Pandora’s Toolbox Or Why Agent Safety Cannot Be Verified

Mario Franco, Carlos Gershenson

Present artificial intelligence agents are constructed by means of a coupling between a fixed language model with a general computing device. We show that this coupling makes safety undecidable. In particular, we show that given some dangerous program, no algorithm can predict whether an agent will run such a program that is semantically equivalent. In general, guardrails, blacklists, whitelists and certifications can only lead to imperfect safety solutions. Under finite constraints (bounded context, values and running time) safety is decidable but still intractable. Specifically, the problem is 𝖯𝖲𝖯𝖠𝖢𝖤-complete even when the description of the agent is known. Finally, safety is not a compositional property, i.e., agents that can be proven safe on their own can still produce unsafe behavior when grouped together. We also discuss relaxed notions of safety that could be used in practice.

Read the full article at: arxiv.org

Symmetry as a common origin of hierarchical emergence and abstraction

Fernando E. Rosas Fernando E. Rosas

SCIENCE ADVANCES

13 Nov 2026
Vol 12, Issue 46

Many systems of interest exhibit nested emergent layers with their own rules and regularities, and our knowledge about them seems naturally organized around these levels. How do such hierarchical structures arise, and why should observers track these hierarchies so closely? This paper proposes symmetry as an answer to both questions. By combining principles from information theory, group theory, and statistical mechanics, it is shown that dynamical processes that are equivariant with respect to a symmetry group give rise to multiple emergent macroscopic levels arranged according to the algebraic structure of the symmetry. The same symmetries shape the Bayesian beliefs of an observer with only partial access to the system, yielding a matching hierarchy of “abstract” beliefs that can be updated autonomously at different levels of resolution. These ideas are illustrated in Hopfield networks, where familiar macroscopic quantities emerge from known symmetries, and in Boolean models of gene regulation, where symmetries are instead found from the dynamics. The computational advantages provided by symmetry are showcased in Ehrenfest diffusion and epidemic modeling, where abstract beliefs enable Bayesian inference over exponentially large state spaces. These results identify symmetry as a sufficient mechanism for hierarchical emergence and—at the same time—for the capacity of abstract beliefs to mirror the systems they track, making feasible inferential problems that would otherwise be intractable.Many systems of interest exhibit nested emergent layers with their own rules and regularities, and our knowledge about them seems naturally organized around these levels. How do such hierarchical structures arise, and why should observers track these hierarchies so closely? This paper proposes symmetry as an answer to both questions. By combining principles from information theory, group theory, and statistical mechanics, it is shown that dynamical processes that are equivariant with respect to a symmetry group give rise to multiple emergent macroscopic levels arranged according to the algebraic structure of the symmetry. The same symmetries shape the Bayesian beliefs of an observer with only partial access to the system, yielding a matching hierarchy of “abstract” beliefs that can be updated autonomously at different levels of resolution. These ideas are illustrated in Hopfield networks, where familiar macroscopic quantities emerge from known symmetries, and in Boolean models of gene regulation, where symmetries are instead found from the dynamics. The computational advantages provided by symmetry are showcased in Ehrenfest diffusion and epidemic modeling, where abstract beliefs enable Bayesian inference over exponentially large state spaces. These results identify symmetry as a sufficient mechanism for hierarchical emergence and—at the same time—for the capacity of abstract beliefs to mirror the systems they track, making feasible inferential problems that would otherwise be intractable.

Read the full article at: www.science.org

Origins:  Cosmos, Life, Mind, by Stuart A. Kauffman

Do the laws of physics convincingly explain how the universe came to be? Is the evolution of the biosphere measurable? Does the human mind operate outside of physical space and time? Origins: Cosmos, Life, Mind asks human nature’s most fundamental questions. Each of these domains are in turmoil.

We are in an age of precision cosmology but do not understand the basic physics of dark matter, dark energy, inflation, and singularities in black holes where spacetime vanishes. Origins proposes a coherent body of theory and new data that may answer these vast, unsettled inquiries.

Stuart Kauffman discusses the dominant and emerging alternative views of how life on earth came to arise; he proposes a new definition of life in terms of which its emergence in the universe is expected. The book weaves a new, coherent, conceptual framework that allows us to think about our most fundamental questions in novel and testable ways for physicists, biologists, and philosophers. Remarkably, whilst Newton and physics teach us that the world is governed by the laws of physics, it seems that evolution cannot be foretold. Evolution is an ever-creative emergence. The world is one of creative emergence. How can we come to flourish as responsible participants in the extractive Anthropocene?

More at: global.oup.com

Collective noise filtering in complex networks

Tingyu Zhao and István A KovácsTingyu Zhao and István A Kovács

Journal of Physics: Complexity, Volume 7, Number 4

Real-world network data often suffer from noise, biases, and missing data in edge weights, all of which can undermine the reliability of downstream network analyses. Standard noise filtering approaches, whether treating individual edges one-by-one or assuming a uniform global noise level, are suboptimal because, in practice, both signal and noise can be heterogeneous and correlated across multiple edges. To address this challenge, we introduce the Network Wiener Filter (NetWF), a principled method for collective edge-level noise filtering that jointly leverages network structure and explicit noise characteristics. It is capable of reducing noise in observed edge weights as well as inferring missing edge weights. It applies broadly across network types and scales efficiently to large empirical systems, as we demonstrate using real-world network data from two distinct settings. We envision the NetWF as a new foundation for error-aware network analysis.Real-world network data often suffer from noise, biases, and missing data in edge weights, all of which can undermine the reliability of downstream network analyses. Standard noise filtering approaches, whether treating individual edges one-by-one or assuming a uniform global noise level, are suboptimal because, in practice, both signal and noise can be heterogeneous and correlated across multiple edges. To address this challenge, we introduce the Network Wiener Filter (NetWF), a principled method for collective edge-level noise filtering that jointly leverages network structure and explicit noise characteristics. It is capable of reducing noise in observed edge weights as well as inferring missing edge weights. It applies broadly across network types and scales efficiently to large empirical systems, as we demonstrate using real-world network data from two distinct settings. We envision the NetWF as a new foundation for error-aware network analysis.

Read the full article at: iopscience.iop.org

Swarm AI: Meet the Next Generation of Artificial Intelligence, by George Rzevski

The book describes, using clear and concise language, the next generation of Artificial Intelligence (AI), which consists of swarms of hundreds of thousands of intelligent digital agents that exhibit collective intelligence greater than the sum of the intelligence of individual agents. These agents have access to domain knowledge stored in the attached ontology and do not need training to operate. Swarms of agents self-organise to adapt to environmental changes, resist hacking, and coevolve with the systems they interact with.

The author explains the basics of Swarm AI and its uses, including adaptive schedulers, pattern recognisers, semantic processors, variable geometry engineering products, and teams of space robots. They show how swarms of agents can handle operational tasks in businesses and protect nations from pandemics, financial crises, and military attacks.

The book is written for curious readers who want to understand the simple principles behind the power of the next generation of AI. It is also aimed at business executives, administrators, politicians, government leaders, regulators and policy makers, offering them insights into the new tools that will transform the organisations they lead and the world they inhabit.

More at: link.springer.com