The findings show that LLMs can approximate structured cybersecurity reasoning under controlled representations, but do not apply it robustly, which has important implications for the design and evaluation of AI-assisted security decision-support systems.
Abstract
Large language models (LLMs) are increasingly used in cybersecurity workflows, yet it remains unclear whether they can perform structured security reasoning or merely rely on superficial cues and prior knowledge. We study this question in the context of defence selection over attack graphs derived from real-world threat scenarios, including ransomware, supply-chain compromise, cloud abuse, Kubernetes attacks, POS malware, and ICS/OT intrusion. Given a budget constraint, LLMs must select security controls to minimise attacker success. We compare their strategies against each other and against a game-theoretic optimization baseline used as a normative reference for structured reasoning. Our results show that LLMs exhibit conditional competence. When explicit attack-graph structure is provided, they often produce coherent strategies close to the optimization baseline. However, their capabilities are fragile. LLM behaviour becomes increasingly fragile with graph complexity and is highly sensitive to framing. Small prompt changes can substantially alter rankings, and merely relabeling a poor strategy as ``optimal''dramatically improves its evaluation. We further observe a non-monotonic relationship between formal risk and LLM judgement: strategies closest to the optimum are not necessarily ranked highest by LLM evaluators. To further probe reasoning ability, we ask LLMs to generate solvers for the same optimization problem. While the generated implementations recover the correct high-level formulation, they scale poorly compared to a purpose-built solver. Overall, our findings show that LLMs can approximate structured cybersecurity reasoning under controlled representations, but do not apply it robustly. This has important implications for the design and evaluation of AI-assisted security decision-support systems.
The growing complexity and frequency of cyberattacks make cybersecurity risk assessment an increasingly demanding task for organisations, requiring substantial expertise, resources, and adherence to established standards. This work explores the applicability of Large Language Model (LLM) to cybersecurity risk assessment, with a focus on threat identification and risk scoring. The paper presents a standalone consistency analysis across five models, measuring accuracy and stability under lexical, structural, and noisy prompt perturbations using an OWASP-oriented rubric. Building on the analysis results, we present a modular LLM-based system that combines Retrieval-Augmented Generation, MITRE ATT&CK-Aligned threat evaluation, rubric-constrained risk scoring, and a Judge Reviewer, orchestrated through a Beliefs–Desires–Intentions control loop. The validation against incidents from the VERIS and EuRepoC datasets highlights limitations and weaknesses, and allows identifying the architectural and structural mitigations that can reduce prompt sensitivity in LLM-based risk assessment.
Large Language Model (LLM) agents are increasingly adopted in real-world security operations with access to host artifacts and command-line interfaces (CLIs), making it critical to thoroughly assess their security capabilities. However, existing cybersecurity benchmarks focus on pre-compromise settings where agents are placed in a clean and idealized environment before an attack occurs. This leaves the post-compromise setting underexplored. To address this gap, we introduce SecRespond, the first benchmark for evaluating LLM agents on the post-compromise incident-response workflow. Given a forensic disk snapshot of a compromised host together with the alerts, vulnerability scans, and baseline checks reported by a host security product, agents are required to produce forensic reports on intrusions, baseline risks, and vulnerability risks, together with a remediation plan. We instantiate this task across 10 cyber ranges, each constructed from a distinct compromised cloud host, spanning 4 entry-point types, 21 ATT&CK techniques, and 5 operating systems. We evaluate 23 frontier LLMs on the OpenCode agent harness. Experimental results show that although current agents can reliably uncover the problems exposed by alerts, they struggle to proactively investigate the disk for silent intrusions and to produce comprehensive, verified remediation plans, with no model achieving complete detection and remediation on any single range. This reveals a fundamental bottleneck in building agents for real-world incident response. The benchmark is publicly available at https://github.com/Alibaba-NLP/qqr/tree/main/data/secrespond.
Lehan Wang, Boli Chen, Ruixue Ding et al.· 0 citations
With the increasingly aggressive cyber threat landscape for governments, businesses, and institutions, as information and/or cybersecurity implementations are increasingly under scrutiny by regulators, it has been pointed out that governance failure is one of the major reasons for a weakened cybersecurity posture. A major component of Cyber/information security governance is the development, adoption, and implementation of a comprehensive information and/or cyber security policy document. The policy document must be in compliance with international or national standards and, if possible, with regulatory guidelines. However, it is often observed that policy documents are often incomplete with respect to industry standards or regulations and require revision when subjected to a thorough audit. Identifying the gaps between the controls and processes documented in the policy and those required in the regulations or standards necessitates extensive manual effort. The advent of Generative AI tools such as Large Language Models (LLMs) led to use of LLMs and Agentic AI tools to automate such compliance checks, as seen in a few research publications in recent times. However, such reported use of LLMs are experimented with high resource environments such as expensive GPUs and memory based servers. For smaller organizations such expensive compute platform may not be easily available. In this article, we benchmark the compliance checking tasks on LLMs that do not require GPU and high memory usage and the effectiveness of such resource constrained LLMs in compliance checking. Our experiments demonstrated that the low resource LLMs can provide good agreement/accuracy in compliance checking of policy documents against standards by experimenting with ISO 27002:2022 controls against multiple policy documents.
R. Negi, Rishika Jain, Soumyo V Chakarborty et al.· 0 citations
Most vulnerability pipelines remain predictioncentric: they output scores or labels and defer decisions to engineers, even when outputs are compressed, imbalanced, or unreliable under Continuous Integration and Continuous Deployment (CI/CD) shift. We introduce AEGIS (Autonomous Enhanced Guardian for Intelligent Security), a policy-governed framework that treats vulnerability management as a constrained CI/CD decision process in which learned signals serve as evidence and are translated into admissible actions under explicit constraints. AEGIS combines graph-based risk estimation, epistemic uncertainty via stochastic inference, and a symbolic policy guard that maps evidence to auditable decisions: Block, Warn, and Pass. A key design principle is separation of concerns: perception estimates risk while governance determines admissible actions. Irreversible automation is permitted only when risk is high and uncertainty is low, while uncertain cases are deferred to controlled review. This separation makes it possible to revise policy thresholds, cost assumptions, and review budgets without retraining the perception model. We evaluate AEGIS on an extreme-imbalance patch stream used as a stress-test setting and an expanded multi-project dataset that enables more stable estimation of decision outcomes. The evaluation reports the policy thresholds, model settings, symbolic predicates, ablations, and sensitivity settings used in policy replay. Results provide preliminary evidence that policygoverned control supports more interpretable decision behavior under uncertainty, while enabling controlled trade-offs between automation, safety, and review load.
Imad Abdallah, Lunjin Lu· Annual International Compute...· 0 citations
To improve cybersecurity across industries, Cyber Threat Intelligence (CTI) is becoming increasingly crucial. This systematic review explores how CTI practices are evolving in response to advancements in Artificial Intelligence (AI), particularly in the context of Large Language Models (LLMs). We examined 61 peer-reviewed studies using the PRISMA methodology, which demonstrates a strict selection procedure founded on specified inclusion, exclusion, and quality standards. This approach aligns with the scope of similar systematic reviews in the field of cyber threat intelligence. The review provides a comparative synthesis of CTI research capabilities across threat detection and prediction, attribution, forecasting, and automated reporting. We classify these approaches into three categories: conventional methods, those enhanced by AI and Machine Learning, and those based on LLMs. Our findings indicate that LLMs offer significant advantages in contextual reasoning, processing unstructured threat intelligence, and generating actionable mitigation plans. However, challenges such as model explainability, data privacy, system interoperability, and standardization impede their integration into operational environments. In addition to highlighting the potential and practical limitations of LLMs in CTI, this study identifies research gaps and proposes methods to create scalable, secure, and flexible CTI systems that support real-time cyber defense.
Hilalah Alturkistani, Abdul Ghafar Jaafar, S. Chuprat et al.· International journal of res...· 0 citations
AI agents are rapidly improving in cybersecurity capabilities when the source code is available for analysis, yet much of the software most consequential to cybersecurity, including malware, firmware, and proprietary applications, is available only as binaries. Analyzing such software requires reverse engineering(RE): recovering program semantics before the analysis can be meaningfully performed. However, evaluating agentic RE poses a fundamental challenge: benchmark instances must be unseen as source code in the LLMs'training data to prevent models from taking shortcuts by recognizing them rather than really analyzing them, while also matching the scale and anti-analysis protections of real software. Unfortunately, however, existing benchmarks do not jointly satisfy these requirements. To this end, we introduce SRE-Bench, the first realistic, contamination-free RE benchmark. Built entirely from scratch by RE experts with over 5,000 hours, SRE-Bench comprises 19 private, real-world-scale programs averaging 16.9K lines of code. We further developed 44 in-house anti-analysis primitives, yielding 262 binary instances and 1572 deterministically graded tasks. Our evaluation across five frontier LLMs (GPT-5.6-sol,Claude-Opus-5,GPT-5.5,Grok-4.5, and GLM-5.2) shows that RE remains largely unsolved: the strongest model, GPT-5.6-sol, scores 61.4% per instance, and fully solves only 31.5% of the instances. Our analysis further reveals that agents behave differently from human engineers, where agents are relatively insensitive to compiler optimization and static linking. Controlled ablations also confirm that both contamination control and realistic scale are essential. These results indicate that strong source-code security capabilities do not yet transfer to binary analysis, highlighting RE as an important frontier for agentic cybersecurity and SRE-Bench as a rigorous testbed to measure progress.
Jeremy Spence, Nicholas Assaderaghi, Jinhao Zhu et al.· 0 citations
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.