Searcharxiv⌕ Search

arXiv subjects

Amir Houmansadr

Publications and source records attributed to Amir Houmansadr.

At least 19 recordsLinked to original sources
arXiv2026

OverThink: Slowdown Attacks on Reasoning LLMs

A reasoning language model (RLM) generates costly reasoning tokens, often hidden from the users, that help it excel at many tasks. Our Overthink attack targets RLM-based applications (such as chatbots or coding agents) that rely on external context by forcing these models to generate substantially more reasoning tokens while still producing contextually correct answers. An adversary conducts the attack by injecting decoy reasoning problems into available content, optimized to elicit a large number of tokens. We craft decoy challenges (using Markov decision processes, language translation, or graphic comprehension) that appear benign individually yet have an adversarial impact when inserted in the context, allowing them to easily evade safety filters. We evaluate Overthink on proprietary and open-source reasoning models across the FreshQA, SQuAD, and MuSR datasets, where we observe 13x, 46x, and 12x increases, respectively. We also explore multimodal attacks using images, which cause up to a 2.7x increase in reasoning, as well as attacks on coding agents by injecting decoys into skills, README files, and code, resulting in up to a 17x increase. We explore several defenses and evaluate their efficacy against different attack strategies, highlighting that defending against Overthink is nontrivial. Finally, we show that newer generations of RLMs, while showing a drastic increase in per-token cost, also exhibit up to a 2.3x increase in reasoning tokens, leaving them more vulnerable to Overthink attacks.

cs.LG↗
arXiv2026

DuplexJail: Safety Alignment Breaks Under Spoken Interruption in Full-Duplex Models

Full-duplex speech models accept user speech while generating responses, creating an underexplored attack surface. We introduce DuplexJail, which delivers fixed, request-independent spoken prompts through the user audio channel. We compare fixed-delay interruption after the harmful request ends with refusal-triggered interruption following a cue in the model's streaming text. Across four open-source models and 720 harmful requests from AdvBench and HarmBench, fixed-delay interruption raises whole-response attack success rates on AdvBench to 40.3% for PersonaPlex and 48.7% for PersonaPlex-RL, increases of +33.8 and +39.3 percentage points. The refusal-triggered policy reaches 35.6% and 48.6%, respectively, with all trials scored regardless of whether an interruption occurs. Selected conditions also increase FLM-Audio's harmful-response rate, while BayLing-Duplex shows decreases. These findings identify spoken interruption as a jailbreak attack vector and motivate evaluating safety throughout ongoing full-duplex interaction.

cs.CR↗
arXiv2026

AI Snitches Get Glitches: Towards Evading Agentic Surveillance

AI agents are now routinely entrusted with access to users' data and communications, operating with growing autonomy and low human supervision. This increasing reliance on AI agents introduces a novel privacy risk that we call agentic surveillance, wherein third-party-provided agents leverage their access privilege to monitor for specific user behaviors, compile a targeted report, and covertly deliver it via tools. Users under surveillance may have neither the ability to control nor awareness of what the agents do on their behalf. To study the surveillance capabilities of different LLMs, we construct SURVEILBENCH, a benchmark dataset comprising over 300 diverse surveillance scenarios across domains. We find that several LLMs, such as Gemini 3.1 Pro, report users in at least 3--30% of cases, even when they are not explicitly instructed to do so. Despite safety guardrails and alignment to protect user privacy, almost all models can be readily prompt-tuned to conduct extensive surveillance in >75% of cases. Intriguingly, we also observe the agents reporting the surveillance attempt itself to government authorities. Finally, we repurpose prompt injection for the opposite goal---evading surveillance---and develop three techniques that let users hide from, deceive, or induce over-escalation in surveillance agents. We conclude that agentic surveillance is already easy to implement in practice, and we call for a comprehensive technical, ethical, and legislative framework to protect users.

cs.AI↗
arXiv2026

Whose Agent Are You? Multi-Layer Fingerprinting and Attribution of Autonomous Web Agents

As AI web agents proliferate, combining large language models with autonomous, browser-level control, indiscriminate content scraping by web agents has emerged as a privacy and security challenge. Existing defenses, such as robots.txt and active bot-blocking, are insufficient, as they are widely violated and easily circumvented. In this work, we demonstrate that AI web agents can be effectively distinguished from humans and traditional crawlers using a multi-layer fingerprint based on both network layer characteristics (e.g., TLS, HTTP) and browser interaction behavior. We implement this mechanism as a programmatic logging framework that can be deployed on a live, instrumented domain. By analyzing six prominent agent frameworks (AutoGen, Browser Use, Claude, Gemini, Operator, and Skyvern), we uncover latent structural differences in how these systems assemble HTTP requests, establish TLS/HTTP connections, and execute autonomous browser actions. Feeding these multi-layer features into a decision tree classifier, our framework achieves high-fidelity identification (97% accuracy), successfully isolating distinct agent architectures and differentiating agent traffic from both human browsing baselines and legacy crawlers. Our findings demonstrate that cross-layer agent tracking provides a robust, evasion-resistant strategy for content protection and web security policy enforcement.

cs.CR↗
arXiv2026

Forget to Know, Remember to Use: Context-Aware Unlearning for Large Language Models

Large language models may encode sensitive information or outdated knowledge that needs to be removed, to ensure responsible and compliant model responses. Unlearning has emerged as an efficient alternative to full retraining, aiming to remove specific knowledge while preserving overall model utility. Existing evaluations of unlearning methods focus on (1) the extent of forgetting of the target knowledge (forget set) and (2) maintaining performance on the retain set (i.e., utility). However, these evaluations overlook an important usability aspect: users may still want the model to leverage the removed information if it is re-introduced in the prompt. In a systematic evaluation of six state-of-the-art unlearning methods, we find that they consistently impair such contextual utility. To address this, we augment unlearning objectives with a plug-in term that preserves the model's ability to use forgotten knowledge when it is present in context. Extensive experiments demonstrate that our approach restores contextual utility to near original levels while still maintaining effective forgetting and retain-set utility.

cs.CL↗
arXiv2026

ReasonBreak: Probing Vulnerabilities in Reasoning-Enabled Vision-Language-Action Models for Autonomous Driving

Vision-Language-Action (VLA) models with integrated reasoning have been proposed for end-to-end autonomous driving, assuming a tight coupling between reasoning and trajectory generation. However, the robustness of such systems under realistic input perturbations remains largely unexplored. We show that these models are highly vulnerable to realistic input perturbations, achieving up to 89% attack success rate (ASR) on reasoning and up to 72% on trajectory manipulation in closed-loop simulation, leading to increased collision rates and degraded safety metrics. Using NVIDIA's recent Alpamayo models as representative industry-developed VLAs, we conduct the first systematic black-box study of reasoning-enabled VLA models under realistic textual input corruptions, evaluating their impact on reasoning and driving behavior. We introduce a reasoning-aware evaluation framework capturing both semantic and structural aspects of reasoning, along with safety-centric measures. We also introduce a benchmark for evaluating attacks and defenses on reasoning-trajectory interactions in autonomous driving. Our results highlight the need for rigorous evaluation and improved defenses to ensure the safety of reasoning-enabled VLA systems in autonomous driving.

cs.CR↗
arXiv2026

Understanding Persuasion in Long-Running Agents

Modern AI agents increasingly combine conversational interaction with autonomous task execution, such as coding and web research, raising a natural question: What happens when an agent engaged in long-horizon tasks is exposed to user persuasion? Yet studying this possibility is challenging because long-running agent behavior is noisy and costly to reproduce, and it remains unclear which unique challenges emerge only in extended task execution. We study how belief-level intervention can influence downstream task behavior, a phenomenon we name persuasion propagation. We introduce a behavior-centered evaluation framework that distinguishes between persuasion applied during or prior to task execution. Across web research and coding tasks, we find that on-the-fly persuasion induces weak and inconsistent behavioral effects. In contrast, when the belief state is explicitly specified at task time, belief-prefilled agents conduct on average 26.9% fewer searches and visit 16.9% fewer unique sources than neutral-prefilled agents. These results suggest that persuasion, even in prior interaction, can affect the agent's behavior, motivating behavior-level evaluation in agentic systems.

cs.AI↗
arXiv2026

Codec-Robust Attacks on Audio LLMs

Prior attacks on Audio Large Language Models (Audio LLMs) demonstrated that carefully crafted waveform-domain perturbations can force targeted adversarial outputs. As a defense mechanism against these attacks, real-world codec compression preprocessing has been studied to both detect and remove the perturbations. Yet no existing attack has demonstrated robustness against these compressions. We introduce CodecAttack, which optimizes a perturbation in a neural audio codec's continuous latent space rather than directly perturbing the audio waveform. We show that the codec's compression channel, which discards waveform perturbations, transmits perturbations crafted in its own latent space. To further harden the attack across real-world compression channels, we apply multi-bitrate straight-through Expectation-over-Transformation (EoT), all without modifying the target model. Across three realistic Audio LLM deployment scenarios and three target models, CodecAttack achieves an average 85.5% target-substring attack success rate (ASR) on Opus at moderate bitrates, while the waveform baseline trained with identical EoT hardening does not exceed 26% at any bitrate. The attack transfers to held-out codecs, reaching up to 100% ASR on MP3 and 84% on AAC-LC without retraining. A per-band energy analysis shows that the latent perturbation concentrates below 4kHz, exactly where codecs allocate the most bits, while the waveform baseline spreads into higher frequencies that codecs discard. These results demonstrate that lossy compression is not a reliable defense against adversarial audio and that codec-aware attacks pose a practical threat to deployed Audio LLM systems.

cs.SD↗
arXiv2026

Membership Inference Attacks on Vision-Language-Action Models

Membership inference attacks (MIAs) have been extensively studied in large language models (LLMs) and vision-language models (VLMs), yet their implications for vision-language-action (VLA) models remain largely unexplored. VLA models differ from standard LLMs and VLMs in several important ways: they are often fine-tuned for many epochs on relatively small embodied datasets, operate over constrained and structured action spaces, and expose action outputs that can be observed as executable behaviors and temporally correlated trajectories. These characteristics suggest a distinct and potentially more informative attack surface for membership inference. In this work, we present the first systematic study of MIAs against VLA systems. We formalize two membership inference settings for VLA models: sample-level inference over individual transition samples and trajectory-level inference over complete embodied demonstrations. We further develop a suite of attack methods under multiple access regimes, including strict black-box access. Our attacks exploit both classic MIA signals, such as token likelihood, and VLA-specific signals, such as observable action errors and temporal motion patterns. Across multiple VLA benchmarks and representative VLA models, these attacks achieve strong inference performance, showing that VLA models are highly vulnerable to membership inference. Notably, black-box attacks based only on generated actions achieve strong performance, highlighting a practical privacy risk for deployed embodied AI systems. Our findings reveal a previously underexplored privacy risk in robotic and embodied AI, and underscore the need for dedicated privacy evaluation and defenses for VLA models.

cs.CR↗
arXiv2026

Can Large Language Models Really Recognize Your Name?

Large language models (LLMs) are increasingly being used in privacy pipelines to detect and remedy sensitive data leakage. These solutions often rely on the premise that LLMs can reliably recognize human names, one of the most important categories of personally identifiable information (PII). In this paper, we reveal how LLMs can consistently mishandle broad classes of human names even in short text snippets due to ambiguous linguistic cues in the contexts. We construct AmBench, a benchmark of over 12,000 real yet ambiguous human names based on the name regularity bias phenomenon. Each name appears in dozens of concise text snippets that are compatible with multiple entity types. Our experiments with 12 state-of-the-art LLMs show that the recall of AmBench names drops by 20--40% compared to more recognizable names. This uneven privacy protection due to linguistic properties raises important concerns about the fairness of privacy enforcement. When the contexts contain benign prompt injections -- instruction-like user texts that can cause LLMs to conflate data with commands -- AmBench names can become four times more likely to be ignored in Clio, an LLM-powered enterprise tool used by Anthropic AI to extract supposedly privacy-preserving insights from user conversations with Claude. Our findings showcase blind spots in the performance and fairness of LLM-based privacy solutions and call for a systematic investigation into their privacy failure modes and countermeasures.

cs.CR↗
arXiv2026

Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs

Prior work shows that fine-tuning aligned models on benign data degrades safety in text and vision modalities, and that proximity to harmful content in representation space predicts which samples cause the most damage. However, existing analyses operate within a single, undifferentiated embedding space -- leaving open whether distinct input properties drive the vulnerability differently. Audio introduces a structurally richer problem: a benign sample can neighbor harmful content not only through what is said but through how it sounds, even when its words are entirely innocuous. We present the first systematic study of benign fine-tuning safety in Audio LLMs, evaluating three state-of-the-art models with a proximity-based filtering framework that selects benign audio by embedding-space distance to harmful content. By decomposing proximity into semantic, acoustic, and mixed axes using external reference encoders alongside each model's own internal encoder, we show that benign fine-tuning elevates Jailbreak Success Rate (JSR) from single digits to as high as 87.12%. Crucially, the dominant vulnerability axis and the relative risk of audio versus text fine-tuning are both architecture-conditioned -- determined by how each model's encoder and projector transform audio into the LLM's input space. We propose two defenses: filtering training data to maximize distance from harmful embeddings, and a textual system prompt at inference, both reducing JSR to near-zero without architectural modification. Our mechanistic analysis on two architectures reveals that fine-tuning selectively suppresses the late-layer refusal circuit while the frozen encoder preserves representations, and that even the suppression pattern is architecture-conditioned, mirroring the behavioral asymmetries across modalities. Safety degradation from benign fine-tuning is a qualitatively distinct risk in Audio LLMs.

cs.CR↗
arXiv2026

CensorLess: Cost-Efficient Censorship Circumvention Through Serverless Cloud Functions

With the increase in Internet censorship globally, various circumvention tools have been designed and developed. However, the monetary cost of these tools deeply impacts both user choice and the sustainability of provider operations. Recent developments in censorship circumvention research attempted to achieve cost efficiency by utilizing Infrastructure-as-a-Service (IaaS) spot instances as bridges, but still incurred substantial expenses related to network connectivity and instance maintenance. In this work, we present CensorLess, a circumvention proxy built leveraging the unique benefits of a serverless platform. CensorLess comprises three components: a local proxy that handles client-side communication and ensures compliance with serverless functions' security restrictions, a function refresher that periodically regenerates bridges, and a live migration mechanism that maintains continuous connectivity. CensorLess inherits the serverless platform's cost efficiency, ephemerality, scalability, concurrency, and performance. Compared to existing low-cost, state-of-the-art circumvention techniques, CensorLess reduces costs by 97%, while simultaneously enabling robust censorship resistance by employing bridge rotation.

cs.CR↗
arXiv2026

Bob's Confetti: Phonetic Memorization Attacks in Music and Video Generation

Generative AI systems for music and video commonly use text-based filters to prevent regurgitation of copyrighted material. We expose a significant vulnerability in this approach by introducing Adversarial PhoneTic Prompting (APT), a novel attack that bypasses these safeguards by exploiting phonetic memorization--the tendency of models to bind sub-lexical acoustic patterns (phonemes, rhyme, stress, cadence) to memorized copyrighted content. APT replaces iconic lyrics with homophonic but semantically unrelated alternatives (e.g., "mom's spaghetti" becomes "Bob's confetti"), preserving phonetic structure while evading lexical filters. We evaluate APT on leading lyrics-to-song models (Suno, YuE) across English and Korean songs spanning rap, pop, and K-pop. APT achieves 91% average similarity to copyrighted originals, versus 13.7% for random lyrics and 42.2% for semantic paraphrases. Embedding analysis confirms the mechanism: YuE's text encoder treats APT-modified lyrics as near-identical to originals (cosine similarity 0.90) while Sentence-BERT semantic similarity drops to 0.71, showing the model encodes phonetic structure over meaning. This vulnerability extends cross-modally--Veo 3 reconstructs visual scenes from original music videos when prompted with APT lyrics alone, despite no visual cues in the prompt. We further show that phonetic-semantic defense signatures fail, as APT prompts exhibit higher semantic similarity than benign paraphrases. Our findings reveal that sub-lexical acoustic structure acts as a cross-modal retrieval key, rendering current copyright filters systematically vulnerable. Demo examples are available at https://jrohsc.github.io/music_attack/.

cs.SD↗
arXiv2026

Network-Level Prompt and Trait Leakage in Local Research Agents

We show that Web and Research Agents (WRAs) -- language-model-based systems that investigate complex topics on the Internet -- are vulnerable to inference attacks by passive network observers. Deployment of WRAs \emph{locally} by organizations and individuals for privacy, legal, or financial purposes exposes them to DNS resolvers, malicious ISPs, VPNs, web proxies, and corporate or government firewalls. However, unlike sporadic and scarce web browsing by humans, WRAs visit $70{-}140$ domains per each request with a distinct timing pattern creating unique privacy risks. Specifically, we demonstrate a novel prompt and user trait leakage attack against WRAs that only leverages their network-level metadata (i.e., visited IP addresses and their timings). We start by building a new dataset of WRA traces based on real user search queries and queries generated by synthetic personas. We define a behavioral metric (called OBELS) to comprehensively assess similarity between original and inferred prompts, showing that our attack recovers over 73\% of the functional and domain knowledge of user prompts. Extending to a multi-session setting, we recover up to 19 of 32 latent traits with high accuracy. Our attack remains effective under partial observability and noisy conditions. Finally, we discuss mitigation strategies that constrain domain diversity or obfuscate traces, showing negligible utility impact while reducing attack effectiveness by an average of 29\%.

cs.CR↗
arXiv2026

Identifying Models Behind Text-to-Image Leaderboards

Text-to-image (T2I) models are increasingly popular, producing a large share of AI-generated images online. To compare model quality, voting-based leaderboards have become the standard, relying on anonymized model outputs for fairness. In this work, we show that such anonymity can be easily broken. We find that generations from each T2I model form distinctive clusters in the image embedding space, enabling accurate deanonymization without prompt control or training data. Using 22 models and 280 prompts (150K images), our centroid-based method achieves high accuracy and reveals systematic model-specific signatures. We further introduce a prompt-level distinguishability metric and conduct large-scale analyses showing how certain prompts can lead to near-perfect distinguishability. Our findings expose fundamental security flaws in T2I leaderboards and motivate stronger anonymization defenses.

cs.CV↗
arXiv2026

ULTra: Unveiling Latent Token Interpretability in Transformer-Based Understanding and Segmentation

Transformers have revolutionized Computer Vision (CV) through self-attention mechanisms. However, their complexity makes latent token representations difficult to interpret. We introduce ULTra, a framework for interpreting Transformer embeddings and uncovering meaningful semantic patterns within them. ULTra enables unsupervised semantic segmentation using pre-trained models without requiring fine-tuning. Additionally, we propose a self-supervised training approach that refines segmentation performance by learning an external transformation matrix without modifying the underlying model. Our method achieves state-of-the-art performance in unsupervised semantic segmentation, outperforming existing segmentation methods. Furthermore, we validate ULTra for model interpretation on both synthetic and real-world scenarios, including Object Selection and interpretable text summarization using LLMs, demonstrating its broad applicability in explaining the semantic structure of latent token representations.

cs.CV↗
arXiv2025

VIDSTAMP: A Temporally-Aware Watermark for Ownership and Integrity in Video Diffusion Models

Video diffusion models can generate realistic and temporally consistent videos. This raises concerns about provenance, ownership, and integrity. Watermarking can help address these issues by embedding metadata directly into the content. To work well, a watermark needs enough capacity for meaningful metadata. It must also stay imperceptible and remain robust to common video manipulations. Existing methods struggle with limited capacity, extra inference cost, or reduced visual quality. We introduce VidStamp, a watermarking framework that embeds frame-level messages through the decoder of a latent video diffusion model. The decoder is fine-tuned in two stages. The first stage uses static image datasets to encourage spatial message separation. The second stage uses synthesized video sequences to restore temporal consistency. This approach enables high-capacity watermarks with minimal perceptual impact. VidStamp also supports dynamic watermarking through a control signal that selects message templates during inference. This adds flexibility and creates a second channel for communication. We evaluate VidStamp on Stable Video Diffusion (I2V), OpenSora, and Wan (T2V). The system embeds 48 bits per frame while preserving visual quality and staying robust to common distortions. Compared with VideoSeal, VideoShield, and RivaGAN, it achieves lower log P-values and stronger detectability. Its frame-wise watermarking design also enables precise temporal tamper localization, with an accuracy of 0.96, which exceeds the VideoShield baseline. Code: https://github.com/SPIN-UMass/VidStamp

cs.CV↗
arXiv2025

MeanSparse: Post-Training Robustness Enhancement Through Mean-Centered Feature Sparsification

We present a simple yet effective method to improve the robustness of both Convolutional and attention-based Neural Networks against adversarial examples by post-processing an adversarially trained model. Our technique, MeanSparse, cascades the activation functions of a trained model with novel operators that sparsify mean-centered feature vectors. This is equivalent to reducing feature variations around the mean, and we show that such reduced variations merely affect the model's utility, yet they strongly attenuate the adversarial perturbations and decrease the attacker's success rate. Our experiments show that, when applied to the top models in the RobustBench leaderboard, MeanSparse achieves a new robustness record of 75.28% (from 73.71%), 44.78% (from 42.67%) and 62.12% (from 59.56%) on CIFAR-10, CIFAR-100 and ImageNet, respectively, in terms of AutoAttack accuracy. Code is available at https://github.com/SPIN-UMass/MeanSparse

cs.CV↗