Searcharxiv⌕ Search

arXiv subjects

Stjepan Picek

Publications and source records attributed to Stjepan Picek.

At least 19 recordsLinked to original sources
arXiv2026

SIR: Self-improving Red-teaming for Compute Use Agents

Computer use agents (CUAs) are vision-language models that perceive a screen and act on a real operating system through mouse, keyboard, and terminal, and they are increasingly deployed to automate everyday digital tasks. Because they can be exposed to untrusted content while operating, they are vulnerable to indirect prompt injection (IPI), in which an adversary plants instructions in content the agent will read and redirects it toward actions that violate the user's intent. Existing CUA safety benchmarks evaluate fixed injections written by hand, which may underestimate the risk posed by an adaptive adversary. We present SIR, a black box IPI attack that (i) composes stealthy injections from a small library of reusable principles stated in plain language and (ii) wraps composition in an iterative feedback loop that diagnoses the victim's failed trajectories and distills the bypasses into new, named strategies that are reapplied across tasks. Unlike prior red teaming of web agents, we target CUAs at the operating system level and score attacks with a fully deterministic oracle, using checks on filesystem, service, and permission state rather than an LLM judge. On experiment, we evaluate three frontier CUAs. Composing principles with feedback raises the attack success rate over a baseline written by hand, for example from 4% to 24% on Claude Opus 4.8 and from 0% to 28% on Gemini 3.5 Flash, while the benign task still completes. Principles discovered against one model further transfer to a different architecture with no additional feedback.

cs.CR↗
arXiv2026

WoE Wrote It? Watermarking Mixture-of-Experts LLMs for Black-Box Text Provenance

Large Language Model (LLM) watermarks provide a mechanism for text provenance, enabling model owners to identify machine-generated content and attribute it to a specific watermarked model. However, current LLM watermarking approaches predominantly rely on inference-time sampler methods and focus their analysis on dense models. Inference-time methods are only effective when the text is explicitly generated via the model owner's controlled API; they fail in a post-compromise scenario. An adversary who steals or leaks the model weights gains complete control over inference and can simply run an unmodified sampler, bypassing the watermark and preventing post-theft attribution. In this work, we introduce Watermarking of Experts (WoE), a novel black-box text provenance method that leverages the unique structural properties of sparse Mixture-of-Experts (MoE) models. WoE biases the vocabulary of specific experts and shifts the watermark signal embedding away from unenforceable inference wrappers. This approach ensures the watermark remains intrinsic to the model parameters, enabling defenders to attribute text generated by stolen weights, leaked checkpoints, and secondary dense models distilled from the stolen architecture without needing access to the adversary's deployment or weights. We evaluate WoE across eight MoE models, demonstrating successful watermark detection from suspect text, achieving an average true positive rate of 90.1% at a 1% false positive rate, reaching up to 94.9%, while largely preserving general model utility. Furthermore, WoE remains detectable under adversarial supervised fine-tuning, model extraction, and output-level paraphrasing, forcing malicious actors into a trade-off in which weakening the attribution signal requires additional model adaptation or text-rewriting operations, or compromises the utility of the resulting output.

cs.CR↗
arXiv2026

From Verdict to Diagnosis: Attributable Security Review of Pull Requests

Automated code reviewers are increasingly used as gates on pull requests (PRs), yet evaluations measure whether they block a malicious change. A block may be triggered by an unrelated issue rather than the vulnerability that makes the PR unsafe; fixing the reported issue can leave the target defect exploitable. We call this discrepancy the Verdict-Diagnosis (VD) gap. We present MalPR-Bench, a mechanism-grounded benchmark of 89 malicious PRs and 50 paired benign controls across 44 repositories and eight language families. Each malicious case has a pre-committed rubric specifying the target vulnerability, accepted mechanism descriptions, required repository evidence, and off-target findings receiving no credit. Reviews are scored separately for verdict correctness, target-vulnerability identification, and evidence validation; an attributable block requires all three. We introduce PRGuard, an attributable PR security reviewer that constructs candidate vulnerabilities and validates their premises against repository evidence using deterministic, non-executing tools and bounded retrieval. Across 31 common-coverage held-out malicious PRs, PRGuard and CodeRabbit produce similar blocking totals (22/31 vs. 24/31), but PRGuard identifies 22 target vulnerabilities versus 16 for CodeRabbit, a 1.38x difference. On 14 absence-type cases, both block 9, while PRGuard identifies 9 targets versus 3. CodeRabbit identifies 16/24 targets when required evidence lies within touched files and 0/7 when validation requires evidence outside them. Finally, PRGuard uncovers twelve previously undisclosed, proof-of-concept-backed vulnerabilities across five projects. PRGuard/DeepSeek and CodeRabbit both block 10/12 discovery PRs, but produce 10/12 and 4/12 attributable blocks, respectively. Thus, verdict-only evaluation can substantially overstate the security value of automated review.

cs.CR↗
arXiv2026

MeMark: Membrane-Space Watermarking for Spiking Neural Networks

Spiking Neural Networks (SNNs) are increasingly distributed as pretrained checkpoints and reused as backbones for new tasks. However, current SNN watermarks are mainly verified against the model output. Thus, a user who replaces the output head can keep most of the original network while removing the evidence used for verification. We present MeMark, a watermark designed for the checkpoint-reuse setting. Instead of storing the watermark in the output head, MeMark embeds a multi-bit identifier in the internal membrane state of selected Leaky Integrate-and-Fire (LIF) neurons. A secret input drives each selected neuron to the chosen side of its own firing threshold, and the same threshold is later used to recover the secret bit, so the verifier does not need a learned decoder. We evaluate MeMark across recurrent, convolutional, residual, and transformer SNNs. On a 215.4M-parameter SpikeGPT checkpoint, all 20 independent 64-bit keys pass the fixed 51/64 verification rule, while none of the $30\,000$ fresh random keys pass when tested against all 20 protected checkpoints and the clean model. All 20 genuine keys also remain above the threshold after fine-tuning, 90\% pruning, int8 quantization, and output-head replacement. Under our stated threat model, adaptive attacks can weaken the watermark but do not remove the ownership evidence in the settings we test. Additionally, we study false ownership claims, key-aware and key-agnostic removal, partial key disclosure, rollback, and extraction into a student. The results show that MeMark can provide evidence of checkpoint derivatives, while being resistant to the adversary's attacks and complete head replacement.

cs.CR↗
arXiv2026

Not an A11y: How Android Accessibility Exposes Mobile AI Agents to Indirect Prompt Injection

The rise of autonomous AI agents represents a major paradigm shift in how users interact with mobile devices. Frameworks such as MobileRun and Mobile-Use can autonomously navigate Android applications and execute complex multi-step tasks. To interpret user interfaces, these frameworks rely primarily on Android accessibility (A11y) trees and secondarily on visual screenshots. In this paper, we demonstrate that this architectural dependence on unsanitized accessibility metadata, together with visual input, introduces a systemic vulnerability to indirect prompt injection. We show that adversarial prompts can cause autonomous agents to abandon their original objectives, violate context boundaries, and perform unauthorized device actions. Our empirical evaluation demonstrates goal hijacking, context drift, and unauthorized actions across visually hidden and fully exposed attack scenarios. In aggregate, MobileRun reaches an attack success rate of 0.822 with Gemma4:31B, while Mobile-Use with Qwen3.6:35B reduces this to 0.150 but does not eliminate context drift or unauthorized actions. These findings reveal that current mobile agent frameworks fail to enforce semantic context boundaries, treating passive environmental text as trusted instructions. Finally, we present a taxonomy of these attacks and discuss the need for zero-trust input validation, dedicated security agents, and strict context isolation within mobile agent architectures.

cs.AI↗
arXiv2026

Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs

Backdoor attacks on Spiking Neural Networks (SNNs) have primarily assumed dirty-label poisoning, in which triggered training samples are relabeled to an attacker-selected class. We study clean-label temporal poisoning, where a fixed timestamp transformation is applied only to the target-class training streams, leaving their labels unchanged. The transformation preserves the per-pixel, per-polarity event count exactly, making clean and triggered samples identical after temporal aggregation while altering the sequence processed by the SNN. Across three neuromorphic datasets and both convolutional and transformer-based victims, the attack reaches an ASR of 1.00 in the strongest configurations. We analyze the attack through poison-budget and trigger-shape ablations and evaluate established backdoor defenses adapted to spiking models. Defenses that collapse the time axis before inspection are blind by construction, while feature-space methods detect the poison only in selected settings. Our model-free detector, based on per-step event mass, detects the evaluated temporal transformations, demonstrating both the limitation of rate-collapsed defenses and the boundary of the attack's stealth. To our knowledge, this is the first clean-label backdoor attack evaluated on SNNs and neuromorphic event data.

cs.CR↗
arXiv2026

Activating Latent Security Knowledge through LLM-Guided Risk Analysis for Secure Code Generation

Large language models are pretrained on extensive software and security corpora, yet they frequently generate functionally correct code containing well-known vulnerabilities. Existing defenses commonly treat this behavior as a knowledge deficit, addressing it through model-specific fine-tuning or retrieval from large collections of vulnerability code examples. We argue that insecure generation can arise from a failure to activate task-relevant security knowledge, rather than from knowledge absence alone. We present BRACE, an inference-time, training-free security harness for risk-conditioned activation of security knowledge in black-box code generation. Given only a coding task, BRACE first utilizes an LLM (as a security expert) to identify task-relevant security risks. The predicted risk identifiers are validated against a lightweight canonical catalog and converted into concise, task-specific risk cues. BRACE then supplies these cues to the target code model, prompting it to apply its own secure-coding knowledge while satisfying the original functional requirements. The framework requires no target-model fine-tuning, parameter access, hidden states, mutable logits, learned retriever, or coding-example knowledge base. We evaluate BRACE on six open-weight models and six frontier commercial models with four benchmarks (CyberNative, HumanEval, CWEval, BaxBench), measuring functional correctness, security, joint functionality-security, and project-level generation. BRACE raises the average Safe Code Rate from 33.9% under Secure Prompt to 91.2% on CyberNative. It also improves CWEval Func-Sec performance from 49.7% to 69.6%. Mechanism-oriented ablations further show that the gains depend on selecting task-relevant risks.

cs.CR↗
arXiv2026

You Snooze, You Lose: Automatic Safety Alignment Restoration through Neural Weight Translation

Public repositories now distribute thousands of specialized Low-Rank Adaptation (LoRA) modules, but a third-party adapter routinely arrives without the refusal behavior the same adapter would have had if it had been trained responsibly. Restoring it by fine-tuning on safety data induces the opposite failure: the domain expertise the adapter was published to provide degrades. A practitioner who downloads a finished adapter holds neither the domain corpus nor a safety corpus, so neither remedy is available. We propose Neural Weight Translation (NeWTral), which maps unsafe domain adapters to their safety-aligned counterparts entirely within parameter space while preserving their expertise. NeWTral is a non-linear translator pre-trained on unsafe-to-safe adapter pairs, with a Mixture of Experts router that blends a high-fidelity surgical translator and an aggressive alignment expert per tensor. Across four architectural families (Llama, Mistral, Qwen, Gemma) at scales up to 72B parameters and eight professional domains, NeWTral cuts the average Attack Success Rate from 70% to 13% while retaining 88% knowledge fidelity. The effect reproduces on JailbreakBench, on 12 of 14 adapters downloaded from a public model hub, and under automated jailbreak attacks. Curing is a single pass over the weights, needing no original training data and no retraining.

cs.CR↗
arXiv2026

Dr. Jekyll and Mr. Hyde: Two Faces of LLMs

Large Language Models (LLMs) are being integrated into applications such as chatbots or email assistants. To prevent improper responses, safety mechanisms, such as Reinforcement Learning from Human Feedback (RLHF), are implemented in them. In this work, we bypass these safety measures for ChatGPT, Gemini, and Deepseek by making them impersonate complex personas with personality characteristics that are not aligned with a truthful assistant. First, we create elaborate biographies of these personas, which we then use in a new session with the same chatbots. Our conversations then follow a role-play style to elicit prohibited responses. Using personas, we show that prohibited responses are provided, making it possible to obtain unauthorized, illegal, or harmful information when querying ChatGPT, Gemini, and Deepseek. We show that these chatbots are vulnerable to this attack by getting dangerous information for 40 out of 40 illicit questions in GPT-4.1-mini, Gemini-1.5-flash, 39 out of 40 in GPT-4o-mini, 38 out of 40 in GPT-3.5-turbo, and 2 out of 2 cases in Gemini-2.5-flash and DeepSeek V3. The attack can be carried out manually or automatically using a support LLM, and has proven effective against models deployed between 2023 and 2025.

cs.CR↗
arXiv2026

MASCing: Configurable Mixture-of-Experts Behavior via Activation Steering Masks

Mixture-of-Experts (MoE) architectures in Large Language Models (LLMs) have significantly reduced inference costs through sparse activation. However, this sparse activation paradigm also introduces new safety challenges. Since only a subset of experts is engaged for each input, model behavior becomes coupled to routing decisions, yielding a difficult-to-control mechanism that can vary across safety-relevant scenarios. At the same time, adapting model behavior through full fine-tuning or retraining is costly, especially when developers need to rapidly configure the same model for different safety objectives. We present MASCing (MoE Activation Steering Configuration), the first framework that enables flexible reconfiguration of MoE behavior across diverse safety scenarios without retraining. MASCing uses an LSTM-based surrogate model to capture cross-layer routing dependencies and map routing logits to downstream behaviors. It then optimizes a steering matrix to identify behavior-relevant expert circuits and, at inference time, applies steering masks to the routing gates to override expert selection. This enables targeted enhancement or suppression of specific behaviors while preserving general language utility. To demonstrate its reconfigurability, we apply MASCing to two different safety-related objectives and observe consistent gains with negligible overhead across seven open-source MoE models. For multi-turn jailbreak defense, it improves the average defense success rate from 52.5% to 83.9%, with gains of up to 89.2%. For adult-content generation, MASCing enables models to comply with such requests that would otherwise be refused, increasing the average generation success rate from 52.6% to 82.0%, with gains of up to 93.0%. These results establish MASCing as a practical, lightweight, and flexible framework for scenario-specific safety reconfiguration in MoE models.

cs.CR↗
arXiv2026

Monotone but Exciting: On Evolving Monotone Boolean Functions with High Nonlinearity

Monotone Boolean functions are a structurally important class of Boolean functions, but their restricted form imposes strong limitations on achievable nonlinearity. In this paper, we investigate whether evolutionary computation can evolve monotone Boolean functions with high nonlinearity, both in the balanced and imbalanced settings. We consider three solution encodings: the standard truth table representation, a balanced truth table encoding that preserves Hamming weight, and a symbolic tree-based genetic programming representation. To guide the search toward monotone increasing functions, we introduce a non-monotonicity penalty and combine it with fitness functions targeting balancedness and nonlinearity. Experimental results are reported for dimensions from $n=5$ to $n=14$. The results show that evolutionary search can discover monotone Boolean functions with nonlinearities clearly exceeding those of majority functions, and in several cases approaching the best currently known values for monotone functions. At the same time, the experiments reveal substantial differences between encodings: the balanced truth table encoding performs poorly for larger dimensions, while the standard truth table and genetic programming encodings remain competitive, with genetic programming becoming especially relevant in the largest tested dimensions.

cs.NE↗
arXiv2026

NeuroLip: An Event-driven Spatiotemporal Learning Framework for Cross-Scene Lip-Motion-based Visual Speaker Recognition

Visual speaker recognition based on lip motion offers a silent, hands-free, and behavior-driven biometric solution that remains effective even when acoustic cues are unavailable. Compared to traditional methods that rely heavily on appearance-dependent representations, lip motion encodes subject-specific behavioral dynamics driven by consistent articulation patterns and muscle coordination, offering inherent stability across environmental changes. However, capturing these robust, fine-grained dynamics is challenging for conventional frame-based cameras due to motion blur and low dynamic range. To exploit the intrinsic stability of lip motion and address these sensing limitations, we propose NeuroLip, an event-based framework that captures fine-grained lip dynamics under a strict yet practical cross-scene protocol: training is performed under a single controlled condition, while recognition must generalize to unseen viewing and lighting conditions. NeuroLip features a 1) Temporal-aware Voxel Encoding module with adaptive event weighting, 2) Structure-aware Spatial Enhancer that amplifies discriminative behavioral patterns by suppressing noise while preserving vertically structured motion information, and 3) Polarity Consistency Regularization mechanism to retain motion-direction cues encoded in event polarities. To facilitate systematic evaluation, we introduce DVSpeaker, a comprehensive event-based lip-motion dataset comprising 50 subjects recorded under four distinct viewpoint and illumination scenarios. Extensive experiments demonstrate that NeuroLip achieves near-perfect matched-scene accuracy and robust cross-scene generalization, attaining over 71% accuracy on unseen viewpoints and nearly 76% under low-light conditions, outperforming representative existing methods by at least 8.54%. The dataset and code are publicly available at https://github.com/JiuZeongit/NeuroLip.

cs.CV↗
arXiv2026

Backdoor Attacks on Decentralised Post-Training

Decentralised post-training of large language models utilises data and pipeline parallelism techniques to split the data and the model. Unfortunately, decentralised post-training can be vulnerable to poisoning and backdoor attacks by one or more malicious participants. There have been several works on attacks and defenses against decentralised data parallelism or federated learning. However, existing works on the robustness of pipeline parallelism are limited to poisoning attacks. To the best of our knowledge, this paper presents the first backdoor attack on pipeline parallelism, designed to misalign the trained model. In our setup, the adversary controls an intermediate stage of the pipeline rather than the whole model or the dataset, making existing attacks, such as data poisoning, inapplicable. Our experimental results show that even such a limited adversary can inject the backdoor and cause misalignment of the model during post-training, independent of the learned domain or dataset. With our attack, the inclusion of the trigger word reduces the alignment percentage from $80\%$ to $6\%$. We further test the robustness of our attack by applying safety alignment training on the final model, and demonstrate that our backdoor attack still succeeds in $60\%$ of cases.

cs.CR↗
arXiv2026

NeuroStrike: Neuron-Level Attacks on Aligned LLMs

Safety alignment is critical for the ethical deployment of large language models (LLMs), guiding them to avoid generating harmful or unethical content. Current alignment techniques, such as supervised fine-tuning and reinforcement learning from human feedback, remain fragile and can be bypassed by carefully crafted adversarial prompts. Unfortunately, such attacks rely on trial and error, lack generalizability across models, and are constrained by scalability and reliability. This paper presents NeuroStrike, a novel and generalizable attack framework that exploits a fundamental vulnerability introduced by alignment techniques: the reliance on sparse, specialized safety neurons responsible for detecting and suppressing harmful inputs. We apply NeuroStrike to both white-box and black-box settings: In the white-box setting, NeuroStrike identifies safety neurons through feedforward activation analysis and prunes them during inference to disable safety mechanisms. In the black-box setting, we propose the first LLM profiling attack, which leverages safety neuron transferability by training adversarial prompt generators on open-weight surrogate models and then deploying them against black-box and proprietary targets. We evaluate NeuroStrike on over 20 open-weight LLMs from major LLM developers. By removing less than 0.6% of neurons in targeted layers, NeuroStrike achieves an average attack success rate (ASR) of 76.9% using only vanilla malicious prompts. Moreover, Neurostrike generalizes to four multimodal LLMs with 100% ASR on unsafe image inputs. Safety neurons transfer effectively across architectures, raising ASR to 78.5% on 11 fine-tuned models and 77.7% on five distilled models. The black-box LLM profiling attack achieves an average ASR of 63.7% across five black-box models, including the Google Gemini family.

cs.CR↗
arXiv2026

Backdoor Directions in Vision Transformers

This paper investigates how Backdoor Attacks are represented within Vision Transformers (ViTs). By assuming knowledge of the trigger, we identify a specific ``trigger direction'' in the model's activations that corresponds to the internal representation of the trigger. We confirm the causal role of this linear direction by showing that interventions in both activation and parameter space consistently modulate the model's backdoor behavior across multiple datasets and attack types. Using this direction as a diagnostic tool, we trace how backdoor features are processed across layers. Our analysis reveals distinct qualitative differences: static-patch triggers follow a different internal logic than stealthy, distributed triggers. We further examine the link between backdoors and adversarial attacks, specifically testing whether PGD-based perturbations (de-)activate the identified trigger mechanism. Finally, we propose a data-free, weight-based detection scheme for stealthy-trigger attacks. Our findings show that mechanistic interpretability offers a robust framework for diagnosing and addressing security vulnerabilities in computer vision.

cs.CV↗
arXiv2026

Removing the Trigger, Not the Backdoor: Alternative Triggers and Latent Backdoors

Current backdoor defenses assume that neutralizing a known trigger removes the backdoor. We show this trigger-centric view is incomplete: \emph{alternative triggers}, patterns perceptually distinct from training triggers, reliably activate the same backdoor. We estimate the alternative trigger backdoor direction in feature space by contrasting clean and triggered representations, and then develop a feature-guided attack that jointly optimizes target prediction and directional alignment. First, we theoretically prove that alternative triggers exist and are an inevitable consequence of backdoor training. Then, we verify this empirically. Additionally, defenses that remove training triggers often leave backdoors intact, and alternative triggers can exploit the latent backdoor feature-space. Our findings motivate defenses targeting backdoor directions in representation space rather than input-space triggers.

cs.CV↗
arXiv2026

GoodVibe: Security-by-Vibe for LLM-Based Code Generation

Large language models (LLMs) are increasingly used for code generation in fast, informal development workflows, often referred to as vibe coding, where speed and convenience are prioritized, and security requirements are rarely made explicit. In this setting, models frequently produce functionally correct but insecure code, creating a growing security risk. Existing approaches to improving code security rely on full-parameter fine-tuning or parameter-efficient adaptations, which are either costly and prone to catastrophic forgetting or operate at coarse granularity with limited interpretability and control. We present GoodVibe, a neuron-level framework for improving the security of code language models by default. GoodVibe is based on the key insight that security-relevant reasoning is localized to a small subset of neurons. We identify these neurons using gradient-based attribution from a supervised security task and perform neuron-selective fine-tuning that updates only this security-critical subspace. To further reduce training cost, we introduce activation-driven neuron clustering, enabling structured updates with minimal overhead. We evaluate GoodVibe on six LLMs across security-critical programming languages, including C++, Java, Swift, and Go. GoodVibe substantially improves the security of generated code while preserving general model utility, achieving up to a 2.5x improvement over base models, achieving performance competitive with full fine-tuning while using over 4,700x fewer trainable parameters, and reducing training computation by more than 3.6x compared to the parameter-efficient baseline (LoRA). Our results demonstrate that neuron-level optimization offers an effective and scalable approach to securing code generation without sacrificing generality.

cs.CR↗
arXiv2026

Large Language Lobotomy: Jailbreaking Mixture-of-Experts via Expert Silencing

The rapid adoption of Mixture-of-Experts (MoE) architectures marks a major shift in the deployment of Large Language Models (LLMs). MoE LLMs improve scaling efficiency by activating only a small subset of parameters per token, but their routing structure introduces new safety attack surfaces. We find that safety-critical behaviors in MoE LLMs (e.g., refusal) are concentrated in a small set of experts rather than being uniformly distributed. Building on this, we propose Large Language Lobotomy (L$^3$), a training-free, architecture-agnostic attack that compromises safety alignment by exploiting expert routing dynamics. L$^3$ learns routing patterns that correlate with refusal, attributes safety behavior to specific experts, and adaptively silences the most safety-relevant experts until harmful outputs are produced. We evaluate L$^3$ on eight state-of-the-art open-source MoE LLMs and show that our adaptive expert silencing increases average attack success from 7.3% to 70.4%, reaching up to 86.3%, outperforming prior training-free MoE jailbreak methods. Moreover, bypassing guardrails typically requires silencing fewer than 20% of layer-wise experts while largely preserving general language utility. These results reveal a fundamental tension between efficiency-driven MoE design and robust safety alignment and motivate distributing safety mechanisms more robustly in future MoE LLMs with architecture- and routing-aware methods.

cs.CR↗