foundational assessment of LLMs in the rule engineering domain

Related Work and System Comparison

Existing research explores various LLM-based frameworks, including RuleMaster+ [35] and RulePilot [51], which use PoC data to generate rules The comparative analysis of recent approaches shows that explicit syntax validation and correction mechanisms remain limited, with only a subset of methods incorporating dedicated parsers or grammar checks tailored to IDS frameworks The paper compares different LLM engines using both an aggregation of statistical metrics and an LLM-as-a-judge setting to identify the best performing agent The results show that DeepSeek-R1 70B emerges as the best performing model for both ranking, its relatively lighter nature – w.r.t.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Grounded Generation Framework: Implement a system that formalizes an LLM-based NIDS rule generation system based on literature review and expert pre-study, ensuring a well-defined interface, workflow, and behavior for consistent assessment.

  2. Contextual Rule Input Prioritization: Design the input layer to prioritize vulnerability specifications (CVE descriptions), Proof-of-Concept (PoC) exploits, and network topology information over raw traffic captures when generating rules to maximize utility in real-world environments.

  3. Multi-Stage Validation Pipeline: Integrate a rigorous two-fold validation process where generated rules undergo syntax verification by a native compiler and an automated correction module utilizing adaptive prompting for error fixing before reaching the human loop.

  4. LLM Model Selection Strategy: Employ a statistically driven metric, such as the weighted sum of normalized parameters, correction effort, and hallucination rates across multiple scenarios, to select the best performing LLM engine for generating rules to be tested by domain experts.

  5. Human-in-the-Loop Refinement: Build an interactive system interface that allows domain experts to perform proactive modifications via adaptive prompting, enabling them to mitigate hallucinations and technical inaccuracies, which is critical for achieving semantic correctness.

  6. Usability and Trust Metrics: Evaluate the system using a System Usability Scale (SUS) score, while simultaneously measuring expert trust in an autonomous mode, finding that participants express a near-unanimous consensus on the necessity of a human-in-the-loop over full autonomy.

  7. Specificity Enhancement: Engineer the generation process to encourage the use of specific rule body fields like DISTANCE and WITHIN, as these fields are shown to be rarely used in LLM outputs, thereby addressing the expert complaint that rules lack the level of specificity that SOC experts require.

  8. Handling Cognitive Load: Model network complexity using three scenarios with progressively increasing structural and informational complexity to systematically evaluate how increasing cognitive load affects the rule generation process.

Abstract

As network threats evolve, manual NIDS rule engineering has become a critical operational bottleneck. While Large Language Models (LLMs) show promise for automating this process, their ability to produce production-ready rules remains unvalidated. This paper presents a human-centered investigation into LLM-based NIDS rule engineering, formalizing a grounded generation framework and evaluating it through a user study with 10 domain experts. Our evaluation reveals a syntax-semantics paradox: although LLMs generate syntactically correct rules, experts find them only partially deployable due to low specificity and logic hallucinations in 12% of cases. While the system received a favorable SUS score of 67, practitioners remain skeptical of its autonomous capabilities, viewing LLMs as support tools for drafting and verification rather than independent generators. Finally, our statistical analysis indicates that while large-scale models (at least 70B) consistently produce syntactically valid rules, small models (at most 4B) are largely ineffective for IDS rule generation.

Sources

Related papers