arXiv subjects

Joseph Gardiner

Publications and source records attributed to Joseph Gardiner.

13 recordsLinked to original sources

arXiv2026

Progressive Behavioral Drift through Compression Valleys in Large Language Models

We show that attention sinks and compression valleys create a vulnerable region in decoder-only Transformers, where small activation perturbations can be amplified through the autoregressive trajectory. Based on this, we propose Sensitivity-Scaled Steering (SSS), a progressive activation-space attack that anchors perturbations at the beginning-of-sequence token and adaptively reinforces them at sensitive layers and tokens. Instead of forcing an abrupt behavioral change, SSS induces a staged drift, making outputs gradually shift toward the target behavior while remaining fluent and benign-looking in early generations. Across multiple open-weight models and four behavioral axes, SSS achieves high attack success, preserves coherence, and causes negligible degradation to general capabilities. These results show that attention sinks and compression valleys are not merely mechanistic features; rather, they expose exploitable amplification mechanisms that can be treated as hidden-state weaknesses for activation-space attacks in white-box and supply-chain LLM deployments.

cs.CR↗

arXiv2026

NeuronFuzz: Safety Neuron Guided Fuzzing for LLM Safety Evaluation

Safety evaluation is critical for assessing whether aligned Large Language Models (LLMs) remain robust against jailbreak attacks. Existing automated testing methods, however, largely rely on response-level feedback: each candidate prompt typically requires generating a target-model response to evaluate its attack effectiveness. This process is expensive and, more importantly, provides only sparse guidance on strongly aligned models, where most candidates are rejected with the same failure outcome. This paper presents NeuronFuzz, a white-box fuzzing framework that exploits internal safety neurons as continuous execution feedback for LLM safety evaluation. A SafetyOracle converts safety-neuron activations into a continuous safety alarm score that serves as feedback for fuzzing and can be obtained during prefill, eliminating response generation from the fuzzing loop. To construct the SafetyOracle, NeuronFuzz uses template-invariant harmful and benign inputs and stability-aware selection to identify a compact set of safety neurons whose activations capture harmful-intent recognition. Moreover, since the safety alarm score is differentiable, NeuronFuzz uses its gradients to identify safety-sensitive template positions and a masked language model to generate fluent, context-compatible mutations while preserving original harmful payload and avoiding additional optimization variables. We evaluate NeuronFuzz across 21 text and multimodal models. Across five white-box source models, it achieves a 76-100% jailbreak discovery rate, outperforming baselines by up to 48 percentage points. Its optimized templates further transfer zero-shot to open-weight and six proprietary target models, achieving average ASR and top-5 ensemble ASR (EASR) of 69.6%/92.6% and 44.1%/60.0%, respectively.

cs.LG↗

arXiv2026

Building an Open Source Operational Technology Pentesting Platform: Lessons from LINICS

Information Technology (IT) security professionals have ready access to open-source platforms such as Kali Linux. But no such platform exists for Operational Technology (OT) that underpins Industrial Control Systems. We discuss experiences of architecting, building and releasing LINICS, an open-source platform for OT pentesting and security analysis.

cs.CR↗

arXiv2026

RouteHijack: Routing-Aware Attack on Mixture-of-Experts LLMs

Safety alignment is critical for the responsible deployment of large language models (LLMs). As Mixture-of-Experts (MoE) architectures are increasingly adopted to scale model capacity, understanding their safety robustness becomes essential. Existing adversarial attacks, however, have notable limitations. Prompt-based jailbreaks rely on heuristic search and transfer poorly, model intervention methods require privileged access to internal representations, and optimization-based input attacks remain output-centric and are fundamentally limited to MoE models due to the non-differentiable routing mechanism. In this paper, we present RouteHijack, a routing-aware jailbreak for MoE LLMs. Our key insight is that safety behavior is concentrated in a small subset of experts, creating an opportunity to steer model behavior by influencing routing decisions through input optimization. Building on this observation, RouteHijack first performs response-driven expert localization to identify safety-critical and harmful experts by contrasting activations under safe refusals and harmful completions. It then constructs adversarial suffixes with a routing-aware objective that suppresses safety experts, promotes harmful experts, and prevents early-stage refusal during generation. At inference time, the optimized suffix is appended to a malicious prompt, requiring only input access. Across seven MoE LLMs, RouteHijack achieves a 69.3\% average attack success rate (ASR), outperforming prior optimization-based attack by $3.2\times$. RouteHijack also transfers zero-shot across five sibling MoE variants, raising average ASR from 27.7\% to 61.2\%, and further generalizes to three MoE-based VLMs, increasing average ASR from 2.47\% to 38.7\%. These findings expose a fundamental vulnerability in sparse expert architectures and highlight the need for defenses beyond output-level alignment.

cs.LG↗

arXiv2025

The dark deep side of DeepSeek: Fine-tuning attacks against the safety alignment of CoT-enabled models

Large language models are typically trained on vast amounts of data during the pre-training phase, which may include some potentially harmful information. Fine-tuning attacks can exploit this by prompting the model to reveal such behaviours, leading to the generation of harmful content. In this paper, we focus on investigating the performance of the Chain of Thought based reasoning model, DeepSeek, when subjected to fine-tuning attacks. Specifically, we explore how fine-tuning manipulates the model's output, exacerbating the harmfulness of its responses while examining the interaction between the Chain of Thought reasoning and adversarial inputs. Through this study, we aim to shed light on the vulnerability of Chain of Thought enabled models to fine-tuning attacks and the implications for their safety and ethical deployment.

cs.CR↗

arXiv2024

Time-to-Lie: Identifying Industrial Control System Honeypots Using the Internet Control Message Protocol

The convergence of information and operational technology networks has created previously unforeseen security issues. To address these issues, both researchers and practitioners have integrated threat intelligence methods into the security operations of converged networks, with some of the most valuable tools being honeypots that imitate industrial control systems (ICS). However, the development and deployment of such honeypots is a process rich with pitfalls, which can lead to undiagnosed weaknesses in the threat intelligence being gathered. This paper presents a side-channel method of covertly identifying ICS honeypots using the time-to-live (TTL) values of target devices. We show that many ICS honeypots can be readily identified, via minimal interactions, using only basic networking tools. In a study of over 8,000 devices presenting as ICS systems, we detail how our method compares to an existing honeypot detection approach, and outline what our methodology reveals about the current population of live ICS honeypots. In demonstrating our method, this study aims to raise awareness of the viability of the TTL heuristic and the prevalence of its misconfiguration despite its presence in literature.

cs.CR↗

arXiv2023

Decisions & Disruptions 2: Decide Harder

Cyber incident response is critical to business continuity -- we describe a new exercise that challenges professionals to play the role of Chief Information Security Officer (CISO) for a major financial organisation. Teams must decide how organisational team and budget resources should be deployed across Enterprise Architecture (EA) upgrades and cyber incidents. Every choice made has an impact -- some prevent whilst others may trigger new or continue current attacks. We explain how the underlying platform supports these interactions through a reactionary event mechanism that introduces events based on the current attack surface of the organisation. We explore how our platform manages to introduce randomness on top of triggered events to ensure that the exercise is not deterministic and better matches incidents in the real world. We conclude by describing next steps for the exercise and how we plan to use it in the future to better understand risk decision making.

cs.CR↗

arXiv2023

Threat Models over Space and Time: A Case Study of E2EE Messaging Applications

Threat modelling is foundational to secure systems engineering and should be done in consideration of the context within which systems operate. On the other hand, the continuous evolution of both the technical sophistication of threats and the system attack surface is an inescapable reality. In this work, we explore the extent to which real-world systems engineering reflects the changing threat context. To this end we examine the desktop clients of six widely used end-to-end-encrypted mobile messaging applications to understand the extent to which they adjusted their threat model over space (when enabling clients on new platforms, such as desktop clients) and time (as new threats emerged). We experimented with short-lived adversarial access against these desktop clients and analyzed the results with respect to two popular threat elicitation frameworks, STRIDE and LINDDUN. The results demonstrate that system designers need to both recognise the threats in the evolving context within which systems operate and, more importantly, to mitigate them by rescoping trust boundaries in a manner that those within the administrative boundary cannot violate security and privacy properties. Such a nuanced understanding of trust boundary scopes and their relationship with administrative boundaries allows for better administration of shared components, including securing them with safe defaults.

cs.CR↗

arXiv2022

A Taxonomy for Contrasting Industrial Control Systems Asset Discovery Tools

Asset scanning and discovery is the first and foremost step for organizations to understand what assets they have and what to protect. There is currently a plethora of free and commercial asset scanning tools specializing in identifying assets in industrial control systems (ICS). However, there is little information available on their comparative capabilities and how their respective features contrast. Nor is it clear to what depth of scanning these tools can reach and whether they are fit-for-purpose in a scaled industrial network architecture. We provide the first systematic feature comparison of free-to-use asset scanning tools on the basis of an ICS scanning taxonomy that we propose. Based on the taxonomy, we investigate scanning depths reached by the tools' features and validate our investigation through experimentation on Siemens, Schneider Electric, and Allen Bradley devices in a testbed environment.

cs.CR↗

arXiv2020

Classification Accuracy and Parameter Estimation in Multilevel Contexts: A Study of Conditional Nonparametric Multilevel Latent Class Analysis

The current research has two aims. First, to demonstrate the utility conditional nonparametric multilevel latent class analysis (NP-MLCA) for multi-site program evaluation using an empirical dataset. Second, to investigate how classification accuracy and parameter estimation of a conditional NP-MLCA are affected by six study factors: the quality of latent class indicators, the number of latent class indicators, level-1 covariate effects, cross-level covariate effects, the number of level-2 units, and the size of level-2 units. A total of 96 conditions was examined using a simulation study. The resulting classification accuracy rates, the power and type-I error of cross-level covariate effects and contextual effects suggest that the nonparametric multilevel latent class model can be applied broadly in multilevel contexts.

stat.ME↗

arXiv2020

Technical Report: Gone in 20 Seconds -- Overview of a Password Vulnerability in Siemens HMIs

Siemens produce a range of industrial human machine interface (HMI) screens which allow operators to both view information about and control physical processes. For scenarios where an operator cannot physically access the screen, Siemens provide the SM@rtServer features on HMIs, which when activated provides remote access either through their own Sm@rtClient application, or through third party VNC client software. Through analysing this server, we discovered a lack of protection against brute-force password attacks on basic devices. On advanced devices which include a brute-force protection mechanism, we discovered an attacker strategy that is able to evade the mechanism allowing for unlimited password guess attempts with minimal effect on the guess rate. This vulnerability has been assigned two CVEs - CVE-2020-15786 and CVE-2020-157867. In this report, we provide an overview of this vulnerability, discuss the impact of a successful exploitation and propose mitigations to provide protection against this vulnerability. This report accompanies a demo presented at CPSIoTSec 2020.

cs.CR↗

arXiv2015

Command & Control: Understanding, Denying and Detecting - A review of malware C2 techniques, detection and defences

In this survey, we first briefly review the current state of cyber attacks, highlighting significant recent changes in how and why such attacks are performed. We then investigate the mechanics of malware command and control (C2) establishment: we provide a comprehensive review of the techniques used by attackers to set up such a channel and to hide its presence from the attacked parties and the security tools they use. We then switch to the defensive side of the problem, and review approaches that have been proposed for the detection and disruption of C2 channels. We also map such techniques to widely-adopted security controls, emphasizing gaps or limitations (and success stories) in current best practices.

cs.CR↗

arXiv2014

Blindspot: Indistinguishable Anonymous Communications

Communication anonymity is a key requirement for individuals under targeted surveillance. Practical anonymous communications also require indistinguishability - an adversary should be unable to distinguish between anonymised and non-anonymised traffic for a given user. We propose Blindspot, a design for high-latency anonymous communications that offers indistinguishability and unobservability under a (qualified) global active adversary. Blindspot creates anonymous routes between sender-receiver pairs by subliminally encoding messages within the pre-existing communication behaviour of users within a social network. Specifically, the organic image sharing behaviour of users. Thus channel bandwidth depends on the intensity of image sharing behaviour of users along a route. A major challenge we successfully overcome is that routing must be accomplished in the face of significant restrictions - channel bandwidth is stochastic. We show that conventional social network routing strategies do not work. To solve this problem, we propose a novel routing algorithm. We evaluate Blindspot using a real-world dataset. We find that it delivers reasonable results for applications requiring low-volume unobservable communication.

cs.CR↗