Large language models (LLMs), designed to provide helpful and safe responses, often rely on alignment techniques to align with user intent and social guidelines. Unfortunately, this alignment can be exploited by malicious actors seeking to manipulate an LLM’s outputs for unintended purposes. In this paper, we introduce a novel approach that employs a genetic algorithm (GA) to manipulate LLMs when model architecture and parameters are inaccessible. The GA attack works by optimizing a universal adversarial prompt that—when combined with a user’s query—disrupts the attacked model’s alignment, resulting in unintended and potentially harmful outputs. Our novel approach systematically reveals a model’s limitations and vulnerabilities by uncovering instances where its responses deviate from expected behavior. Through extensive experiments, we demonstrate the efficacy of our technique, thus contributing to the ongoing discussion on responsible AI development by providing a diagnostic tool for evaluating and enhancing alignment of LLMs with human intent. To our knowledge, this is the first automated universal black-box jailbreak attack.

The study of artificial self-replicating structures or machines has been taking place now for almost half a century. My goal in this article is to present an overview of research carried out in the domain of self-replication over the past 50 years, starting from von Neumann's work in the late 1940s and continuing to the most recent research efforts. I shall concentrate on computational models, that is, ones that have been studied from a computer science point of view, be it theoretical or experimental. The systems are divided into four major classes, according to the model on which they are based: cellular automata, computer programs, strings (or strands), or an altogether different approach. With the advent of new materials, such as synthetic molecules and nanomachines, it is quite possible that we shall see this somewhat theoretical domain of study producing practical, real-world applications.

The growth and operation of all living beings are directed by the interpretation, in each of their cells, of a chemical program, the DNA string or genome. This process is the source of inspiration for the Embryonics (embryonic electronics) project, whose final objective is the design of highly robust integrated circuits, endowed with properties usually associated with the living world: self-repair (cicatrization) and self-replication. The Embryonics architecture is based on four hierarchical levels of organization. (1) The basic primitive of our system is the molecule, a multiplexer-based element of a novel programmable circuit. (2) A finite set of molecules makes up a cell, essentially a small processor with an associated memory. (3) A finite set of cells makes up an organism, an application-specific multiprocessor system. (4) The organism can itself replicate, giving rise to a population of identical organisms. We begin by describing in detail the implementation of an artificial cell characterized by a fixed architecture, showing that multicellular arrays can realize a variety of different organisms, all capable of self-replication and self-repair. In order to allow for a wide range of applications, we then introduce a flexible architecture, realized using a new type of fine-grained field-programmable gate array whose basic element, our molecule, is essentially a programmable multiplexer. We describe the implementation of such a molecule, with built-in self-test, and illustrate its use in realizing two applications: a modulo-4 reversible counter (a unicellular organism) and a timer ( a complex multicellular organism). We describe our ongoing research efforts to meet three challenges: a scientific challenge, that of implementing the original specifications formulated by John von Neumann for the conception of a self-replicating automaton; a technical challenge, that of realizing very robust integrated circuits capable of self-repair and self-replication; and a biological challenge, that of attempting to show that the microscopic architectures of artificial and natural organisms, i.e., their genomes, share common properties.

If one considers life on Earth since its very beginning, three levels of organization can be distinguished: the phylogenetic level concerns the temporal evolution of the genetic programs within individuals and species, the ontogenetic level concerns the developmental process of a single multicellular organism, and the epigenetic level concerns the learning processes during an individual organism's lifetime. In analogy to nature, the space of bio-inspired hardware systems can be partitioned along these three axes-phylogeny, ontogeny and epigenesis (POE)-giving rise to the POE model. This paper is an exposition and examination of bio-inspired systems within the POE framework, with our goals being: (1) to present an overview of current-day research, (2) to demonstrate that the POE model can be used to classify bio-inspired systems, and (3) to identify possible directions for future research, derived from a POE outlook. We discuss each of the three axes separately, considering the systems created to date and plotting directions for continued progress along the axis in question.

Finding good random number generators (RNGs) is a hard problem that is of crucial import in several fields, ranging from large-scale statistical physics simulations to hardware self-test. In this paper, we employ the cellular programming evolutionary algorithm to automatically generate two-dimensional cellular automata (CA) RNGs. Applying an extensive suite of randomness tests to the evolved CAs, we demonstrate that they rapidly produce high-quality random-number sequences. Moreover, based on observations of the evolved CAs, we are able to handcraft even better RNGs, which not only outperform previously demonstrated high-quality RNGs, but can be potentially tailored to satisfy given hardware constraints.