RedCyber: Coupled Adversarial Cyber Operations

• RedCyber is a layered research domain that couples red-team adversarial operations with blue-team responses across strategic, operational, and technical cyber environments.
• It integrates formal game-theoretic models, cyber range assessments, and adversary emulation tools to simulate synthetic conflicts like China–Taiwan cyber campaigns.
• The framework leverages AI assistance, autonomous agents, and human behavior analysis to enhance detection, response, and cyber-physical recovery under continuous contestation.

RedCyber denotes a body of work centered on adversarial cyber operations in which red-team activity, blue-team response, cyber-range instrumentation, intelligence collection, and increasingly AI-driven decision support are treated as a coupled system. In one explicit formulation, RedCyber is a synthetic China–Taiwan cyber conflict used to illustrate a multi-echelon meta-game-theoretic paradigm; across adjacent literature, the term aligns with cyber ranges, red-vs.-blue training, adversary emulation, autonomous attack and defense agents, CTI pipelines, and cyber-physical recovery under continuing adversarial pressure (Yang et al., 10 Sep 2025, Kim et al., 10 Apr 2025, Bianchi et al., 2023, Landauer et al., 2024).

1. Scope, meanings, and doctrinal foundations

In the most explicit formalization, RedCyber is a hypothetical coordinated cyber campaign against Taiwan in which China acts as attacker and Taiwan as defender. The chapter models it as a stage-structured game spanning policy, reconnaissance, disruption, escalation, sustained assault, and strategic messaging, with campaign evolution written as

$$s^{k+1}\!\sim\!T(s^k,a_d^k,a_a^k;\lambda,\theta), \qquad U^O_d=\sum_{k=1}^K u^O_d(s^k,a_d^k,a_a^k,\lambda),$$

where $\lambda$ is the deception index and $\theta$ is the technology lever. The same chapter states that larger $\lambda$ expands belief-shaping options and that improved $\theta$ lowers detection latency and alters feasible action sets, changing equilibria between phases. RedCyber is therefore not presented there as a benchmark or simulator, but as a synthetic conflict embedded in a hierarchy of policy, strategic, operational, tactical, and technical games (Yang et al., 10 Sep 2025).

A broader doctrinal precursor appears in Juan Manuel R. Mosso’s “Ciberseguridad Inteligente”, which proposes Ciberseguridad Inteligente (CSI), a Capacidad de Ciberinteligencia (CCI), and a Modelo de Sistema Ofensivo de Referencia (MSOR). CSI is defined as the capacity to “detectar, denegar, interrumpir, degradar, engañar y destruir las capacidades del adversario en el ciberespacio.” CCI is defined as a strategic organizational resource that transforms indicators and precursors into knowledge, and MSOR is decomposed into Proceso de Surgimiento del Conflicto (PSC), Proceso de Planificación Estratégica (PPE), Proceso de Operaciones en el Ciberespacio (POC), and Proceso de Evaluación de Resultados (PER). Within POC, the operational chain is Análisis de Campañas → Militarización → Validación de blancos → Entrega → Explotación → Instalación → Comando y Control (C2) → Acciones y Objetivos (Mosso, 2015).

This suggests that RedCyber can be read at two levels simultaneously. At the narrow level, it is a named synthetic conflict with formal state transitions and echelon coupling. At the broader level, it denotes a research program concerned with how adversarial cyber campaigns are planned, executed, measured, and countered across strategic, operational, and technical layers.

2. Cyber ranges, adversary emulation, and blue-team assessment

A central RedCyber substrate is the cyber range: a secure environment for Red Team, Blue Team, and often White Team activity. One line of work addresses the scaling problem of post-exercise evaluation. The framework in “Scalable and automated Evaluation of Blue Team cyber posture in Cyber Ranges” defines a four-stage pipeline: Collection of reports from Blue and Red Teams, Definition of Reference/Response Graphs from reports, Automatic evaluation of multiple intermediate scores from defined graphs, and Computation of the final score and the Cyber Posture. It uses MITRE ATT&CK, STIX, and CAPEC, represents reports as a ReportADTree, compares graphs using breadth-first search, and computes Comprehension Score, Defense Score, Implementation Score, and Responsiveness Score before averaging them into a unique Final Score. The paper is explicit that the Red Team’s reported action is the adversarial ground truth against which Blue Team understanding and response are assessed (Bianchi et al., 2023).

A second line of work addresses the operational toolchain of adversary emulation itself. “Red Team Redemption” compares nine open-source adversary emulation tools using an 80-question questionnaire across Installation & Configuration (IC), Community & Support (CS), Documentation (D), Usability (U), and Features & Capabilities (FC), then weights results using an online survey of 20 domain experts conducted between January 15 and March 15, 2024. The study finds MITRE Caldera, Metasploit, and Atomic Red Team on top, and shows that the most important sub-categories across respondents were a broad set of predefined attack procedures (FC-7), the ability to create new procedures (FC-8), and support for automation and chaining (FC-5 and FC-6) (Landauer et al., 2024).

Tool	Brief characterization	Reported position
MITRE Caldera	Browser-based GUI, attack chaining, agent orchestration, cleanup support, JSON/CSV/text/PDF reporting	Top-ranked overall
Metasploit	Massive documentation, broad OS support, flexible payloads, strong community/support	Second overall
Atomic Red Team	18/18 installation score, no agent requirement, large ATT&CK-style test library	Third overall

The range literature also shows that RedCyber is not limited to scoring or tooling. It is concerned with whether blue-team performance is measured against actual red-team behavior, whether attack procedures are repeatable and automatable, and whether exercise infrastructure can scale beyond manual White Team intervention.

3. AI assistance for defenders and operational workflows

A prominent development is the insertion of LLM-based assistance directly into live red-vs.-blue exercises. CyberAlly is designed for a maritime cyber range that emulates the IT and OT systems of a shipping port. It is integrated with Wazuh as SIEM, Slack as the collaboration interface, Cydarm as case management, GPT-4o as the generative model, and LlamaIndex for KG-RAG. Its pipeline is concrete: Wazuh posts alerts to Slack; CyberAlly performs duplicate filtering using sentence embeddings computed as averages of word vectors and cosine distance; non-duplicate alerts are classified by a k-nearest neighbors model with $k=15$; suspicious alerts are enriched via a static graph built from previous Red-vs.-Blue Team events and a dynamic graph built from the current event; then the system produces a Slack message containing an alert summary, recommended actions, and explanation/reasoning, and can create task tickets in Cydarm (Kim et al., 10 Apr 2025).

The quantitative evidence in CyberAlly is concentrated in alert reduction and classification. During one 2-day training event the authors collected more than 60,000 alert messages and observed around 2,500 distinct alert types, excluding background noise. The duplication burden is severe: priority-5 alerts number 1,461 including duplicates but only 78 excluding duplicates, and priority-10 alerts number 612 including duplicates but 16 excluding duplicates. For the benign-versus-suspicious classifier, 10-fold cross-validation reports precision 0.9904, recall 0.9974, and F1 0.9939 with malicious samples upsampled 1x, rising to precision 0.9992, recall 0.9997, and F1 0.9994 at 10x upsampling. The paper is equally explicit about what it does not specify: no detailed ontology, no exact prompt template, no graph database product, and no explicit command-execution guardrails (Kim et al., 10 Apr 2025).

A related operational problem appears in CSOC alert inspection. “Two Can Play That Game” studies an RL-based cyber-alert inspection policy under red-team evaluation in a Cyber Security Operations Center. The queue is tied to operational thresholds: 1175 alerts ≈ 1 hour AvgTTA and 4350 alerts ≈ 4 hours AvgTTA. Under equal attacker and defender budgets, the learned attacker could not drive the defender into the unacceptable region, and with equal budget plus no daily bound 97% of the $500N$ hourly states remained in the green band. But when the authors relaxed an assumption and allowed the attacker finer alert granularity, the attacker used few additional alerts to elicit a disproportionate defender response; after adversarial retraining using a double oracle approach, the retrained defender kept backlog in the green zone for 95% of the $500N$ hours (Shah et al., 2018).

Beyond single-workflow copilots, open cyber LLMs broaden the support layer around RedCyber operations. RedSage is an open-source, locally deployable cybersecurity assistant built from 11.8B tokens of cybersecurity-focused continual pretraining data and 266K multi-turn cybersecurity samples for supervised fine-tuning. It introduces RedSage-Bench with 30K multiple-choice and 240 open-ended Q&A items, and at the 8B scale reports gains of up to +5.59 points on cybersecurity benchmarks and +5.05 points on Open LLM Leaderboard tasks. The paper positions it as a generalist assistant covering frameworks, offensive techniques, CTI, CLI tooling, and incident-response-relevant workflows, while emphasizing local deployment over proprietary APIs (Suryanto et al., 29 Jan 2026).

4. Autonomous red and blue agents

Autonomous RedCyber research has been shaped by the CybORG family of environments. The 2020 paper introduces CybORG as a research gym for Autonomous Cyber Operations (ACO) with both simulation and emulation, a common interface, YAML scenario definitions, and explicit support for blue team and red team decision-making models in adversarial settings. The environment is designed so that the same scenario and agent logic can operate in a fast FSM-based simulator or in a higher-fidelity emulator on virtual infrastructure, and the red-team use case is a capture-the-flag / penetration-testing task involving exploitation, credential gathering, and pivoting through a network (Baillie et al., 2020).

The 2021 CybORG paper makes the red side concrete. In its demonstration scenario, the red agent’s goal is to get a session on the Internal host as the System user. Available actions are SSH Bruteforce, Portscan, Pingsweep, Upgrade to Meterpreter, IPConfig, MS17-010-PSExec, Autoroute, and Sleep. The agent is trained in simulation with DQN and LSTM under partial observability, then evaluated in emulation; across 21 trained RL agents and 10 emulator trials each, the paper reports 139 successful tests, i.e. a 66% rate of success. The failure analysis is equally important: some agents overfit simulation artifacts and failed in emulation because they did not use autoroute to route around the firewall (Standen et al., 2021).

Multi-agent red-vs.-blue training appears in “A Multiagent CyberBattleSim for RL Cyber Operation Agents.” The MARLon extension adds an AttackerWrapper, DefenderWrapper, and simultaneous alternating-turn training to CyberBattleSim. A key result is methodological: in an undefended network, an A2C red agent achieved cumulative reward 359 when invalid actions were assigned reward 0, compared with 283 when invalid actions were redirected with no modification, while a -1 invalid-action penalty made the agent effectively conservative. When joint training was repaired by removing premature reset on blue availability violations, a PPO MARL red agent scored 309.7 against no blue and 96.4 against a PPO MARL blue agent, showing that joint training strengthened blue defense without completely destroying red competence (Kunz et al., 2023).

A complementary direction infers red behavior from defender-side observables rather than training the red policy directly. “Learning Red Agent Policy from Observations for Neurosymbolic Autonomous Cyber Agents” models the environment as

$\mathcal{M} := (S, A, O, T, R),$

and predicts hidden red actions from $o^b_{t-1}$, $\lambda$0, and $\lambda$1 using a latent-action imitation pipeline integrated into an Evolving Behavior Tree. In CybORG CAGE Challenge 2, the method reports subnet-level prediction accuracies of 99.82% for $\lambda$2, 97.49% for $\lambda$3, and 95.85% for $\lambda$4, with corresponding host-level accuracies of 95.28%, 77.71%, and 83.83%. This places attacker-policy inference inside a neurosymbolic blue agent rather than in a purely offline analytics layer (Samaddar et al., 16 Jun 2026).

Agentic LLMs extend autonomous red behavior into interactive terminal use. RedTeamLLM defines a seven-component architecture—Launcher, RedTeamAgent, Memory Manager, ADaPT Enhanced, Plan Corrector, ReAct, and Planner—and evaluates a proof of concept centered on summarizing, reasoning and act with GPT-4o. On five VulnHub targets, the paper states that RedTeamLLM outperforms PentestGPT in 3 out of 5 use cases; reasoning reduces tool calls in 4 out of 5 use cases and improves achieved steps in 4 of 5 use cases. The headline architectural claims about memory management and plan correction are broader than the validated implementation, but the ablation on reasoning provides direct evidence that explicit deliberation improves terminal-centric offensive workflows (Challita et al., 11 May 2025).

5. Intelligence from online discussion and network-scale telemetry

RedCyber research also treats the online discourse surrounding cyber operations as an intelligence surface. “Early Warnings of Cyber Threats in Online Discussions” combines hourly collection from 69 well known and reliable Twitter experts with a crawler covering nearly 200 darkweb and deepweb sites. From September 1, 2016 to January 31, 2017, the system generated 661 alerts, of which about 84% were judged legitimate cyber-threats by majority vote of five annotators, and the full workflow averaged 0.6 seconds. Its strongest case study is Mirai: warnings for “mirai” were generated 49 days before the October 21 Dyn-related DDoS disruption, and from October 1 onward the warnings were reinforced by darkweb/deepweb references to exploit code sharing (Sapienza et al., 2018).

Lexicon induction and forum classification generalize this CTI logic across platforms. “Detecting Cyber-Related Discussions in Online Social Platforms” builds a multi-platform corpus of 1,070,882 documents from Twitter, Reddit, Stack Overflow, Cyber War News, and The Hacker News, ranks terms using TF-IDF, FDR, and APMIS, and proposes

$\lambda$5

The reported evaluation is internally uneven but still informative: FDR achieves precision 85%, while APMIS achieves recall 78% and F1 67%, indicating a precision–recall trade-off between conservative filtering and broader cyber-discussion capture (Ikwu et al., 2019).

Dark-web classification in a specific linguistic ecosystem is addressed by “Identification of Malicious Posts on the Dark Web Using Supervised Machine Learning.” The paper constructs three Brazilian Portuguese datasets from Hidden Answers and Deep Answer(s), labels posts through a multi-stage process combining IoCs, contextual keywords, and manual review, and trains several classical models. The best-performing configuration is LightGBM + TF-IDF Unigram, which on the final labeled dataset reaches Accuracy 94%, Precision 74%, Recall 83%, and F1-score 79%. When applied to 7,498 new posts, it marks 1,158 posts (15%) as relevant and further stratifies them into 595 (8%) High Relevance, 849 (11%) Medium Relevance, and 6,054 (81%) Low Relevance (Filho et al., 28 Nov 2025).

At larger scale, RedCyber intelligence also includes enriched network observatories. “Large Scale Enrichment and Statistical Cyber Characterization of Network Traffic” fuses CAIDA UCSD Network Telescope data with GreyNoise enrichment using Python GraphBLAS and PyD4M, anonymizes source IPs with CryptoPAN, and characterizes feature prevalence through sparse associative arrays and heavy-tail models. The paper reports that the enriched variables are consistent with Zipf-Mandelbrot forms such as

$\lambda$6

with fitted exponents around 1.75 and 2.0 for source-packet counts, ports, and ASNs. Operationally, the dominant classes are concentrated: the classification label malicious accounts for fractions 0.587, 0.588, 0.639, and 0.569 across four snapshots, and among records with CVE labels, CVE-2017-0144 accounts for fractions 0.906, 0.936, 0.882, and 0.892. The paper’s “exemplar” adversary is an unknown actor, using Windows 7/8, sending 64–127 packets, over TCP/445, targeting EternalBlue-related behavior (Kawaminami et al., 2022).

Taken together, these works show that RedCyber intelligence is heterogeneous by design: expert social media, dark-web forums, multilingual text classification, and enriched telescope traffic all function as complementary sensors rather than substitutes.

6. Human behavior, cyber-physical recovery, and future directions

A distinctive recent development is the explicit treatment of human attacker cognition. The GAMBiT project releases three human-subject red-team cyber range datasets, HSR1, HSR2, and HSR3, collected between July 2024 and March 2025. Across the experiments, 19–20 skilled attackers per experiment each conducted two 8-hour sessions, producing a combined 944 participant-hours on the SimSpace Cyber Force Platform. Data include self-reports, psychometrics, operational notes, terminal histories, keylogs, PCAP, Suricata alerts, and curated derivative labels. The experiments embed triggers for loss aversion, base-rate neglect, availability heuristic, confirmation bias, and sunk-cost fallacy, making the dataset directly relevant to deception, attacker-behavior modeling, and bounded-rationality analysis (Beltz et al., 28 Aug 2025).

RedCyber also extends into cyber-physical restoration. “Cyber Recovery from Dynamic Load Altering Attacks” studies recovery after a Dynamic Load Altering Attack (DLAA) by jointly optimizing repair crew routes and adaptive adjustment of system operation across electricity, transportation, and cyber networks. The resulting MILP prioritizes resetting emergency droop gains while repairing all compromised loads. In the modified IEEE 39-bus case study, the integrated CRDA resets all IBR droop gains to zero at time step 5, whereas the decoupled benchmark requires 5 additional time steps; under the stated operating-cost assumption this yields about $\lambda$7 of savings. The contribution is specifically recovery-centric: detection and mitigation are assumed to have occurred already, and the problem is how to restore secure normalcy while the attacker may still exploit remaining compromised resources (Liu et al., 2023).

At the strategic end, the RedCyber chapter concludes with directions for resilience, cross-echelon planning, and the evolving role of AI in cyber warfare. Read together with GAMBiT and CRDA, this suggests that RedCyber research is expanding in three directions at once: downward into operator cognition and human-subject telemetry, outward into interdependent critical infrastructure recovery, and upward into multi-echelon theories in which policy, strategy, operations, tactics, and technical implementations are recursively linked (Yang et al., 10 Sep 2025).

RedCyber is therefore best understood not as a single artifact but as a layered research domain. It encompasses cyber-range training and assessment, adversary emulation tooling, AI assistance for blue teams, autonomous red and blue agents, discourse- and telemetry-driven intelligence, human attacker behavior, and recovery of cyber-physical systems under ongoing contestation. The common thread is that red activity is not merely background noise or a benchmark label; it is the organizing reference against which defense, resilience, and strategic adaptation are designed and evaluated.