arXiv subjects

Guy Shtar

Publications and source records attributed to Guy Shtar.

5 recordsLinked to original sources

arXiv2026

SALLIE: Generation-Free Hidden-State Detection of Jailbreaks and Prompt Injections Across Text and Vision

Large Language Models (LLMs) and Vision-Language Models (VLMs) are vulnerable to jailbreaks and prompt injections delivered through text or images. Existing defenses often narrow threat coverage or add inference cost through input transformations, repeated model queries, generated responses, or threat-specific detectors. We present SALLIE (Single-Pass Activation Lookup for Layerwise Input Evaluation), a white-box, generation-free detector with a common architecture for textual and visual jailbreaks and prompt injections. Using one backbone forward pass, SALLIE extracts last-token residual-stream representations, applies layer-wise k-nearest-neighbor probes, and averages their scores across a contiguous layer range, without modifying the backbone or generating a response. To distinguish reusable detector configurations from deployment-specific tuning, we compare shared, threshold-only, and fully calibrated regimes across backbones and modalities. We evaluate SALLIE on compact open-weight VLMs (Gemma-3-4b-it, Phi-3.5-vision-instruct, SmolVLM2-2.2B-Instruct) using heterogeneous held-out test data spanning both threat families and modalities. Fully calibrated SALLIE-Phi achieves an aggregate F1 of 0.87 and higher aggregate balanced accuracy than RCS-KCD on all three backbones under a matched-calibration protocol. On Phi-3.5 visual inputs, threshold-only calibration yields an F1 of 0.99 with zero observed false positives, versus 0.54 for zero-shot Gemini-2.5-Flash-Lite. On Gemma-4-E2B-it, a backbone held out from representation-hyperparameter selection, the same procedure raises F1 from 0.62 to 0.79 on text and 0.68 to 0.98 on visual inputs. Representation hyperparameters can thus transfer across some backbones, while operating points remain backbone-, modality-, and distribution-dependent. We further report FPR/FNR tradeoffs, latency, and representation ablations.

cs.CR↗

arXiv2025

ASTRA: Agentic Steerability and Risk Assessment Framework

Securing AI agents powered by Large Language Models (LLMs) represents one of the most critical challenges in AI security today. Unlike traditional software, AI agents leverage LLMs as their "brain" to autonomously perform actions via connected tools. This capability introduces significant risks that go far beyond those of harmful text presented in a chatbot that was the main application of LLMs. A compromised AI agent can deliberately abuse powerful tools to perform malicious actions, in many cases irreversible, and limited solely by the guardrails on the tools themselves and the LLM ability to enforce them. This paper presents ASTRA, a first-of-its-kind framework designed to evaluate the effectiveness of LLMs in supporting the creation of secure agents that enforce custom guardrails defined at the system-prompt level (e.g., "Do not send an email out of the company domain," or "Never extend the robotic arm in more than 2 meters"). Our holistic framework simulates 10 diverse autonomous agents varying between a coding assistant and a delivery drone equipped with 37 unique tools. We test these agents against a suite of novel attacks developed specifically for agentic threats, inspired by the OWASP Top 10 but adapted to challenge the ability of the LLM for policy enforcement during multi-turn planning and execution of strict tool activation. By evaluating 13 open-source, tool-calling LLMs, we uncovered surprising and significant differences in their ability to remain secure and keep operating within their boundaries. The purpose of this work is to provide the community with a robust and unified methodology to build and validate better LLMs, ultimately pushing for more secure and reliable agentic AI systems.

cs.CR↗

arXiv2023

AMFPMC -- An improved method of detecting multiple types of drug-drug interactions using only known drug-drug interactions

Adverse drug interactions are largely preventable causes of medical accidents, which frequently result in physician and emergency room encounters. The detection of drug interactions in a lab, prior to a drug's use in medical practice, is essential, however it is costly and time-consuming. Machine learning techniques can provide an efficient and accurate means of predicting possible drug-drug interactions and combat the growing problem of adverse drug interactions. Most existing models for predicting interactions rely on the chemical properties of drugs. While such models can be accurate, the required properties are not always available.

cs.LG↗

arXiv2019

Detecting drug-drug interactions using artificial neural networks and classic graph similarity measures

Drug-drug interactions are preventable causes of medical injuries and often result in doctor and emergency room visits. Computational techniques can be used to predict potential drug-drug interactions. We approach the drug-drug interaction prediction problem as a link prediction problem and present two novel methods for drug-drug interaction prediction based on artificial neural networks and factor propagation over graph nodes: adjacency matrix factorization (AMF) and adjacency matrix factorization with propagation (AMFP). We conduct a retrospective analysis by training our models on a previous release of the DrugBank database with 1,141 drugs and 45,296 drug-drug interactions and evaluate the results on a later version of DrugBank with 1,440 drugs and 248,146 drug-drug interactions. Additionally, we perform a holdout analysis using DrugBank. We report an area under the receiver operating characteristic curve score of 0.807 and 0.990 for the retrospective and holdout analyses respectively. Finally, we create an ensemble-based classifier using AMF, AMFP, and existing link prediction methods and obtain an area under the receiver operating characteristic curve of 0.814 and 0.991 for the retrospective and the holdout analyses. We demonstrate that AMF and AMFP provide state of the art results compared to existing methods and that the ensemble-based classifier improves the performance by combining various predictors. These results suggest that AMF, AMFP, and the proposed ensemble-based classifier can provide important information during drug development and regarding drug prescription given only partial or noisy data. These methods can also be used to solve other link prediction problems. Drug embeddings (compressed representations) created when training our models using the interaction network have been made public.

cs.LG↗

arXiv2019

Clustering Wi-Fi Fingerprints for Indoor-Outdoor Detection

This paper presents a method for continuous indoor-outdoor environment detection on mobile devices based solely on WiFi fingerprints. Detection of indoor outdoor switching is an important part of identifying a user's context, and it provides important information for upper layer context aware mobile applications such as recommender systems, navigation tools, etc. Moreover, future indoor positioning systems are likely to use Wi-Fi fingerprints, and therefore Wi-Fi receivers will be on most of the time. In contrast to existing research, we believe that these fingerprints should be leveraged, and they serve as the basis of the proposed method. Using various machine learning algorithms, we train a supervised classifier based on features extracted from the raw fingerprints, clusters, and cluster transition graph. The contribution of each of the features to the method is assessed. Our method assumes no prior knowledge of the environment, and a training set consisting of the data collected for just a few hours on a single device is sufficient in order to provide indoor-outdoor classification, even in an unknown location or when using new devices. We evaluate our method in an experiment involving 12 participants during their daily routine, with a total of 828 hours' worth of data collected by the participants. We report a predictive performance of the AUC (area under the curve) of 0.94 using the gradient boosting machine ensemble learning method. We show that our method can be used for other context detection tasks such as learning and recognizing a given building or room.

eess.SP↗