Kurate.org
Rankings
Back to Rankings

LanG 2.0: Energy-Efficient Agentic AI for Semi-Automated Security Operations at Scale

Anes Abdennebi, Nadjia Kara, Laaziz Lahlou, Hakima Ould-Slimane

Oct 29, 2026arXiv:2610.41072v1
cs.AIcs.CR
Share
Scorecard· 16/16
6.0/10 impact

Thorough, honest, practically useful SOC system evaluation with open artifacts, but incremental over the authors' prior platform and with synthetic benchmarks and a narrow frontier comparison.

Abstract

Security operations centers (SOCs) need to turn terabytes of alerts, logs and flows into trusted, affordable decisions to unveil hidden threats. We present LanG 2.0, an agentic system that detects threats, correlates evidence and writes rules, investigating whether small local models can replace frontier models affordably and safely when structural safeguards, not model accuracy, provide the guarantees. It incorporates entity-anchored correlation, human approval of rule changes and a tamper-evident ledger. Tests span eight flow datasets from five networks, real XDR alerts and multistage campaigns. Against its predecessor, campaign correlation precision rises from 18.75% to 51.81%, and rule acceptance from at most 60.00% to 98.33%. Against Claude Opus 5.5, scoring flows against each network's normal traffic detects attacks on unseen networks at 73.42% F1 without false positives, versus 58.10% with 49.00% false positives, for at least 650 times less energy. Finetuned models statistically match it on alert classification and rule writing for 5.6 to 59 times less energy. The work demonstrates LanG's practical scalability, processing 3.33 million records in 6.4 minutes, over 30 million hourly, and its safeguards resisted prompt injection, label poisoning and ledger tampering, so small local models can carry most SOC detection affordably and safely.

AI Impact Assessments

(1 model)

Core Contribution

LanG 2.0 is a systems-plus-evaluation paper on agentic AI for SOCs. Its thesis is that small, locally trained models can carry most of the detection phase affordably and safely, provided *structural safeguards* provide the guarantees rather than model accuracy. The safeguards are:

  • an entity witness required on every correlation edge,
  • analyst approval of every change,
  • a single deploying agent,
  • a hash-chained, externally anchored ledger.

    • The concrete technical contributions are:

    1. Diagnosis of "phase collapse" in the authors' previous multihook correlation graph, fixed by requiring an entity witness on every edge. Benchmark B precision rises from 18.75% to 51.81%.

    2. A per-tenant "site baseline" anomaly detector (rank-transformed features plus kNN) for cross-network detection. The paper shows that supervised flow classifiers do not transfer across NetFlow datasets and that scoring against a network's own normal traffic does.

    3. A rule generator (Qwen3-1.7B LoRA) validated by actual Snort/Suricata/YARA/Sigma engines, with 98.33% acceptance and detection statistically indistinguishable from Claude Opus 5.5.

    4. A preregistered comparison of finetuned 1.7B–8B agents against a frontier model and classical baselines, with an escalation ladder.

    5. Safeguard stress tests (indirect prompt injection, label poisoning, ledger tampering) and per-operation energy accounting.

    Methodological Rigor

    Strengths:

  • Statistical practice is unusually careful: Wilson intervals, cluster bootstrap, paired McNemar tests, Holm correction, and equivalence margins.
  • A preregistered protocol is used for the frontier comparison.
  • Several negative or unflattering findings are reported openly: classical models match the finetuned agents on two of three tasks, the frontier model wins on family naming, union-by-address beats the anchored graph on real alerts, and adaptation raised the false positive rate on one dataset.
  • The ablations are informative, for example the unanchored-hooks ablation and the removal of individual hooks.
  • The ledger tests are systematic, with 200 trials per attack type.

    • Weaknesses:

  • The frontier comparison uses one hosted model, one query per item, and subsets of 217–397 items. The frontier prompt was also minimal, while the local ones were long, so the comparison is not apples-to-apples (acknowledged).
  • Frontier energy is not measured. It is taken from published figures for other models (Gemini, ChatGPT, GPT-4o), so the headline "650× less energy" is a CPU-at-190W-bound versus a hearsay estimate. It is also a per-flow comparison between a trivial kNN scorer and an LLM, which is not a like-for-like pairing.
  • The headline detection comparison (73.42% F1 versus 58.10%) uses 200 flows, with the frontier model prompted with raw flow features. It is a weak baseline, since no tuned unsupervised or supervised baseline was run on those same 200 flows. The strongest framing, "LLMs are bad at flow detection," is already known (cited GPT-4 result).
  • Correlation benchmarks are synthetic. On real GUIDE alerts the proposed method is worse than the simple IP-union baseline, which undercuts the correlation claim.
  • Injection robustness rests on a design where the verdict follows a text-blind detector, so the "resistance" is largely by construction. Adaptive attacks are untested.
  • Several "targeted improvements" were made post hoc, outside the preregistration.
  • Detection on unseen networks varies enormously (0.27% to 92.48%), and the average is dragged by easy datasets.
  • No real SOC deployment and no analyst study; the review-effort reduction is an analytical estimate.

    • Potential Impact

    The most reusable messages are:

  • Per-tenant normal-traffic baselines beat cross-network supervised models.
  • Small finetuned models, or even TF-IDF/LightGBM, suffice for many SOC triage subtasks.
  • Engine-validated rule generation is practical.
  • Structural safeguards can make agentic SOC tooling auditable.

    • The energy-per-operation accounting and the explicit escalation ladder, which says when to use a frontier model, a classical model, or an analyst, are practically useful for SOC architects. The paper is a thorough integration effort with an open-source release, which helps adoption. However, the work is incremental on the authors' own prior platform paper, and its novelty is mostly in engineering and evaluation. Individual components (kNN anomaly scoring, Louvain correlation, hash-chained logs, LoRA finetuning) are standard. Expected influence is moderate and mainly among applied security/AIOps practitioners and "green AI" researchers.

    Timeliness

    The paper is highly timely: agentic SOC automation, cost and energy concerns about LLM inference, prompt-injection risk in security agents, and skepticism about benchmark-inflated NIDS results. Cross-network generalisation failure of supervised flow models is a known but important issue, and the paper adds careful evidence.

    Strengths and Limitations

    Strengths:

  • Honest, statistically careful, broad evaluation (8 flow datasets, real XDR alerts, rule engines).
  • Practical system with scale numbers (3.33M records in 6.4 minutes on one workstation).
  • Reports negative results and scope limits.
  • Code, benchmark, and prompts released; weights only on request.

    • Limitations:

  • Dense, system-heavy presentation that is hard to follow, with much critical material in the supplement.
  • Mixed results: the correlation method does not beat simple baselines on real data, and the family-naming task fails without labels.
  • Frontier-comparison framing is somewhat favorable (energy figures from different models; prompt asymmetry).
  • Single-workstation scale claims rest on filtered suspects and only the top 25 work units per subsystem, so "30 million per hour" is front-stage filtering throughput, not full agentic analysis.
  • The self-citing predecessor makes the contribution delta modest.
  • The safeguards claim is "guarantees," but the evidence is empirical and non-adaptive.

    • Overall: a solid, honest, practically oriented systems/evaluation paper with moderate novelty and a broad but not decisive impact.

    Rating:6/ 10
    Significance 5.5Rigor 6.5Novelty 5Clarity 5.5

    Generated Nov 2, 2026

    Comparison History (0)

    No comparisons yet.