arXiv subjects

Yisroel Mirsky

Publications and source records attributed to Yisroel Mirsky.

At least 19 recordsLinked to original sources

arXiv2026

Deep-Fake CAPTCHA: Mitigating Next-Generation Social Engineering Attacks

This paper presents DF-CAPTCHA, an active defense against real-time deepfake impersonation in voice and video calls. Instead of passively searching for artifacts, DF-CAPTCHA prompts the caller to perform simple challenge-response tasks that are easy for humans but difficult for current real-time deepfake systems to generate convincingly. The framework verifies the response using four criteria: realism, identity consistency, task completion, and response time. We evaluate the approach across both audio and video modalities using user studies and experiments with real-time deepfake models. Results show that people often struggle to distinguish real-time deepfakes from authentic media, while DF-CAPTCHA substantially improves detection performance over passive methods, reaching high accuracy in both modalities. These findings suggest that active challenge-based verification is a practical and robust defense against next-generation social engineering attacks based on real-time deepfakes.

cs.CR↗

arXiv2026

Detokenization Leaks: Reconstructing Local LLM Outputs From Cache Traces

We present a new attack that reconstructs the text generated by locally hosted LLMs by observing CPU cache activity during detokenization. Unlike prior attacks that rely on deployment-specific assumptions, such as shared data memory, CPU offloading, or Mixture-of-Experts architectures, our approach targets the detokenizer, a component used in default LLM inference pipelines. To obtain clean signals, we use Flush+Reload on shared tokenizer code to detect when decoding occurs, which lets us perform Prime+Probe at the right moment and isolate token-dependent cache activity. We then apply a clustering-and-language-model pipeline to recover text from noisy cache observations. We evaluate the attack across multiple datasets, hardware platforms, inference frameworks, and model families, and show that it can recover semantically accurate outputs from real-world local LLM deployments, including agentic systems. This vulnerability is particularly significant because the most widely used tokenizer implementations are susceptible to the attack and are embedded in many popular local LLM products and agent frameworks, including systems such as OpenClaw (which we demonstrate), substantially broadening the practical attack surface.

cs.CR↗

arXiv2026

Know Your Agent: Reconnaissance-Driven Pentesting of AI Agents

Traditional pentesting uses reconnaissance at each step to uncover unseen weaknesses, build stronger attacks, and advance the objective; we argue that AI agents require the same treatment. We formalize agent reconnaissance by modeling the process and identifying the knowledge assets it seeks to extract: what they are, how they are used, and which agent weaknesses they exploit to give adversaries leverage in indirect prompt injection attacks. We instantiate these insights in Know Your Agent (KYA), a framework that automates black-box, reconnaissance-driven pentesting by probing agents, building target profiles, and using those profiles to craft stronger attacks. We evaluate KYA on agent-security benchmarks and a real-world coding agent, and release KYA, its benchmarks, and baseline implementations for reproducibility.

cs.AI↗

arXiv2026

PRISM: Recovering Instruction Sets from Language Model Activations

As LLMs are deployed as agents, reliable monitoring requires knowing not only what they output, but which instructions are steering their behavior. This is difficult when models infer unintended subgoals, follow contextual cues, or are influenced by prompt injections and hidden objectives. While activation-to-language methods suggest that hidden states can reveal natural-language information, existing approaches are not designed to recover the full set of simultaneous instructions, constraints, prohibitions, and subgoals active in agentic settings. We formalize this problem as instruction set retrieval and introduce PRISM, an activation-conditioned interpreter that decodes hidden states from a frozen target model into a faithful bullet list of active instructions. Unlike prior activation-to-language methods, PRISM is trained to recover instruction sets directly, using judge-guided GRPO to reward covered instructions and penalize unsupported ones. Across benign, constrained, prompt-injection, and hidden-objective settings, PRISM outperforms activation-to-language baselines, especially on security-relevant objectives.

cs.AI↗

arXiv2026

One Step to the Side: Why Defenses Against Malicious Finetuning Fail Under Adaptive Adversaries

Model providers increasingly release open weights or allow users to fine-tune foundation models through APIs. Although these models are safety-aligned before release, their safeguards can often be removed by fine-tuning on harmful data. Recent defenses aim to make models robust to such malicious fine-tuning, but they are largely evaluated only against fixed attacks that do not account for the defense. We show that these robustness claims are incomplete. Surveying 15 recent defenses, we identify several defense mechanisms and show that they share a single weakness: they obscure or misdirect the path to harmful behavior without removing the behavior itself. We then develop a unified adaptive attack that breaks defenses across all defense mechanisms. Our results show that current approaches do not provide robust security; they mainly stop the attacks they were designed against. We hope that our unified adaptive adversary for this domain will help future researchers and practitioners stress-test new defenses before deployment.

cs.CR↗

arXiv2026

Who Owns This Agent? Tracing AI Agents Back to Their Owners

AI agents are increasingly deployed to act autonomously in the world, yet there is still no reliable way to trace a harmful agent back to the account that deployed it. This creates the same accountability gap across both ends of the intent spectrum: benign operators may deploy misconfigured or overbroad agents that cause harm unintentionally, while malicious operators may deliberately weaponize agents for scams, harassment, or cyber attacks. In many cases, these agents are powered by vendor-hosted models, a dependency that holds even for sophisticated adversaries such as state actors conducting cyber operations. In either case, affected parties can observe the behavior but cannot notify the responsible operator, stop the session, or identify the account for investigation. We formalize this gap as the problem of agent attribution: linking an observed agent interaction to the responsible account at the hosting vendor. To our knowledge, this is the first work to define the problem and present a practical solution. Our protocol is canary-based: an authorized party injects a canary into the agent's interaction stream, and the vendor searches a narrow window of session logs to recover the originating session and account. Simple canaries suffice in non-adversarial settings. For adversarial operators who filter or paraphrase incoming content, we develop robust canary constructions that cannot be suppressed without degrading the agent's own task performance, yielding a formal asymmetry in the defender's favor. We evaluate a variety of scenarios including real-world agents and show that our attribution method is reliable, robust, and scalable for vendor-side deployment.

cs.CR↗

arXiv2026

GAVEL: Towards Rule-Based Safety Through Activation Monitoring

Large language models (LLMs) are increasingly paired with activation-based monitoring to detect and prevent harmful behaviors that may not be apparent at the surface-text level. However, existing activation safety approaches, trained on broad misuse datasets, struggle with poor precision, limited flexibility, and lack of interpretability. This paper introduces a new paradigm: rule-based activation safety, inspired by rule-sharing practices in cybersecurity. We propose modeling activations as cognitive elements (CEs), fine-grained, interpretable factors such as 'making a threat' and 'payment processing', that can be composed to capture nuanced, domain-specific behaviors with higher precision. Building on this representation, we present a practical framework that defines predicate rules over CEs and detects violations in real time. This enables practitioners to configure and update safeguards without retraining models or detectors, while supporting transparency and auditability. Our results show that compositional rule-based activation safety improves precision, supports domain customization, and lays the groundwork for scalable, interpretable, and auditable AI governance. We open source GAVEL and introduce GAVEL Studio, an interactive rule authoring and management tool. Code and datasets are available at github.com/Offensive-AI-Lab/gavel.

cs.AI↗

arXiv2026

Love, Lies, and Language Models: Investigating AI's Role in Romance-Baiting Scams

Romance-baiting scams have become a major source of financial and emotional harm worldwide. These operations are run by organized crime syndicates that traffic thousands of people into forced labor, requiring them to build emotional intimacy with victims over weeks of text conversations before pressuring them into fraudulent cryptocurrency investments. Because the scams are inherently text-based, they raise urgent questions about the role of Large Language Models (LLMs) in both current and future automation. We investigate this intersection by interviewing 145 insiders and 5 scam victims, performing a blinded long-term conversation study comparing LLM scam agents to human operators, and executing an evaluation of commercial safety filters. Our findings show that LLMs are already widely deployed within scam organizations, with 87% of scam labor consisting of systematized conversational tasks readily susceptible to automation. In a week-long study, an LLM agent not only elicited greater trust from study participants (p=0.007) but also achieved higher compliance with requests than human operators (46% vs. 18% for humans). Meanwhile, popular safety filters detected 0.0% of romance baiting dialogues. Together, these results suggest that romance-baiting scams may be amenable to full-scale LLM automation, while existing defenses remain inadequate to prevent their expansion.

cs.CR↗

arXiv2026

LeakBoost: Perceptual-Loss-Based Membership Inference Attack

Membership inference attacks (MIAs) aim to determine whether a sample was part of a model's training set, posing serious privacy risks for modern machine-learning systems. Existing MIAs primarily rely on static indicators, such as loss or confidence, and do not fully leverage the dynamic behavior of models when actively probed. We propose LeakBoost, a perceptual-loss-based interrogation framework that actively probes a model's internal representations to expose hidden membership signals. Given a candidate input, LeakBoost synthesizes an interrogation image by optimizing a perceptual (activation-space) objective, amplifying representational differences between members and non-members. This image is then analyzed by an off-the-shelf membership detector, without modifying the detector itself. When combined with existing membership inference methods, LeakBoost achieves substantial improvements at low false-positive rates across multiple image classification datasets and diverse neural network architectures. In particular, it raises AUC from near-chance levels (0.53-0.62) to 0.81-0.88, and increases TPR at 1 percent FPR by over an order of magnitude compared to strong baseline attacks. A detailed sensitivity analysis reveals that deeper layers and short, low-learning-rate optimization produce the strongest leakage, and that improvements concentrate in gradient-based detectors. LeakBoost thus offers a modular and computationally efficient way to assess privacy risks in white-box settings, advancing the study of dynamic membership inference.

cs.AI↗

arXiv2025

Memory Backdoor Attacks on Neural Networks

Neural networks are often trained on proprietary datasets, making them attractive attack targets. We present a novel dataset extraction method leveraging an innovative training time backdoor attack, allowing a malicious federated learning server to systematically and deterministically extract complete client training samples through a simple indexing process. Unlike prior techniques, our approach guarantees exact data recovery rather than probabilistic reconstructions or hallucinations, provides precise control over which samples are memorized and how many, and shows high capacity and robustness. Infected models output data samples when they receive a patternbased index trigger, enabling systematic extraction of meaningful patches from each clients local data without disrupting global model utility. To address small model output sizes, we extract patches and then recombined them. The attack requires only a minor modification to the training code that can easily evade detection during client-side verification. Hence, this vulnerability represents a realistic FL supply-chain threat, where a malicious server can distribute modified training code to clients and later recover private data from their updates. Evaluations across classifiers, segmentation models, and large language models demonstrate that thousands of sensitive training samples can be recovered from client models with minimal impact on task performance, and a clients entire dataset can be stolen after multiple FL rounds. For instance, a medical segmentation dataset can be extracted with only a 3 percent utility drop. These findings expose a critical privacy vulnerability in FL systems, emphasizing the need for stronger integrity and transparency in distributed training pipelines.

cs.CR↗

arXiv2025

Trust Me, I Know This Function: Hijacking LLM Static Analysis using Bias

Large Language Models (LLMs) are increasingly trusted to perform automated code review and static analysis at scale, supporting tasks such as vulnerability detection, summarization, and refactoring. In this paper, we identify and exploit a critical vulnerability in LLM-based code analysis: an abstraction bias that causes models to overgeneralize familiar programming patterns and overlook small, meaningful bugs. Adversaries can exploit this blind spot to hijack the control flow of the LLM's interpretation with minimal edits and without affecting actual runtime behavior. We refer to this attack as a Familiar Pattern Attack (FPA). We develop a fully automated, black-box algorithm that discovers and injects FPAs into target code. Our evaluation shows that FPAs are not only effective against basic and reasoning models, but are also transferable across model families (OpenAI, Anthropic, Google), and universal across programming languages (Python, C, Rust, Go). Moreover, FPAs remain effective even when models are explicitly warned about the attack via robust system prompts. Finally, we explore positive, defensive uses of FPAs and discuss their broader implications for the reliability and safety of code-oriented LLMs.

cs.LG↗

arXiv2025

ProxyPrints: From Database Breach to Spoof, A Plug-and-Play Defense for Biometric Systems

Fingerprint recognition systems are widely deployed for authentication and forensic applications, but the security of stored fingerprint data remains a critical vulnerability. While many systems avoid storing raw fingerprint images in favor of minutiae-based templates, recent research shows that these templates can be reverse-engineered to reconstruct realistic fingerprint images, enabling physical spoofing attacks that compromise user identities with no means of remediation. We present ProxyPrints, the first practical defense that brings cancellable biometrics to existing fingerprint recognition systems without requiring modifications to proprietary matching software. ProxyPrints acts as a transparent middleware layer between the fingerprint scanner and the matching algorithm, transforming each scanned fingerprint into a consistent, unlinkable alias. This transformation allows biometric identities to be revoked and replaced in the event of a breach, without affecting authentication accuracy. Additionally, ProxyPrints provides organizations with breach detection capabilities by enabling the identification of out-of-band spoofing attempts involving compromised aliases. We evaluate ProxyPrints on standard benchmark datasets and commercial fingerprint recognition systems, demonstrating that it preserves matching performance while offering strong security and revocability. Our open-source implementation includes tools for alias generation and deployment in real-world pipelines, making ProxyPrints a drop-in, scalable solution for fingerprint data protection.

cs.CR↗

arXiv2025

ASRJam: Human-Friendly AI Speech Jamming to Prevent Automated Phone Scams

Large Language Models (LLMs), combined with Text-to-Speech (TTS) and Automatic Speech Recognition (ASR), are increasingly used to automate voice phishing (vishing) scams. These systems are scalable and convincing, posing a significant security threat. We identify the ASR transcription step as the most vulnerable link in the scam pipeline and introduce ASRJam, a proactive defence framework that injects adversarial perturbations into the victim's audio to disrupt the attacker's ASR. This breaks the scam's feedback loop without affecting human callers, who can still understand the conversation. While prior adversarial audio techniques are often unpleasant and impractical for real-time use, we also propose EchoGuard, a novel jammer that leverages natural distortions, such as reverberation and echo, that are disruptive to ASR but tolerable to humans. To evaluate EchoGuard's effectiveness and usability, we conducted a 39-person user study comparing it with three state-of-the-art attacks. Results show that EchoGuard achieved the highest overall utility, offering the best combination of ASR disruption and human listening experience.

cs.CL↗

arXiv2025

PaniCar: Securing the Perception of Advanced Driving Assistance Systems Against Emergency Vehicle Lighting

The safety of autonomous cars has come under scrutiny in recent years, especially after 16 documented incidents involving Teslas (with autopilot engaged) crashing into parked emergency vehicles (police cars, ambulances, and firetrucks). While previous studies have revealed that strong light sources often introduce flare artifacts in the captured image, which degrade the image quality, the impact of flare on object detection performance remains unclear. In this research, we unveil PaniCar, a digital phenomenon that causes an object detector's confidence score to fluctuate below detection thresholds when exposed to activated emergency vehicle lighting. This vulnerability poses a significant safety risk, and can cause autonomous vehicles to fail to detect objects near emergency vehicles. In addition, this vulnerability could be exploited by adversaries to compromise the security of advanced driving assistance systems (ADASs). We assess seven commercial ADASs (Tesla Model 3, "manufacturer C", HP, Pelsee, AZDOME, Imagebon, Rexing), four object detectors (YOLO, SSD, RetinaNet, Faster R-CNN), and 14 patterns of emergency vehicle lighting to understand the influence of various technical and environmental factors. We also evaluate four SOTA flare removal methods and show that their performance and latency are insufficient for real-time driving constraints. To mitigate this risk, we propose Caracetamol, a robust framework designed to enhance the resilience of object detectors against the effects of activated emergency vehicle lighting. Our evaluation shows that on YOLOv3 and Faster RCNN, Caracetamol improves the models' average confidence of car detection by 0.20, the lower confidence bound by 0.33, and reduces the fluctuation range by 0.33. In addition, Caracetamol is capable of processing frames at a rate of between 30-50 FPS, enabling real-time ADAS car detection.

cs.CV↗

arXiv2024

Are You Human? An Adversarial Benchmark to Expose LLMs

Large Language Models (LLMs) have demonstrated an alarming ability to impersonate humans in conversation, raising concerns about their potential misuse in scams and deception. Humans have a right to know if they are conversing to an LLM. We evaluate text-based prompts designed as challenges to expose LLM imposters in real-time. To this end we compile and release an open-source benchmark dataset that includes 'implicit challenges' that exploit an LLM's instruction-following mechanism to cause role deviation, and 'exlicit challenges' that test an LLM's ability to perform simple tasks typically easy for humans but difficult for LLMs. Our evaluation of 9 leading models from the LMSYS leaderboard revealed that explicit challenges successfully detected LLMs in 78.4% of cases, while implicit challenges were effective in 22.9% of instances. User studies validate the real-world applicability of our methods, with humans outperforming LLMs on explicit challenges (78% vs 22% success rate). Our framework unexpectedly revealed that many study participants were using LLMs to complete tasks, demonstrating its effectiveness in detecting both AI impostors and human misuse of AI tools. This work addresses the critical need for reliable, real-time LLM detection methods in high-stakes conversations.

cs.CL↗

arXiv2024

Back-in-Time Diffusion: Unsupervised Detection of Medical Deepfakes

Recent progress in generative models has made it easier for a wide audience to edit and create image content, raising concerns about the proliferation of deepfakes, especially in healthcare. Despite the availability of numerous techniques for detecting manipulated images captured by conventional cameras, their applicability to medical images is limited. This limitation stems from the distinctive forensic characteristics of medical images, a result of their imaging process. In this work we propose a novel anomaly detector for medical imagery based on diffusion models. Normally, diffusion models are used to generate images. However, we show how a similar process can be used to detect synthetic content by making a model reverse the diffusion on a suspected image. We evaluate our method on the task of detecting fake tumors injected and removed from CT and MRI scans. Our method significantly outperforms other state of the art unsupervised detectors with an increased AUC of 0.9 from 0.79 for injection and of 0.96 from 0.91 for removal on average. We also explore our hypothesis using AI explainability tools and publish our code and new medical deepfake datasets to encourage further research into this domain.

eess.IV↗

arXiv2024

The Best Defense is a Good Offense: Countering LLM-Powered Cyberattacks

As large language models (LLMs) continue to evolve, their potential use in automating cyberattacks becomes increasingly likely. With capabilities such as reconnaissance, exploitation, and command execution, LLMs could soon become integral to autonomous cyber agents, capable of launching highly sophisticated attacks. In this paper, we introduce novel defense strategies that exploit the inherent vulnerabilities of attacking LLMs. By targeting weaknesses such as biases, trust in input, memory limitations, and their tunnel-vision approach to problem-solving, we develop techniques to mislead, delay, or neutralize these autonomous agents. We evaluate our defenses under black-box conditions, starting with single prompt-response scenarios and progressing to real-world tests using custom-built CTF machines. Our results show defense success rates of up to 90\%, demonstrating the effectiveness of turning LLM vulnerabilities into defensive strategies against LLM-driven cyber threats.

cs.CR↗

arXiv2024

PEAS: A Strategy for Crafting Transferable Adversarial Examples

Black box attacks, where adversaries have limited knowledge of the target model, pose a significant threat to machine learning systems. Adversarial examples generated with a substitute model often suffer from limited transferability to the target model. While recent work explores ranking perturbations for improved success rates, these methods see only modest gains. We propose a novel strategy called PEAS that can boost the transferability of existing black box attacks. PEAS leverages the insight that samples which are perceptually equivalent exhibit significant variability in their adversarial transferability. Our approach first generates a set of images from an initial sample via subtle augmentations. We then evaluate the transferability of adversarial perturbations on these images using a set of substitute models. Finally, the most transferable adversarial example is selected and used for the attack. Our experiments show that PEAS can double the performance of existing attacks, achieving a 2.5x improvement in attack success rates on average over current ranking methods. We thoroughly evaluate PEAS on ImageNet and CIFAR-10, analyze hyperparameter impacts, and provide an ablation study to isolate each component's importance.

cs.LG↗