Ana-Maria Creţu

Luca Beurer-Kellner, Beat Buesser, Ana-Maria Creţu et al. · eth-zurich

As AI agents powered by Large Language Models (LLMs) become increasingly versatile and capable of addressing a broad spectrum of tasks, ensuring their security has become a critical challenge. Among the most pressing threats are prompt injection attacks, which exploit the agent's resilience on natural language inputs -- an especially dangerous threat when agents are granted tool access or handle sensitive information. In this work, we propose a set of principled design patterns for building AI agents with provable resistance to prompt injection. We systematically analyze these patterns, discuss their trade-offs in terms of utility and security, and illustrate their real-world applicability through a series of case studies.

Klim Kireev, Ana-Maria Creţu, Raphael Meier et al.

Platforms and the law regulate digital content depicting minors (defined as individuals under 18 years of age) differently from other types of content. Given the sheer amount of content that needs to be assessed, machine learning-based automation tools are commonly used to detect content depicting minors. To our knowledge, no dataset or benchmark currently exists for detecting these identification methods in a multi-modal environment. To fill this gap, we release the Image-Caption Children in the Wild Dataset (ICCWD), an image-caption dataset aimed at benchmarking tools that detect depictions of minors. Our dataset is richer than previous child image datasets, containing images of children in a variety of contexts, including fictional depictions and partially visible bodies. ICCWD contains 10,000 image-caption pairs manually labeled to indicate the presence or absence of a child in the image. To demonstrate the possible utility of our dataset, we use it to benchmark three different detectors, including a commercial age estimation system applied to images. Our results suggest that child detection is a challenging task, with the best method achieving a 75.3% true positive rate. We hope the release of our dataset will aid in the design of better minor detection methods in a wide range of scenarios.

Ana-Maria Creţu, Florent Guépin, Yves-Alexandre de Montjoye

Despite machine learning models being widely used today, the relationship between a model and its training dataset is not well understood. We explore correlation inference attacks, whether and when a model leaks information about the correlations between the input variables of its training dataset. We first propose a model-less attack, where an adversary exploits the spherical parametrization of correlation matrices alone to make an informed guess. Second, we propose a model-based attack, where an adversary exploits black-box model access to infer the correlations using minimal and realistic assumptions. Third, we evaluate our attacks against logistic regression and multilayer perceptron models on three tabular datasets and show the models to leak correlations. We finally show how extracted correlations can be used as building blocks for attribute inference attacks and enable weaker adversaries. Our results raise fundamental questions on what a model does and should remember from its training set.