Matteo Boffa

Giovanni Dettori, Matteo Boffa, Danilo Giordano et al.

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% retrieval score on denser ones. To rule out task-type confounds, we vary and control the density within each benchmark while keeping all other properties unchanged. Reducing density generally restores performance, especially in the high-density regimes where degradation appears. These results show that effective context capacity is a function of lexical density, with direct implications for real-world LLM systems operating on compact, information-rich inputs.

Matteo Boffa, Rodolfo Vieira Valentim, Luca Vassio et al.

The collection of security-related logs holds the key to understanding attack behaviors and diagnosing vulnerabilities. Still, their analysis remains a daunting challenge. Recently, Language Models (LMs) have demonstrated unmatched potential in understanding natural and programming languages. The question arises whether and how LMs could be also useful for security experts since their logs contain intrinsically confused and obfuscated information. In this paper, we systematically study how to benefit from the state-of-the-art in LM to automatically analyze text-like Unix shell attack logs. We present a thorough design methodology that leads to LogPrécis. It receives as input raw shell sessions and automatically identifies and assigns the attacker tactic to each portion of the session, i.e., unveiling the sequence of the attacker's goals. We demonstrate LogPrécis capability to support the analysis of two large datasets containing about 400,000 unique Unix shell attacks. LogPrécis reduces them into about 3,000 fingerprints, each grouping sessions with the same sequence of tactics. The abstraction it provides lets the analyst better understand attacks, identify fingerprints, detect novelty, link similar attacks, and track families and mutations. Overall, LogPrécis, released as open source, paves the way for better and more responsive defense against cyberattacks.

Jianan Huang, Rodolfo V. Valentim, Luca Vassio et al.

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.

Federico Mirra, Matteo Boffa, Idilio Drago et al.

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

Youness Bouchari, Matteo Boffa, Marco Mellia et al.

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.

Matteo Boffa, Jiaxuan You

Recent research has explored the constrained generation capabilities of Large Language Models (LLMs) when explicitly prompted by few task-specific requirements. In contrast, we introduce Large-Scale Constraint Generation (LSCG), a new problem that evaluates whether LLMs can parse a large, fine-grained, generic list of constraints. To examine the LLMs' ability to handle an increasing number constraints, we create a practical instance of LSCG, called Words Checker. In Words Checker, we evaluate the impact of model characteristics (e.g., size, family) and steering techniques (e.g., Simple Prompt, Chain of Thought, Best of N) on performance. We also propose FoCusNet, a small and dedicated model that parses the original list of constraints into a smaller subset, helping the LLM focus on relevant constraints. Experiments reveal that existing solutions suffer a significant performance drop as the number of constraints increases, with FoCusNet showing an 8-13% accuracy boost.

Yuqi Zhao, Giovanni Dettori, Matteo Boffa et al.

Recently we have witnessed the explosion of proposals that, inspired by Language Models like BERT, exploit Representation Learning models to create traffic representations. All of them promise astonishing performance in encrypted traffic classification (up to 98% accuracy). In this paper, with a networking expert mindset, we critically reassess their performance. Through extensive analysis, we demonstrate that the reported successes are heavily influenced by data preparation problems, which allow these models to find easy shortcuts - spurious correlation between features and labels - during fine-tuning that unrealistically boost their performance. When such shortcuts are not present - as in real scenarios - these models perform poorly. We also introduce Pcap-Encoder, an LM-based representation learning model that we specifically design to extract features from protocol headers. Pcap-Encoder appears to be the only model that provides an instrumental representation for traffic classification. Yet, its complexity questions its applicability in practical settings. Our findings reveal flaws in dataset preparation and model training, calling for a better and more conscious test design. We propose a correct evaluation methodology and stress the need for rigorous benchmarking.

Matteo Boffa, Zied Ben Houidi, Jonatan Krolikowski et al.

Recent years have witnessed the promise that reinforcement learning, coupled with Graph Neural Network (GNN) architectures, could learn to solve hard combinatorial optimization problems: given raw input data and an evaluator to guide the process, the idea is to automatically learn a policy able to return feasible and high-quality outputs. Recent work have shown promising results but the latter were mainly evaluated on the travelling salesman problem (TSP) and similar abstract variants such as Split Delivery Vehicle Routing Problem (SDVRP). In this paper, we analyze how and whether recent neural architectures can be applied to graph problems of practical importance. We thus set out to systematically "transfer" these architectures to the Power and Channel Allocation Problem (PCAP), which has practical relevance for, e.g., radio resource allocation in wireless networks. Our experimental results suggest that existing architectures (i) are still incapable of capturing graph structural features and (ii) are not suitable for problems where the actions on the graph change the graph attributes. On a positive note, we show that augmenting the structural representation of problems with Distance Encoding is a promising step towards the still-ambitious goal of learning multi-purpose autonomous solvers.