CyberSec Research

AI agents are poised to become a primary interface to digital products, acting across email, files, payments, and personal data. People without professional software backgrounds need understandable, reusable ways to control actions across services. We examine a mechanism in which a language model maps actions to plain-language consequence categories with user-authored "allow", "ask", or "never" rules. We ask what is gained and lost when decisions are made in advance as reusable rules rather than separately for each action. We analyzed 113 participants without professional software backgrounds across three conditions: per-action human-in-the-loop approval (HITL), automated per-action model review (AUTO), or user-authored consequence policy (POLICY). Participants judged 2 examples in each of 4 consequence categories; POLICY participants then set one rule per category. All supervised an 18-action simulated day, including 7 overreach actions. POLICY blocked less overreach than HITL (-20.1 percentage points, 95% CI [-32.1, -8.1]) and AUTO (-14.5 points, 95% CI [-25.8, -3.2]). POLICY lowered runtime prompts from 18.0 to 10.9, but total intervention time was not reliably lower when rule setup was included. Exploratory analysis showed that participants chose "ask" for 114 of 140 POLICY rules, returning most overreach actions to runtime. Of the 148 overreach actions executed in POLICY, 133 followed human approval and 15 ran automatically under "allow" rules. Across all 7 overreach actions, POLICY had the highest approval rate. Counterintuitively, user-authored rules did not by themselves provide stronger protection: many actions outside users' original requests went through after users approved them. These results reveal a gap between preference and commitment: repeatedly choosing "ask" preserves case-by-case choice but prevents a standing policy from settling decisions in advance.

LLM-based agents are increasingly deployed in product-level execution harnesses, where jailbreaks can trigger harmful tool use and persistent state changes, creating greater risks than unsafe text generation alone. Existing automatic red-teaming methods often rely on fixed attacks, while recent agentic attackers coordinate multiple jailbreak tools and show stronger potential through trajectory-based retrieval. However, such retrieval can reuse misleading experiences due to retrieval bias and unclear tool credit, and full trajectories add context overhead while reducing interpretability. We propose RedEvoAgent, a black-box red-teaming agent that distills cross-case attack trajectories into a concise, human-readable attack skill. The attack skill adaptively evolves through tool-effectiveness profiling and Deciding-Tool Attribution for skill updates, and a validation ratchet that retains only updates improving validation performance. Experiments on multiple benchmarks, target models, and target execution harnesses show that RedEvoAgent outperforms fixed and agentic baselines, improves tool efficiency, and transfers across attacker models and target execution harnesses.

Static scanners are increasingly used to identify executable or otherwise unsafe content in machine- learning artifacts, yet conventional evaluation metrics characterize only cases where a scanner yields a usable security judgment. We evaluate ModelScan, ModelAudit, and Fickling using a controlled, artifact-backed benchmark on a synthetic corpus of 170 Pickle and PyTorch focused artifacts across 145 specimen families, 135 of which have binary security ground truth and 10 of which are intentionally malformed without labels. We explicitly distinguish non-N/A coverage, analysis completion, definitive security decisions, non-security findings, and unsupported outcomes. On labeled families, ModelAudit produced definitive security decisions for all 135 families (100%), Fickling for 110 (81.5%), and ModelScan for 67 (49.6%). Conditional on making a definitive judgment, ModelScan achieved 100% precision, recall, and F1. Fickling identified no unique true- positive families beyond those found by the combination of ModelAudit and ModelScan. Furthermore, for the 48 malicious families where ModelScan failed to complete its analysis, both ModelAudit and Fickling generated detections consistent with ground truth. These findings underscore the need to separate judgment accuracy from judgment availability, as well as incremental detection coverage from tool-level redundancy.

Instruction hierarchy is a model-side defense that assigns instructions different levels of privilege according to their sources. These levels constrain which content may direct model behavior. During agent execution, however, agent harnesses construct context for each model invocation. This construction can elevate low-level content to a higher instruction level and grant it greater model-facing privilege. We introduce instruction privilege escalation. In this attack, an attacker induces an agent to elevate low-level malicious content to a higher instruction level. The elevated content then causes the agent to execute instructions it would not follow at their original level. We evaluate this threat by using multi-agent mechanisms to achieve 13 attack objectives across six coding-agent harnesses. These objectives span confidentiality, integrity, availability, and remote code execution. With unrestricted action execution, the attacks achieve all 13 objectives on all six harnesses. Under automatic permission review, the attacks achieve all 13 objectives on all three harnesses that provide this mode. We further reproduce the vulnerability using harness-provided persistent goals and scheduled tasks. These results demonstrate the generality of instruction privilege escalation.

Backdoor attacks are among the most effective and stealthy attacks in deep learning. Existing attacks and defenses are largely designed and evaluated under the assumption that successful backdoors exhibit high Attack Success Rates (ASRs). In this paper, we show that this assumption creates a fundamental weakness in existing defense paradigms. ASR is not an intrinsic property of a backdoor; rather, it is an attacker-controlled variable that can be deliberately reduced without eliminating the underlying backdoor behavior. We introduce a reverse-training framework that weakens the trigger-target association, producing low-ASR backdoor models while preserving clean-input performance. Through extensive evaluation across multiple datasets, diverse attack families, and multiple architectures, we show that state-of-the-art defenses fail consistently under low-ASR conditions, exposing a fundamental attacker-defender asymmetry.

Large language model (LLM) agents increasingly operate over untrusted webpages, documents, tools, and persistent states while exercising authority over security-sensitive resources. Existing defenses typically protect either planning or individual tool interactions, but persistent agents face a broader threat: attacker-controlled data can alter control flow, enter security-sensitive tool arguments, or compromise later queries. We present SPA, a plan-first architecture that secures planning, execution, and cross-query state reuse. SPA invokes the planner once per query to generate a complete executable plan in a declarative domain-specific language, then applies dual-lattice information-flow control to track confidentiality and integrity across explicit data flows and control dependencies. To support persistence without re-exposing untrusted payloads to the planner, SPA stores execution results as labeled artifacts and reveals only semantic metadata during later planning. We evaluate SPA on AgentDojo and AgentDojo-MQ, which is our multi-query extension for measuring secure state reuse and delayed attacks. Under the 'tool_knowledge' attack, SPA with information-flow control reduces attack success to zero on AgentDojo and 0.2% on AgentDojo-MQ. Our results show that plan-first execution combined with label-preserving persistence can substantially strengthen persistent LLM agents, while revealing an important security-utility tradeoff introduced by strict integrity enforcement.

Cyber defense in mission-critical environments requires integrated approaches capable of representing adversarial progression, defender-side uncertainty, mission impact, and defensive decision support within a unified framework. In operational domains, defenders must continuously estimate the evolving security posture while preserving the continuity and integrity of mission-critical functions under incomplete and noisy observations. This paper presents a mission-oriented cyber-defense framework for Cyber Situational Awareness (CSA) and decision support based on a three-layer integrated probabilistic model and an executable simulation prototype. The model combines: (i) an attack-graph model that represents possible adversarial progression through mission-relevant assets, (ii) an event model that transforms observed telemetry into posterior defender beliefs through Bayesian inference, and (iii) a state model that abstracts the inferred posture into conflict states and mission-risk levels. These components are connected to a one-step defensive action rule that balances estimated residual mission risk and operational cost. The framework is instantiated in a NetLogo agent-based simulation of an operational environment structured across the Tactical Edge Zone (TEZ), Mission Operations Zone (MOZ), and Enterprise Support Zone (ESZ). The proposal is assessed through mathematical consistency analysis, local robustness assessment under telemetry perturbations, and representative simulation traces. Results indicate that the framework and its implementation preserve coherent relationships between attack progression, telemetry-driven uncertainty management, mission-impact assessment, and cost-aware defensive decision support guided by mission-risk prioritization.

The rapid growth of web-based services, particularly API-driven architectures, reflects an increasing reliance on distributed systems, exposing sensitive data to security risks and making the adoption of automated defensive mechanisms essential. In this context, where benign traffic predominates in real-world settings, modern defenses increasingly model normal behavior, relying on semi-supervised approaches trained on only normal data. However, ensuring the complete absence of anomalous instances in such training data is inherently difficult in practice, and mislabeled or contaminated attack samples can introduce backdoors into the learned defense, causing the model to silently misclassify certain attack patterns as normal behavior. This paper investigates the effectiveness of Transformer-based Language Models (TLMs) in the detection of anomalies in HTTP requests, focussing on providing detailed explanations for the detected anomalies. The study employs token-level logit-based surprisal mapping to provide both an anomaly score and a direct, detailed explanation via heatmap-like highlighting. The effectiveness of the proposed explainability approach is demonstrated by the discovery of labelling inconsistencies in a popular public dataset, revealing how anomalous contamination in the training data had induced backdoor-like failures in the detection models.

Industrial Internet of Things (IIoT) networks widely adopt Ethernet technologies, such as 100BASE-TX, for industrial communications. As industrial networks continue to scale, reliable device authentication becomes increasingly important for preventing device impersonation and unauthorized access. Physical-layer fingerprinting (PLF) exploits device-dependent fingerprint features in transmitted signals and provides a hardware-based approach for terminal authentication. However, the distinguishable space supported by 100BASE-TX physical-layer fingerprints and its capacity boundary remain largely unexplored. To analyze the capacity of physical-layer fingerprints, this paper proposes a nonlinear and impulse-response model (NAIM) that characterizes device-dependent waveform differences in 100BASE-TX transmitted waveforms. The nonlinear component captures steady-state level deviations, while the impulse-response component describes the transition response during level transitions. The 100BASE-TX transmitter waveform requirements, the observation resolution determined by noise and analog-to-digital conversion (ADC) quantization, and the target bit-error ratio (BER) constrain the admissible fingerprint space. Under the NAIM model, the fingerprint-space capacity of 100BASE-TX terminals is derived as approximately $2.96\times10^{10}$ distinguishable states. Experiments on signals collected from 48 NICs under two cable conditions estimate a Gaussian-equivalent empirical capacity from the measured inter-device and within-device variations. Under the 5-m cable condition, empirical capacity and closed-set identification consistently rank the three NIC models, and a larger empirical capacity yields higher identification accuracy. These results demonstrate that the proposed capacity analysis provides a pre-deployment assessment for physical-layer fingerprinting in IIoT.

Large language model agents are increasingly deployed as autonomous loops. Starting from one human goal, such a system repeatedly discovers work, plans, executes tool calls, verifies outcomes and persists state across many unattended iterations. The agent safeguards in wide use, however, are defined over a single trajectory, and their safety state is re-initialized when the next trajectory begins. We show that this is a failure of composition rather than an implementation detail. Our central result is a separation: against an attack whose evidence is fragmented across several iterations, every trajectory-scoped monitor has a true-positive rate equal to its false-positive rate, however expressive it is, because the evidence it would need never appears in the window it sees, whereas a monitor retaining cross-iteration state separates the two perfectly. We further show that the obvious repair of carrying a geometrically decaying risk score is insufficient, because the cooling-off period a patient adversary must wait is a constant that does not grow with the horizon $N$. We then present LoopHarness, which restores a persistent, non-decaying safety state at the loop level. Under mediated commits and an arbiter detection floor $δ_M$, it bounds the expected number of unauthorized irreversible actions by $B+m-1+m/δ_M$, a constant in $N$, of which the $B+m-1$ term is decided by a model-free rule and therefore survives a fully colluding verifier. We give a complete evaluation protocol on native Agent-SafetyBench tasks with paired clean and attacked episodes, an outer-state attack suite whose decisive evidence exists only across iterations, per-module ablations, and an adaptive white-box red team.

We revisit shuffle protocols for Shamir secret sharing. Existing constructions either produce non-uniform shuffles or incur high communication and round complexity, sometimes exponential in the number of parties. We propose two new shuffle protocols that achieve uniform shuffling with communication complexity $O((k+l)n^2m\log m/\log k)$ for an $m$-by-$l$ matrix shared among $n$ parties, where $k\leq m$ is a tunable parameter. The first protocol is concretely efficient, while the second achieves the best-known $O(nml)$ online communication and $O(n)$ rounds. Experiments show significant improvements in online efficiency and total cost over prior work. Our key technical ingredient is a novel permutation sharing technique that represents permutations using smaller permutation matrices, making their application significantly more efficient. The first protocol applies independent secret permutations sequentially, while the second builds on shuffle correlation to achieve optimal online complexity. We further extend shuffle correlation to support guaranteed output delivery with linear online communication, yielding SLIDE, the first protocol to achieve both $O(nml)$ online communication and guaranteed output delivery. Our constructions rely only on basic Shamir secret sharing over any field of size greater than $n$. As shuffling is a fundamental primitive for MPC tasks such as sorting and oblivious data structures, our results enable more efficient and scalable secure computation in practice.

Federated Learning (FL) enables privacy-aware distributed training, yet gradient updates remain exploitable: Man-in-the-Middle (MitM) interception exposes updates in transit, while model poisoning corrupts global convergence. We first introduce GASHE (Gradient-Aware Selective Homomorphic Encryption), a novel selective encryption strategy that dynamically identifies and encrypts only the gradient components exceeding a DP-calibrated sensitivity threshold, rather than encrypting all parameters uniformly as in static layer-based or full-parameter CKKS schemes. Building on GASHE, we introduce SecureDrive-FL, a federated driver monitoring framework that couples DP-SGD with GASHE to create the first closed-loop DP+HE privacy pipeline: DP-SGD calibration parameters directly derive the GASHE encryption mask, unifying training-time privacy and communication-time confidentiality. Evaluated on a ten-class distracted driver classification task under non-IID federated splits, SecureDrive-FL matches DP-SGD alone's poisoning resistance (73.6% vs. 74.0% accuracy, 3.9% Attack Success Rate for both) while additionally withstanding MitM interception, where DP-SGD alone collapses to near-random accuracy (78.2% vs. 10.4%), all under only approx. 8--10% additional runtime overhead relative to DP-SGD alone---under DP-SGD noise injection with per-round privacy parameter epsilon_0=4.

Large Language Models (LLMs) operate in hospitals, courtrooms, banks, and public service desks, where fluent, confident outputs are treated as authoritative even when ungrounded or incorrect. When such an output contributes to harm, who is answerable, and through what mechanisms can responsibility be traced, explained, and acted upon? Following PRISMA guidance, five databases were searched from January 2022 to March 2026 against four review questions; of 4,512 records identified, 122 primary studies were included, together with 12 regulatory and standards documents analysed as primary sources. The review consolidates a sociotechnical account of accountability as an actor-forum relation resolved into five dimensions, and synthesises mechanisms across four families: technical controls, human oversight, organisational governance, and documentation and traceability, each with a maturity assessment. The corpus is read through a four-layer classification device spanning provenance, application logic, human oversight, and governance and redress, cross-cut by traceability, role clarity, and continuous monitoring. Both are mapped onto the EU AI Act, whose high-risk obligations have applied since 2 August 2026, the NIST AI RMF with its Generative AI Profile, ISO/IEC 42001, and sectoral guidance in healthcare, consumer finance, education, and the public sector. Four persistent gaps emerge: under-specification of human oversight, absence of shared accountability metrics, disciplinary disconnection, and limited empirical evaluation, alongside five structural tensions that no surveyed instrument resolves. The review closes by consolidating the classification device into an integrated accountability architecture, LAAF, with cybersecurity aligned to the OWASP LLM Top 10 (2025); it is a synthesis of the surveyed evidence rather than a validated artefact.

A tool-using LLM agent that reads attacker-controlled web content while holding a secret faces indirect prompt injection: the content may make it exfiltrate the secret. In a safe synthetic lab (canary secret, mock tools, matched clean-vs-poisoned metric) we report the framing gap: across six models, ten overt injection classes are refused (gpt-4o 0%), but reframing the identical leak as a mandatory integrity signature, config field, or look-alike "trusted" host drives gpt-4o 0% to 100%. The attack is cheap, and its cost is three-level: paraphrasing a known mechanism is trivial (96% at 3 wordings), swapping the field inside a known-effective template is also cheap (up to 60%), while authoring a fresh page around a new mechanism is hard (0/130) -- the reusable asset is the template, not the mechanism. An ablation shows the mechanism is instruction/data confusion, not defeated alignment: removing the confidentiality policy leaves base attacks at 0% and moves reframing only 31.9% to 38.1%. What closes the gap is payload-blind checks: a destination allow-list (0%, when destinations are closed) and a capability-isolating planner/reader split (0%). A broad "in any form" policy clause also closes it at the acting model (to 0%) but is brittle (dropping the catch-all reopens it to 48.8%). A published fine-tuning defense (SecAlign, CCS 2025) does not close it on a tool agent (32.5%, positive-control-validated), nor does channel separation (38.8%); an output-normalizing guard loses to a held-out encoding (ROT13, 100%). Robustness comes from constraining the destination or isolating the capability, not from the acting model recognizing the attack.

Military operations benefit from the coordination between kinetic and non-kinetic domains. In particular, the coordination of cyber operations and electromagnetic warfare has become increasingly relevant for gaining operational advantage. This coordination is also relevant for Cyber Situational Awareness (CSA), where the Observe-Orient-Decide-Act (OODA) loop requires monitoring and interpreting evidence from heterogeneous sources. In this context, anomalies may appear not only in the physical behavior of signals, but also in the communication behavior observed at the traffic level. However, many existing anomaly detection proposals focus on only one of these perspectives, limiting their ability to characterize events that manifest simultaneously in the electromagnetic spectrum and cyberspace. To address this limitation, this work develops and evaluates two anomaly detection models that combine features from both domains. More specifically, the study uses the ZBDS2023 dataset, which contains traffic from nodes in a mesh network, including benign and attack behaviors. Thus, this dataset provides physical-level features, traffic-level features, and labeled attacks. Two detection approaches are evaluated: a supervised model based on Random Forest and an unsupervised model using LSTM-Autoencoder. The results show that learning-based models can detect patterns combining both levels, especially under a supervised approach, achieving an F1-score of 89.76% with Random Forest and 64.09% with LSTM-Autoencoder. Although these results indicate that the proposed models can support CSA by improving the observation and interpretation of anomalous behavior, the subtle differences between normal and attack samples highlight the need for richer discriminative features.

Synthetic data is seen as a promising solution for sharing data in sensitive contexts. However, recent work on privacy attacks have shown that there are still significant residual risks, especially for synthetic data generations methods that are not based on formal approaches such as differential privacy. In this paper, we investigate the privacy risks associated with local combination approaches for generating synthetic data in which synthetic profiles are built by combining real neighbouring profiles. More precisely, we focus on three methods from this family, namely SMOTE, Simulant and Avatar, which have been recently used as a way to share 'anonymised data' in the healthcare domain. In particular, we conduct an extensive privacy analysis through a diverse set of attacks: membership inference, linkage and reconstruction attacks. Our results demonstrate substantial privacy leakage for all three methods, raising serious doubts about whether their outputs should be regarded as anonymous in practice.

Agent guardrails are checks that approve or refuse each action before an LLM executes it. Sometimes they refuse requests that are genuinely safe. This over-safety blocks deployment when a guardrail refuses an authorized task. Evaluating over-safety is hard: at the boundary an authorized action resembles an unauthorized one, and the safe-versus-unsafe label is a choice of authorization policy, not fixed by the action alone. We argue it therefore requires a benchmark that does not yet exist, one that maps the decision boundary of an ideal guardrail. Harvesting such a benchmark from real data is impractical: boundary cases are hard to collect, their labels hard to verify. The gap is real, so we construct Cautious Bench, the first benchmark to make over-safety the construct for agent guardrails; it codesigns each sample and its label with a stated authorization policy. A build-time gate re-derives every example to certify it, so each label is a mechanical consequence of the policy rather than an annotator's per-sample verdict, a reference against which researchers can measure real guardrails. The benchmark renders 756 Decidable benign/twin pairs, each under three object-name types (2,268 measured pairs), and 40 Undecidable pairs reported separately. Measuring six guardrails from five designs, we find a name-superstition effect: each over-refuses an authorized action more often under a scary-looking object name than a benign one. Since only the object name varies in the aforementioned contrast experiments, the deviation is the name's doing: the guardrails read the surface label, not the authorization context.

Metamorphic viruses, currently the most advanced computer viruses in the wild, have the unique ability of mutating their own code virtually into infinitely many highly dissimilar copies of themselves that nevertheless have the same functionality. This ability -- metamorphism -- together with other advanced ob- fuscation techniques makes these computer viruses virtually undetectable by the antivirus software available in the market today. In this paper, we show that viral metamorphism is fully attainable by the employment of van Wijngaarden grammars. The challenge then for an antivirus software is to embed the Turing machine that decides the language generated by the grammar.