The Proceedings of the ACM on Networking (PACMNET) series showcases top-tier research in emerging computer networks and their applications. We welcome submissions introducing new technologies, innovative experiments, creative applications of networking technologies, and fresh insights gained through analysis. Supported by the ACM Special Interest Group on Communications and Computer Networks (SIGCOMM), the journal is backed by a distinguished Editorial Board composed of leading researchers in the field. This issue constitutes the third issue of the fourth volume of PACMNET and features articles submitted to the December 2025 deadline that underwent a “one-shot major” revision. Originally, more than 90 Editors evaluated each of the December 2025 submissions, coordinated by two Associate Editors. In the initial phase, every article received a minimum of three reviews. For those that advanced to the second phase, the Editors produced at least two additional reviews per article. After a second discussion phase, the Editors met online to decide which articles to accept after a minor revision with shepherding, which to offer a one-shot major revision opportunity, and which to reject. Thirteen articles were offered the one-shot major revision option. Authors received about 3 months to revise their work and prepare a major revision addressing reviewers’ comments and feedback. The same Editors then reviewed and discussed the revised version. After this re-review phase, three articles were rejected, and ten articles were finally accepted and appear in this issue of PACMNET. The articles in this issue showcase the breadth of contemporary networking research, spanning networked systems, transport protocols, Internet measurement, security, and data-plane innovation. Contributions advance the performance and efficiency of modern communication infrastructures through new approaches to forward error correction, congestion control, in-network acceleration, SmartNIC architectures, and high-speed stream processing, while also improving the robustness of transport protocols and programmable hardware. The issue further highlights progress in Internet measurement and operations, including large-scale studies of IPv6 scanning, BGP community semantics, and congestion control in virtualized environments, alongside research on machine learning– driven network security, adaptive intrusion detection, and privacy compliance. Collectively, these works underscore the growing convergence of networking, systems design, measurement, and artificial intelligence in building more efficient, resilient, and trustworthy networked infrastructures. As usual, we would like to express our sincere gratitude to all those who contributed to this issue of PACMNET, especially the Authors for submitting their finest work and the Associate Editors for offering valuable feedback in their reviews and engaging in the discussions. Our thanks also go to the SIGCOMM Executive Committee Chair and the CoNEXT Steering Committee members for their continued support and guidance, providing essential suggestions and insights throughout the article selection process.

Large Language Model (LLM) agents are increasingly proposed to automate offensive security tasks, with recent studies reporting near human-level success rates in Capture-the-Flag (CTF) challenges. We here revisit these results, providing a second look at these claims. We engineer different agent architectures of increasing complexity and modularity on 30 web-based CTFs challenges spanning 14 vulnerability classes. We instantiate these agents with multiple LLM backbones, and compare them with claude-code, a general-purpose agent that automatically determines its internal architecture. Our evaluation yields three main findings. First, claude-code achieves performance comparable to the engineered architectures (19/30 solved tasks), suggesting that general-purpose agents are strong baselines for offensive security tasks. Second, both our architectures and claude-code struggle in the same challenge categories, revealing persistent barriers that keep current agents below human-level capability. Third, by leveraging our manually designed architectures we can systematically measure the impact of additional components, finding that structured orchestration of specialized roles outperforms monolithic designs, improving run-to-run consistency, and reducing execution costs.

The Proceedings of the ACM on Networking (PACMNET) series showcases top-tier research in emerging computer networks and their applications. We welcome submissions introducing new technologies, innovative experiments, creative applications of networking technologies, and fresh insights gained through analysis. Supported by the ACM Special Interest Group on Communications and Computer Networks (SIGCOMM), the journal is backed by a distinguished Editorial Board composed of leading researchers in the field.

Honeypots are deception systems that emulate vulnerable services to collect threat intelligence. While deploying many honeypots increases the opportunity to observe attacker behaviour, in practise network and computational resources limit the number of honeypots that can be exposed. Hence, practitioners must select the assets to deploy, a decision that is typically made statically despite attackers' tactics evolving over time. This work investigates an AI-driven agentic architecture that autonomously manages honeypot exposure in response to ongoing attacks. The proposed agent analyses Intrusion Detection System (IDS) alerts and network state to infer the progression of the attack, identify compromised assets, and predict likely attacker targets. Based on this assessment, the agent dynamically reconfigures the system to maintain attacker engagement while minimizing unnecessary exposure. The approach is evaluated in a simulated environment where attackers execute Proof-of-Concept exploits for known CVEs. Preliminary results indicate that the agent can effectively infer the intent of the attacker and improve the efficiency of exposure under resource constraints

A network telescope is a range of IP addresses that host no services. Millions of bots and scanners contact it to look for vulnerable systems, and the traffic it exposes is fundamental to understanding malicious activities. The visibility a telescope offers depends on its size and geolocation, and merging the information from multiple telescopes could help increase visibility and uncover more malicious activities. However, sharing raw telescope data is complicated, calling for solutions that allow one to directly share the knowledge rather than the data obtained from multiple deployments. In this paper, we explore the application of Federated Learning (FL) to create and share such global knowledge from the malicious activities seen in distributed telescopes. For that, we introduce FedScope, an FL-based solution for generating host embeddings in a distributed way. We compare FedScope to local and distributed alternatives in downstream tasks, such as sender classification or coordinated activities detection. We show that FedScope 1) produces embeddings of equal or higher quality than those of a single telescope; 2) increases coverage, allowing the global model to monitor more malicious actors; 3) avoids the sharing of the raw data, limiting exchanged data.

Satellite Communication (SatCom) offers internet connectivity where traditional infrastructures are too expensive to deploy. When using satellites in a geostationary orbit, the distance from Earth forces a round-trip time of at least 550 ms. Coupled with the constrained capacity of the physical link, this challenges the traditional internet access quality we are used to.In this paper, we present a complete passive characterization of the traffic carried by an operational SatCom provider. With this unique vantage point, we observe the performance of the SatCom technology, as well as the usage habits of subscribers in different countries in Europe and Africa. We highlight the implications of such technology on Internet usage and functioning, and we pinpoint technical challenges due to the CDN and DNS resolution issues, while discussing possible optimizations that the ISP could implement to improve the service offered to SatCom subscribers. We complete the characterization of the adoption and performance of newer protocols with a focus on IPv6 and QUIC.

Traffic visibility remains a key component for management and security operations. Observing unsolicited and erroneous traffic, such as unanswered traffic or errors, is fundamental to detect misconfiguration, temporary failures or attacks. ChamaleoNet transforms any production network into a transparent monitor to let administrators collect unsolicited and erroneous traffic directed to hosts, whether offline or active, hosting a server or a client, protected by a firewall, or unused addresses. ChamaleoNet is programmed to ignore well-formed traffic and collect only erroneous packets, including those generated by misconfigured or infected internal hosts, and those sent by external actors which scan for services. Engineering such a system poses several challenges, from scalability to privacy. Leveraging the SDN paradigm, ChamaleoNet processes the traffic flowing through a campus/corporate network and focuses on erroneous packets only, lowering the pressure on the collection system while respecting privacy regulations by design. ChamaleoNet enables the seamless integration with active deceptive systems like honeypots that can impersonate unused hosts/ports/services and engage with senders. The SDN in-hardware filtering reduces the traffic to the controller by 96

Machine learning has given encrypted traffic classification a new momentum. Yet, once deployed, models often fail due to hidden shortcut features, i.e., spurious correlations learned from training data that do not hold in new environments. Prior work has shown their negative impact through costly manual intervention. Here, we present ShortcutCatcher, an automated, model-agnostic framework that detects and mitigates shortcuts with the help of explainable AI. The key idea is to contrast model behaviour on two datasets: a large training dataset and a separate verification dataset that differs in scenario but shares the same feature schema. ShortcutCatcher integrates feature explanation with cross-scenario evaluation in a closed loop, iteratively removing those critical features that would not be valid in deployment. Across multiple encrypted traffic classification tasks and model architectures, ShortcutCatcher uncovers shortcut dependencies and improves cross-scenario generalisation, up to three times over standard training. In addition, ShortcutCatcher exposes dataset limitations where collection artefacts act as silent shortcuts that have gone so far unnoticed, allowing us to finally expose realistic performance without assuming that the underlying task is intrinsically easy

In an era of effortless data collection, the impact of machine learning - especially neural networks (NNs) - is undeniable. As datasets grow in size and complexity, efficiently handling mixed data types, including categorical and numerical features, becomes critical. Feature encoding and selection play a key role in improving NN performance, efficiency, interpretability, and generalisation. This paper presents GLEm-Net (Grouped Lasso with Embeddings Network), a novel NN-based approach that seamlessly integrates feature encoding and selection directly into the training process. GLEm-Net uses embedding layers to process categorical features with high cardinality, simplifying the model and improving generalisation. By extending the grouped Lasso regularisation to explicitly consider categorical features, GLEm-Net automatically identifies the most relevant features during training and returns them to the analyst. We evaluate GLEm-Net on open and proprietary industry datasets and compare it to state-of-the-art feature selection methodologies. Results show that GLEm-Net adapts to each dataset by allowing the NN to directly select subsets of most important features, offering on par performance with the best state-of-the-art feature selection methods, while eliminating the need for the external feature encoding and selection steps that are now incorporated in the NN training stage.

The Microsoft 365 (M365) ecosystem hosts thousands of third-party applications that integrate with enterprise tenants via fine-grained OAuth permissions, potentially granting access to sensitive organisational resources such as emails, files, calendars, chats, and user directories. Despite the security implications of these permission grants, the M365 ecosystem has not been systematically studied. We present the first privacy- and security-oriented measurement of M365 third-party applications. By combining public marketplace APIs with automated tenant-side deployment, we crawl over 8,000 applications. We find that only 1,069 of them expose both descriptions and permission sets, with significant inconsistencies in transparency across official distribution channels. Next, we leverage a topic-aware anomaly detection framework to assess whether requested permissions align with declared functionality. We cluster applications via Neural Topic Modelling and apply unsupervised anomaly detection within each topic to identify deviations from peer permission profiles. LLM-assisted analysis of the most anomalous cases and blind manual inspection reveal a correlation between anomalous permission profiles and the risk associated with the requested permissions. We find that many applications request overly broad tenant-wide scopes (e.g., directory-wide read/write access), violating least-privilege principles and increasing the organisational attack surface. Our pipeline provides tenant administrators with actionable insights by identifying anomalous applications and the permissions that most contribute to their anomalousness. Overall, our findings expose systemic opacity and structural immaturity in the M365 app ecosystem, where permission disclosure is inconsistent and over-privileged access is common.

Input length and the position of relevant information are widely cited as the primary causes of degraded LLM long-context performance. Here, we study lexical density – the rate at which a context introduces distinct information – as a third, largely overlooked factor that systematically reduces the effective context window of LLMs. We quantify the impact of lexical density on open-weight LLMs (9B-685B) using three "find-the-needle" style benchmarks with identical length ( 12k tokens) and controlled needle position, but increasing density of information. We observe a sharp performance collapse in higher-density benchmarks: models that are near-perfect in sparse contexts drop below 60

In recent years, cybersecurity threats have increasingly exploited human behaviour rather than purely technical vulnerabilities, exposing the limits of traditional awareness programmes delivered outside real-world contexts. To bridge this gap, we introduce TrainShield, an interaction paradigm for contextual cybersecurity training that embeds adaptive learning interventions directly within user workflows. The system integrates real-time risk detection (e.g., phishing and data loss prevention) with event-triggered hypermedia overlays that dynamically connect users to context-specific learning nodes embedded within their browsing workflow to deliver personalised micro-learning content and structured feedback tailored to the user's knowledge level and current context. This approach operationalises behavioural theories by transforming security incidents into immediate learning opportunities, shifting users from automatic to reflective decision-making at critical moments. We further formalise a design model that maps detected events to adaptive training instances, combining user modelling, context extraction, and large language model (LLM)-based content generation. A preliminary study indicates that the approach is perceived as useful in increasing risk awareness and is preferred over lengthy and asynchronous traditional training formats, while also highlighting challenges in aligning generated content with user expectations. Overall, the results suggest that embedding contextual, event-driven training within everyday interactions is a promising direction for behaviour-oriented cybersecurity education.

The complexity and scale of Internet attacks call for distributed, cooperative observatories capable of monitoring malicious traffic across diverse networks. Holoscope is an open, lightweight, and cloud-native platform designed to simplify the deployment and management of telescope (passive) and honeypot (active) sensors. Built upon K3s and WireGuard, Holoscope offers secure connectivity, automated sensor onboarding, and resilient operation even in resource-constrained environments. Through modular design and Infrastructure-as-Code principles, it supports dynamic sensor orchestration, automated recovery, and data processing. We build, deploy, and operate Holoscope across multiple institutions and cloud networks in Europe and Brazil, enabling unified visibility into large-scale attack phenomena while maintaining ease of integration and security compliance.

Passive measurement has traditionally focused on inbound traffic to detect malicious activity, based on the assumption that threats originate externally. In this paper, we offer a complementary perspective by examining outbound traffic, and argue that a narrow subset – what we term erroneous outbound traffic – is a lighter and revealing yet overlooked data source for identifying a broad range of security threats and network problems. This traffic consists of packets sent by internal hosts that either receive no response, trigger ICMP errors, or are ICMP error messages themselves generated in response to unsolicited requests. To demonstrate its potential, we collect and analyse erroneous traffic from a large network, uncovering a variety of previously unnoticed issues, including misconfigurations, obsolete deployments and compromised hosts.

The use of ML in cybersecurity has long been impaired by generalization issues: Models that work well in controlled scenarios fail to maintain performance in production. The root cause often lies in ML algorithms learning superficial patterns (shortcuts) rather than underlying cybersecurity concepts. We investigate contrastive multi-modal learning as a first step towards improving ML performance in cybersecurity tasks. We aim at transferring knowledge from data-rich modalities, such as text, to data-scarce modalities, such as payloads. We set up a case study on threat classification and propose a two-stage multi-modal contrastive learning framework that uses textual vulnerability descriptions to guide payload classification. First, we construct a semantically meaningful embedding space using contrastive learning on descriptions. Then, we align payloads to this space, transferring knowledge from text to payloads. We evaluate the approach on a large-scale private dataset and a synthetic benchmark built from public CVE descriptions and LLM-generated payloads. The methodology appears to reduce shortcut learning over baselines on both benchmarks. We release our synthetic benchmark and source code as open source.

Web services, as the most ubiquitous form of online services, have consistently attracted research attention due to privacy concerns. Although VPNs and anonymous communication methods can partially protect users’ online privacy, advancements in website fingerprinting (WF) attacks still exploit the spatio-temporal characteristics of web resource transmission to identify web services. The challenge lies in defending against WF attacks efficiently, with limited bandwidth costs. Server-side WF defenses, deployed on web servers, can achieve end-to-end obfuscation across both clients and servers. However, existing defenses often consume significant bandwidth and require additional removal operations on the client side. Given the growing use of QUIC with HTTP/3 and the need for robust privacy protections, this paper introduces an asymmetric server-side WF defense scheme using State-Transition Adversarial Perturbations (STAP). STAP introduces the concept of latent resource-state transitions, which represent hidden patterns in resource transmission. Utilizing perturbation models containing these transitions, STAP subtly alters traffic through packet padding and insertion, with inherent transport layer encryption enhancing the concealment. STAP can operate independently, removing the necessity for user involvement. Experimental results demonstrate that STAP outperforms other schemes, achieving reductions in True Positive Rate (TPR) by up to 22% and reductions in bandwidth overhead by up to 30%.

This paper investigates the capabilities and limitations of Large Language Model (LLM) agents in performing cybersecurity forensic tasks, including incident response, digital evidence correlation, and threat attribution. To enable a fair comparison of agents and LLMs, we introduce CFA-bench, a novel benchmark designed to evaluate their forensic reasoning abilities. We leverage a controlled testbed where vulnerable services are instantiated, attacked, and monitored, generating forensic evidence in the form of packet captures and log traces. Using this setup, we generate 20 curated incidents targeting 13 distinct services, focusing on recent vulnerabilities. Each incident presents progressively complex checkpoints, culminating in the identification of the specific Common Vulnerabilities and Exposure (CVE). We evaluate different LLM-powered agent architectures, equipping them with essential forensic tools such as a PCAP Reader and an Information Retriever. Each agent is asked to analyse the incidents to systematically track their performance across different forensic checkpoints. While preliminary, our findings demonstrate the potential of LLM agents in cybersecurity forensics, revealing their strengths and critical areas for improvement. This study underscores the need for standardized benchmarks to assess LLM agents in cyber threat analysis rigorously. For this, we make CFA-bench open to the research community. Our results provide a foundation for future research aimed at refining agent architectures and enhancing their forensic reasoning capabilities.

The Proceedings of the ACM on Networking (PACMNET) series showcases top-tier research in emerging computer networks and their applications. We welcome submissions introducing new technologies, innovative experiments, creative applications of networking technologies, and fresh insights gained through analysis. Supported by the ACM Special Interest Group on Communications and Computer Networks (SIGCOMM), the journal is backed by a distinguished Editorial Board composed of leading researchers in the field. This issue begins the third volume of PACMNET. It features 6 articles, all submitted by the June 2024 deadline when 121 submissions in total were received. Each submission underwent a thorough review process involving over 80 Editors, coordinated by two Associate Editors. In the initial phase, every article received a minimum of three reviews. Following an online discussion, roughly half of the submissions were rejected, while the other half advanced to a second review phase. In this phase, Editors produced at least two additional reviews per article. After further discussion and remote Editors' meeting, 8 articles were given one-shot major revision. The same Editors reviewed the revised version the Authors prepared, and 6 out of 8 articles were finally selected. These 6 articles appear in this issue. Topics include network support for large language models and deep learning, security, and wireless networking. All papers include a thorough set of experiments to validate the proposed solutions. From a methodological perspective, machine learning and artificial intelligence-based solutions are becoming central in developing novel networking solutions. We want to express our gratitude to all those who contributed to this issue of PACMNET, especially the Authors for submitting their finest work and the Associate Editors for offering valuable feedback in their reviews and engaging in the discussions. Our thanks also go to the SIGCOMM Executive Committee Chair and the CoNEXT Steering Committee members for their continued support and guidance, providing essential suggestions and insights throughout the article selection process.

LLM agents show promise for vulnerability testing. We however lack benchmarks to evaluate and compare solutions. AutoPenBench covers this need offering an open benchmark for the evaluation of vulnerability testing agents. It includes 33 tasks, ranging from introductory exercises to actual vulnerable systems. It supports MCP, enabling the comparison of agent capabilities. We introduce milestones per task, allowing the comparison of intermediate steps where agents struggle. To illustrate the use of AutoPenBench we evaluate autonomous and human-assisted agent architectures. The former achieves 21% success rates, insufficient for production, while human-assisted agents reach 64% success, indicating a viable industrial path. AutoPenBench is offered as open source and enables fair comparison of agents.