Agentic AI for Cybersecurity: A Meta-Cognitive Architecture for Governable Autonomy

Andrei Kojukhov; Arkady Bovshover

arxiv: 2602.11897 · v3 · submitted 2026-02-12 · 💻 cs.CR · cs.AI

Agentic AI for Cybersecurity: A Meta-Cognitive Architecture for Governable Autonomy

Andrei Kojukhov , Arkady Bovshover This is my paper

Pith reviewed 2026-05-16 03:06 UTC · model grok-4.3

classification 💻 cs.CR cs.AI

keywords cybersecurityagentic AImeta-cognitionuncertainty managementmulti-agent systemsSOARintrusion detectiondecision calibration

0 comments

The pith

A meta-cognitive agent framework for cybersecurity improves robustness by judging uncertainty before committing to action.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a probabilistic architecture that decomposes cybersecurity tasks into specialized agents for detection, hypothesis formation, contextualization, explanation, and governance. These agents are overseen by a meta-cognitive judgement mechanism that evaluates uncertainty, disagreements, and constraints to select responses such as automated action, escalation, deferral, or further evidence gathering. Traditional deterministic SOAR systems are limited in uncertain or adversarial settings; this approach aims to produce more reliable and accountable decisions. Evaluations on CICIDS2017 and NSL-KDD datasets with added noise show gains in accuracy, lower false positives, and better-calibrated confidence compared to single-agent and baseline methods. A sympathetic reader would care because the work reframes security automation as a form of governable cognitive problem-solving rather than rigid rule-based triggering.

Core claim

The central claim is that cybersecurity orchestration can be modeled as a meta-cognitive process in which interacting agents are coordinated by a judgement mechanism that monitors uncertainty, agent disagreement, and operational constraints to determine decision readiness and enable adaptive strategies including automated action, escalation, deferral, and evidence refinement, yielding higher accuracy under noise, reduced false positives, and better-calibrated confidence on benchmark datasets.

What carries the argument

The meta-cognitive judgement mechanism, which evaluates uncertainty, agent disagreement, and operational constraints to determine decision readiness and select among adaptive response strategies.

If this is right

The system achieves higher accuracy under noise than deterministic or single-agent baselines.
It reduces false positive rates while producing better-calibrated confidence estimates.
It supports adaptive strategies such as deferral and escalation when conditions warrant.
It enables more accountable autonomy and effective human-AI collaboration in adversarial settings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar meta-cognitive layers could be tested in other uncertain domains such as medical diagnosis or autonomous driving to check transferability.
Detailed specifications of how the judgement mechanism weighs inputs would be required for independent replication or scaling to live networks.
The architecture might reduce the volume of escalations reaching human operators by handling routine uncertainty internally.

Load-bearing premise

That the meta-cognitive judgement mechanism can be implemented to reliably assess uncertainty, disagreements, and constraints in a way that produces the claimed improvements.

What would settle it

A head-to-head test on the same augmented CICIDS2017 or NSL-KDD datasets in which the proposed framework shows no accuracy gain, no reduction in false positives, or worse calibration than deterministic baselines when uncertainty or adversarial manipulation is increased.

read the original abstract

Cybersecurity decision-making increasingly occurs in environments characterized by uncertainty, partial observability, and adversarial manipulation, where heterogeneous signals from multiple sources are often incomplete, ambiguous, or conflicting. Traditional Security Orchestration, Automation, and Response (SOAR) systems rely on deterministic pipelines and threshold-based triggers, limiting their ability to support reliable decision-making under such conditions. This paper proposes a probabilistic, agentic framework for cybersecurity orchestration that models decision-making as a meta-cognitive process. The framework decomposes cybersecurity functions into interacting agents responsible for detection, hypothesis formation, contextualization, explanation, and governance, coordinated through a meta-cognitive judgement mechanism. This mechanism evaluates uncertainty, agent disagreement, and operational constraints to determine decision readiness, enabling adaptive strategies including automated action, escalation, deferral, and evidence refinement. Empirical evaluation on benchmark datasets (CICIDS2017 and NSL-KDD), augmented with adversarial and uncertain conditions, demonstrates that the proposed approach improves robustness and decision quality compared to deterministic and single-agent baselines. The framework achieves higher accuracy under noise, reduces false positive rates, and produces better-calibrated confidence estimates, while enabling more adaptive and context-aware decision behavior. By explicitly modeling meta-cognitive processes - monitoring, evaluation, control, and reflection - the proposed approach reframes cybersecurity as an instance of AI-mediated cognitive problem solving, supporting accountable autonomy and more effective human-AI collaboration in adversarial environments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper sketches a multi-agent meta-cognitive setup for cybersecurity orchestration but supplies no implementation details or numbers to support its performance claims.

read the letter

The paper's main contribution is a proposed decomposition of cybersecurity decision-making into five interacting agents—detection, hypothesis formation, contextualization, explanation, and governance—plus a meta-cognitive judgement layer that weighs uncertainty, disagreement, and constraints to pick among action, escalation, deferral, or more evidence gathering. This is framed as a way to move beyond the rigid pipelines in current SOAR systems when signals are noisy or adversarial. The structure gives a clean way to think about accountable autonomy in uncertain environments, and the authors correctly identify partial observability and conflicting inputs as core problems that deterministic thresholds handle poorly. That framing is useful and shows honest engagement with real operational limits. The choice to test on augmented versions of CICIDS2017 and NSL-KDD is also reasonable for the domain. What the paper does not deliver is any concrete evidence. The abstract states that the approach yields higher accuracy under noise, lower false positives, and better-calibrated confidence than deterministic and single-agent baselines, yet no numbers, error bars, ablation results, or description of the judgement mechanism appear. There are no equations, pseudocode, uncertainty model, or disagreement metric. Without those, it is impossible to tell whether any reported gains come from the architecture or from unstated implementation choices. The stress-test note is accurate on this point. The work is aimed at researchers exploring agentic methods for security orchestration who want a high-level blueprint rather than a finished method. A reader seeking reproducible results or validated improvements will find it thin. I would bring it to a reading group to discuss the agent decomposition and how meta-cognition might be realized in practice, but I would not cite it until the missing details are supplied. It shows clear thinking on the problem, so it qualifies as serious engagement with the literature. Even so, the absence of any supporting data or mechanism means it does not yet deserve referee time; the authors would need to add the implementation specifics and actual results before a serious editor should send it out.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a probabilistic agentic framework for cybersecurity orchestration that decomposes decision-making into specialized agents (detection, hypothesis formation, contextualization, explanation, governance) coordinated by a meta-cognitive judgement mechanism. This mechanism evaluates uncertainty, agent disagreement, and operational constraints to select adaptive strategies (automated action, escalation, deferral, evidence refinement). The central claim is that evaluations on noise- and adversarial-augmented CICIDS2017 and NSL-KDD benchmarks demonstrate improved accuracy, reduced false-positive rates, and better-calibrated confidence estimates relative to deterministic and single-agent baselines.

Significance. If the empirical claims can be substantiated with reproducible implementation details, the work could advance governable autonomy in adversarial domains by reframing cybersecurity as meta-cognitive problem-solving. The explicit modeling of monitoring, evaluation, control, and reflection offers a structured path toward accountable human-AI collaboration. However, the current absence of technical specification limits immediate significance and reproducibility.

major comments (2)

[Abstract / Framework] Abstract and framework description: the meta-cognitive judgement mechanism is presented only in prose without equations, pseudocode, uncertainty quantification method, disagreement metric, or decision rule. This is load-bearing for the central claim that the mechanism produces the reported gains in accuracy, FPR, and calibration; without it, improvements cannot be attributed to the architecture rather than unstated implementation choices.
[Empirical Evaluation] Empirical evaluation: the abstract asserts higher accuracy under noise, lower false-positive rates, and improved calibration on augmented CICIDS2017/NSL-KDD, yet supplies no numerical results, error bars, ablation studies, baseline implementations, or data-handling details. This prevents assessment of whether the claimed robustness holds.

minor comments (1)

[Abstract] The abstract would benefit from explicit statement of the quantitative improvements (e.g., accuracy deltas or calibration scores) rather than qualitative descriptors.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments that identify key areas where additional formalization and empirical detail are needed to strengthen the manuscript. We address each major comment below and will make the corresponding revisions.

read point-by-point responses

Referee: [Abstract / Framework] Abstract and framework description: the meta-cognitive judgement mechanism is presented only in prose without equations, pseudocode, uncertainty quantification method, disagreement metric, or decision rule. This is load-bearing for the central claim that the mechanism produces the reported gains in accuracy, FPR, and calibration; without it, improvements cannot be attributed to the architecture rather than unstated implementation choices.

Authors: We agree that the meta-cognitive judgement mechanism requires formal specification. In the revised manuscript we will add a mathematical formulation of the mechanism, including explicit equations for uncertainty quantification (entropy over agent posteriors), a disagreement metric (pairwise Jensen-Shannon divergence), and the decision rules that map these quantities plus operational constraints to the four adaptive strategies. Pseudocode for the full coordination loop will also be included. revision: yes
Referee: [Empirical Evaluation] Empirical evaluation: the abstract asserts higher accuracy under noise, lower false-positive rates, and improved calibration on augmented CICIDS2017/NSL-KDD, yet supplies no numerical results, error bars, ablation studies, baseline implementations, or data-handling details. This prevents assessment of whether the claimed robustness holds.

Authors: We acknowledge the absence of quantitative results and implementation details. The revised version will include tables reporting accuracy, false-positive rate, and expected calibration error (with standard deviations over repeated runs) on the noise- and adversarial-augmented CICIDS2017 and NSL-KDD datasets. Ablation studies isolating the meta-cognitive component, descriptions of the deterministic and single-agent baselines, and full details on data augmentation, preprocessing, and hyper-parameters will be added. revision: yes

Circularity Check

0 steps flagged

No circularity: conceptual framework evaluated on external benchmarks

full rationale

The paper proposes a high-level probabilistic agentic framework decomposing cybersecurity functions into agents coordinated by a meta-cognitive judgement mechanism, with no equations, derivations, fitted parameters, or mathematical reductions presented. Performance claims rest on empirical comparisons to external benchmark datasets (CICIDS2017 and NSL-KDD) and deterministic/single-agent baselines under augmented conditions, which are independent of the paper's own definitions or inputs. No self-citations, uniqueness theorems, or ansatzes are invoked to support core claims, rendering the architecture self-contained as a novel descriptive proposal.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the domain assumption that cybersecurity environments are dominated by uncertainty and conflicting signals, plus the introduction of new agent roles and a meta-cognitive coordinator whose effectiveness is asserted without independent external validation beyond the benchmark claims.

axioms (1)

domain assumption Cybersecurity decision-making occurs in environments characterized by uncertainty, partial observability, and adversarial manipulation where signals are incomplete, ambiguous, or conflicting.
Explicitly stated in the opening of the abstract as the setting that traditional SOAR systems cannot handle.

invented entities (1)

Meta-cognitive judgement mechanism no independent evidence
purpose: Evaluates uncertainty, agent disagreement, and operational constraints to determine decision readiness and select among automated action, escalation, deferral, or evidence refinement.
New coordinating component introduced to manage the agent interactions; no independent falsifiable evidence outside the paper's claims is provided.

pith-pipeline@v0.9.0 · 5554 in / 1411 out tokens · 117805 ms · 2026-05-16T03:06:56.436990+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 1 internal anchor

[1]

Introduction Cybersecurity has undergone a profound transformation over the past decade. Once dominated by signature-based detection and rule-driven automation, modern cyber defence increasingly relies on machine learning, large-scale data analytics, and automated response mechanisms to cope with the growing volume and velocity of security events. These a...

work page
[2]

This metaphor assumes that decision-making can be decomposed into independent steps and that uncertainty can be resolved locally at each stage

Theoretical Foundations and Meta-Cognitive Framework 2.1 From Linear Pipelines to Distributed Cognition Traditional cybersecurity architectures are rooted in a pipeline metaphor [1,2]: data flows through predefined stages, each performing a narrowly defined function. This metaphor assumes that decision-making can be decomposed into independent steps and t...

work page
[3]

The architecture consists of five interacting agent classes coordinated through an orchestration layer powered by generative AI, with humans integrated as active participants

Architecture and Framework: Agentic Cybersecurity Orchestration 3.1 Architectural Overview Building on the theoretical framing, we propose an Agentic Cybersecurity Orchestration Framework that operationalizes cybersecurity as a multi-agent cognitive system. The architecture consists of five interacting agent classes coordinated through an orchestration la...

work page
[4]

Survey papers systematically organize this literature by algorithm, dataset, and performance metric

Related Work 4.1 AI in Cybersecurity: Model-Centric Foundations Extensive prior research has explored the application of machine learning and deep learning to cybersecurity tasks, including intrusion detection, malware classification, phishing detection, and anomaly detection [1,2]. Survey papers systematically organize this literature by algorithm, datas...

work page
[5]

The objective is to assess how agentic orchestration with meta- cognitive judgement improves decision-making under uncertainty compared to traditional pipeline-based approaches

Empirical Evaluation of Agentic Cybersecurity Orchestration To complement the conceptual and architectural contribution, we evaluate the proposed framework using a combination of benchmark cybersecurity datasets and controlled simulation scenarios. The objective is to assess how agentic orchestration with meta- cognitive judgement improves decision-making...

work page
[6]

Deterministic SOAR Pipeline o rule-based orchestration o threshold-triggered actions

work page
[7]

Single-Agent ML System o centralized classifier o no explicit uncertainty integration

work page
[8]

Detection Performance under Noise Figure 1 shows detection accuracy under increasing noise levels (0%–30%)

Proposed Agentic Orchestration Framework o multi-agent belief aggregation o meta-cognitive decision function o adaptive decision strategies 5.5 Evaluation Metrics We evaluate both predictive and cognitive decision properties: Detection Metrics • accuracy • precision / recall / F1 Decision Metrics • correct action rate • false action rate Cognitive Metrics...

work page
[9]

This shift has important implications for both system design and operational practice

Discussion The proposed framework reframes cybersecurity from a pipeline-oriented automation problem to a distributed cognitive process involving artificial and human agents. This shift has important implications for both system design and operational practice. First, the results suggest that many limitations of current cybersecurity systems are not prima...

work page
[10]

Conclusion This paper proposed a conceptual re-framing of cybersecurity orchestration as an agentic, multi-agent cognitive system. By making explicit the roles of interpretation, explanation, governance, and meta-cognitive judgement, the framework moves beyond pipeline-centric architectures toward reflective and accountable autonomy. The introduction of a...

work page
[11]

In: 2024 IEEE 25st Int

Sommer, R.; Paxson, V. Outside the Closed World: On Using Machine Learning for Network Intrusion Detection. Proc. IEEE Symp. Security and Privacy (SP) 2010, 305–316. https://doi.org/10.1109/SP .2010.25

work page doi:10.1109/sp 2010
[12]

On the Effectiveness of Machine and Deep Learning for Cyber Security

Apruzzese, G.; Colajanni, M.; Ferretti, L.; Guido, A.; Marchetti, M. On the Effectiveness of Machine and Deep Learning for Cyber Security. In Proc. 10th Int. Conf. on Cyber Conflict (CyCon); 2018

work page 2018
[13]

An Introduction to MultiAgent Systems, 2nd ed.; Wiley: Chichester, UK, 2009

Wooldridge, M. An Introduction to MultiAgent Systems, 2nd ed.; Wiley: Chichester, UK, 2009

work page 2009
[14]

Cognition in the Wild; MIT Press: Cambridge, MA, USA, 1995

Hutchins, E. Cognition in the Wild; MIT Press: Cambridge, MA, USA, 1995. https://doi.org/10.7551/mitpress/1881.001.0001

work page doi:10.7551/mitpress/1881.001.0001 1995
[15]

Joint Cognitive Systems: Foundations of Cognitive Systems Engineering; CRC Press/Taylor & Francis: Boca Raton, FL, USA, 2005

Hollnagel, E.; Woods, D.D. Joint Cognitive Systems: Foundations of Cognitive Systems Engineering; CRC Press/Taylor & Francis: Boca Raton, FL, USA, 2005

work page 2005
[16]

Human–Machine Reconfigurations: Plans and Situated Actions, 2nd ed.; Cambridge University Press: Cambridge, UK, 2012

Suchman, L. Human–Machine Reconfigurations: Plans and Situated Actions, 2nd ed.; Cambridge University Press: Cambridge, UK, 2012. https://doi.org/10.1017/CBO9780511808418

work page doi:10.1017/cbo9780511808418 2012
[17]

Explanation in Artificial Intelligence: Insights from the Social Sciences

Miller, T. Explanation in Artificial Intelligence: Insights from the Social Sciences. Artif. Intell. 2019, 267, 1–38

work page 2019
[18]

Towards A Rigorous Science of Interpretable Machine Learning

Doshi-Velez, F.; Kim, B. Towards a Rigorous Science of Interpretable Machine Learning. arXiv 2017, arXiv:1702.08608

work page internal anchor Pith review Pith/arXiv arXiv 2017
[19]

Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell

Bender, E.M.; Gebru, T.; McMillan-Major, A.; Shmitchell, S. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proc. ACM Conf. on Fairness, Accountability, and Transparency (FAccT ’21); 2021; pp. 610–623. https://doi.org/10.1145/3442188.3445922

work page doi:10.1145/3442188.3445922 2021
[20]

2023, doi: 10.6028/NIST.AI.100-1

National Institute of Standards and Technology (NIST). Artificial Intelligence Risk Management Framework (AI RMF 1.0); NIST: Gaithersburg, MD, USA, 2023. https://doi.org/10.6028/NIST.AI.100-1

work page doi:10.6028/nist.ai.100-1 2023
[21]

Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1); NIST ‘‘NVLAP’’: 2024

National Institute of Standards and Technology (NIST). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1); NIST ‘‘NVLAP’’: 2024

work page 2024
[22]

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)

European Union. 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). OJ L 2024, 2024/1689

work page 2024
[23]

Metamemory: A Theoretical Framework and New Findings

Nelson, T.O.; Narens, L. Metamemory: A Theoretical Framework and New Findings. In Psychology of Learning and Motivation; Bower, G.H., Ed.; Academic Press: San Diego, CA, USA, 1990; Volume 26, pp. 125–173

work page 1990
[24]

Metacognition in Computation: A Selected Research Review

Cox, M.T. Metacognition in Computation: A Selected Research Review. Artificial Intelligence 2005, 169, 104–141. https://doi.org/10.1016/j.artint.2005.10.009

work page doi:10.1016/j.artint.2005.10.009 2005
[25]

AI4People—An Ethical Framework for a Good AI Society

Floridi, L.; Cowls, J.; Beltrametti, M.; Chatila, R.; Chazerand, P .; Dignum, V.; Luetge, C.; Madelin, R.; Pagallo, U.; Rossi, F.; Schafer, B.; Valcke, P.; Vayena, E. AI4People—An Ethical Framework for a Good AI Society. Minds Mach. 2018, 28, 689–707. https://doi.org/10.1007/s11023-018-9482-5

work page doi:10.1007/s11023-018-9482-5 2018
[26]

Artificial Intelligence in Cybersecurity: A Comprehensive Multidomain Review of Techniques, Applications, Challenges, and Future Directions

Rafy, A.; Ahmed, M. Artificial Intelligence in Cybersecurity: A Comprehensive Multidomain Review of Techniques, Applications, Challenges, and Future Directions. Electronics 2023, 12, 4040. https://doi.org/10.3390/electronics12194040

work page doi:10.3390/electronics12194040 2023
[27]

The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines

Vinay, V. The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines. arXiv 2025, arXiv:2512.06659

work page arXiv 2025
[28]

Hammers and others , title =

Jestus Lazer, S.; Aryal, K.; Gupta, M.; Bertino, E. A Survey of Agentic AI and Cybersecurity: Challenges, Opportunities and Use-Case Prototypes. arXiv 2026, arXiv:2601.05293

work page arXiv 2026
[29]

Securing Agentic AI Systems — A Multilayer Security Framework

Arora, S.; Hastings, J. Securing Agentic AI Systems — A Multilayer Security Framework. arXiv 2025, arXiv:2512.18043

work page arXiv 2025
[30]

Sentinel Agents for Secure and Trustworthy Agentic AI in Multi-Agent Systems

Gosmar, D.; Dahl, D.A. Sentinel Agents for Secure and Trustworthy Agentic AI in Multi-Agent Systems. arXiv 2025, arXiv:2509.14956

work page arXiv 2025
[31]

Metacognitive AI: Framework and the Case for a Neurosymbolic Approach

Wei, H.; Shakarian, P .; Lebiere, C.; Draper, B.; Krishnaswamy, N.; Sreedharan, S. Metacognitive AI: Framework and the Case for a Neurosymbolic Approach. In Proc. Metacognitive AI Workshop; Springer, 2024

work page 2024

[1] [1]

Introduction Cybersecurity has undergone a profound transformation over the past decade. Once dominated by signature-based detection and rule-driven automation, modern cyber defence increasingly relies on machine learning, large-scale data analytics, and automated response mechanisms to cope with the growing volume and velocity of security events. These a...

work page

[2] [2]

This metaphor assumes that decision-making can be decomposed into independent steps and that uncertainty can be resolved locally at each stage

Theoretical Foundations and Meta-Cognitive Framework 2.1 From Linear Pipelines to Distributed Cognition Traditional cybersecurity architectures are rooted in a pipeline metaphor [1,2]: data flows through predefined stages, each performing a narrowly defined function. This metaphor assumes that decision-making can be decomposed into independent steps and t...

work page

[3] [3]

The architecture consists of five interacting agent classes coordinated through an orchestration layer powered by generative AI, with humans integrated as active participants

Architecture and Framework: Agentic Cybersecurity Orchestration 3.1 Architectural Overview Building on the theoretical framing, we propose an Agentic Cybersecurity Orchestration Framework that operationalizes cybersecurity as a multi-agent cognitive system. The architecture consists of five interacting agent classes coordinated through an orchestration la...

work page

[4] [4]

Survey papers systematically organize this literature by algorithm, dataset, and performance metric

Related Work 4.1 AI in Cybersecurity: Model-Centric Foundations Extensive prior research has explored the application of machine learning and deep learning to cybersecurity tasks, including intrusion detection, malware classification, phishing detection, and anomaly detection [1,2]. Survey papers systematically organize this literature by algorithm, datas...

work page

[5] [5]

The objective is to assess how agentic orchestration with meta- cognitive judgement improves decision-making under uncertainty compared to traditional pipeline-based approaches

Empirical Evaluation of Agentic Cybersecurity Orchestration To complement the conceptual and architectural contribution, we evaluate the proposed framework using a combination of benchmark cybersecurity datasets and controlled simulation scenarios. The objective is to assess how agentic orchestration with meta- cognitive judgement improves decision-making...

work page

[6] [6]

Deterministic SOAR Pipeline o rule-based orchestration o threshold-triggered actions

work page

[7] [7]

Single-Agent ML System o centralized classifier o no explicit uncertainty integration

work page

[8] [8]

Detection Performance under Noise Figure 1 shows detection accuracy under increasing noise levels (0%–30%)

Proposed Agentic Orchestration Framework o multi-agent belief aggregation o meta-cognitive decision function o adaptive decision strategies 5.5 Evaluation Metrics We evaluate both predictive and cognitive decision properties: Detection Metrics • accuracy • precision / recall / F1 Decision Metrics • correct action rate • false action rate Cognitive Metrics...

work page

[9] [9]

This shift has important implications for both system design and operational practice

Discussion The proposed framework reframes cybersecurity from a pipeline-oriented automation problem to a distributed cognitive process involving artificial and human agents. This shift has important implications for both system design and operational practice. First, the results suggest that many limitations of current cybersecurity systems are not prima...

work page

[10] [10]

Conclusion This paper proposed a conceptual re-framing of cybersecurity orchestration as an agentic, multi-agent cognitive system. By making explicit the roles of interpretation, explanation, governance, and meta-cognitive judgement, the framework moves beyond pipeline-centric architectures toward reflective and accountable autonomy. The introduction of a...

work page

[11] [11]

In: 2024 IEEE 25st Int

Sommer, R.; Paxson, V. Outside the Closed World: On Using Machine Learning for Network Intrusion Detection. Proc. IEEE Symp. Security and Privacy (SP) 2010, 305–316. https://doi.org/10.1109/SP .2010.25

work page doi:10.1109/sp 2010

[12] [12]

On the Effectiveness of Machine and Deep Learning for Cyber Security

Apruzzese, G.; Colajanni, M.; Ferretti, L.; Guido, A.; Marchetti, M. On the Effectiveness of Machine and Deep Learning for Cyber Security. In Proc. 10th Int. Conf. on Cyber Conflict (CyCon); 2018

work page 2018

[13] [13]

An Introduction to MultiAgent Systems, 2nd ed.; Wiley: Chichester, UK, 2009

Wooldridge, M. An Introduction to MultiAgent Systems, 2nd ed.; Wiley: Chichester, UK, 2009

work page 2009

[14] [14]

Cognition in the Wild; MIT Press: Cambridge, MA, USA, 1995

Hutchins, E. Cognition in the Wild; MIT Press: Cambridge, MA, USA, 1995. https://doi.org/10.7551/mitpress/1881.001.0001

work page doi:10.7551/mitpress/1881.001.0001 1995

[15] [15]

Joint Cognitive Systems: Foundations of Cognitive Systems Engineering; CRC Press/Taylor & Francis: Boca Raton, FL, USA, 2005

Hollnagel, E.; Woods, D.D. Joint Cognitive Systems: Foundations of Cognitive Systems Engineering; CRC Press/Taylor & Francis: Boca Raton, FL, USA, 2005

work page 2005

[16] [16]

Human–Machine Reconfigurations: Plans and Situated Actions, 2nd ed.; Cambridge University Press: Cambridge, UK, 2012

Suchman, L. Human–Machine Reconfigurations: Plans and Situated Actions, 2nd ed.; Cambridge University Press: Cambridge, UK, 2012. https://doi.org/10.1017/CBO9780511808418

work page doi:10.1017/cbo9780511808418 2012

[17] [17]

Explanation in Artificial Intelligence: Insights from the Social Sciences

Miller, T. Explanation in Artificial Intelligence: Insights from the Social Sciences. Artif. Intell. 2019, 267, 1–38

work page 2019

[18] [18]

Towards A Rigorous Science of Interpretable Machine Learning

Doshi-Velez, F.; Kim, B. Towards a Rigorous Science of Interpretable Machine Learning. arXiv 2017, arXiv:1702.08608

work page internal anchor Pith review Pith/arXiv arXiv 2017

[19] [19]

Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell

Bender, E.M.; Gebru, T.; McMillan-Major, A.; Shmitchell, S. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proc. ACM Conf. on Fairness, Accountability, and Transparency (FAccT ’21); 2021; pp. 610–623. https://doi.org/10.1145/3442188.3445922

work page doi:10.1145/3442188.3445922 2021

[20] [20]

2023, doi: 10.6028/NIST.AI.100-1

National Institute of Standards and Technology (NIST). Artificial Intelligence Risk Management Framework (AI RMF 1.0); NIST: Gaithersburg, MD, USA, 2023. https://doi.org/10.6028/NIST.AI.100-1

work page doi:10.6028/nist.ai.100-1 2023

[21] [21]

Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1); NIST ‘‘NVLAP’’: 2024

National Institute of Standards and Technology (NIST). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1); NIST ‘‘NVLAP’’: 2024

work page 2024

[22] [22]

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)

European Union. 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). OJ L 2024, 2024/1689

work page 2024

[23] [23]

Metamemory: A Theoretical Framework and New Findings

Nelson, T.O.; Narens, L. Metamemory: A Theoretical Framework and New Findings. In Psychology of Learning and Motivation; Bower, G.H., Ed.; Academic Press: San Diego, CA, USA, 1990; Volume 26, pp. 125–173

work page 1990

[24] [24]

Metacognition in Computation: A Selected Research Review

Cox, M.T. Metacognition in Computation: A Selected Research Review. Artificial Intelligence 2005, 169, 104–141. https://doi.org/10.1016/j.artint.2005.10.009

work page doi:10.1016/j.artint.2005.10.009 2005

[25] [25]

AI4People—An Ethical Framework for a Good AI Society

Floridi, L.; Cowls, J.; Beltrametti, M.; Chatila, R.; Chazerand, P .; Dignum, V.; Luetge, C.; Madelin, R.; Pagallo, U.; Rossi, F.; Schafer, B.; Valcke, P.; Vayena, E. AI4People—An Ethical Framework for a Good AI Society. Minds Mach. 2018, 28, 689–707. https://doi.org/10.1007/s11023-018-9482-5

work page doi:10.1007/s11023-018-9482-5 2018

[26] [26]

Artificial Intelligence in Cybersecurity: A Comprehensive Multidomain Review of Techniques, Applications, Challenges, and Future Directions

Rafy, A.; Ahmed, M. Artificial Intelligence in Cybersecurity: A Comprehensive Multidomain Review of Techniques, Applications, Challenges, and Future Directions. Electronics 2023, 12, 4040. https://doi.org/10.3390/electronics12194040

work page doi:10.3390/electronics12194040 2023

[27] [27]

The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines

Vinay, V. The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines. arXiv 2025, arXiv:2512.06659

work page arXiv 2025

[28] [28]

Hammers and others , title =

Jestus Lazer, S.; Aryal, K.; Gupta, M.; Bertino, E. A Survey of Agentic AI and Cybersecurity: Challenges, Opportunities and Use-Case Prototypes. arXiv 2026, arXiv:2601.05293

work page arXiv 2026

[29] [29]

Securing Agentic AI Systems — A Multilayer Security Framework

Arora, S.; Hastings, J. Securing Agentic AI Systems — A Multilayer Security Framework. arXiv 2025, arXiv:2512.18043

work page arXiv 2025

[30] [30]

Sentinel Agents for Secure and Trustworthy Agentic AI in Multi-Agent Systems

Gosmar, D.; Dahl, D.A. Sentinel Agents for Secure and Trustworthy Agentic AI in Multi-Agent Systems. arXiv 2025, arXiv:2509.14956

work page arXiv 2025

[31] [31]

Metacognitive AI: Framework and the Case for a Neurosymbolic Approach

Wei, H.; Shakarian, P .; Lebiere, C.; Draper, B.; Krishnaswamy, N.; Sreedharan, S. Metacognitive AI: Framework and the Case for a Neurosymbolic Approach. In Proc. Metacognitive AI Workshop; Springer, 2024

work page 2024