As AI systems move into areas that have real social consequences, trust stops being something you can attribute to an algorithm in isolation. It starts to look more like an outcome of how the whole system behaves over time, including how feedback loops form, how quickly responses arrive, and how institutions react when things go wrong. In this work, we propose a simulation-based approach to study how trust evolves in AI-enabled socio-technical systems. We use Markovian Agent Models (MAMs) to represent users, recommender algorithms, and oversight bodies as interacting populations that change continuously over time. We apply this framework to a misinformation scenario in online recommender systems, with particular attention to situations where engagement-driven optimization allows harmful content to spread before corrective actions take effect. Trust is modeled explicitly as an evaluative index derived from the system’s harmful and corrective states, not as an abstract notion. The simulations suggest that trust erosion is driven mainly by slow or delayed institutional responses, more than by the behavior of individual algorithms on their own. Experiments that remove or delay parts of the governance machinery show that moderation and policy enforcement cannot replace one another: disabling either leads to long-lasting trust loss, and even moderate delays fail to change the system’s trajectory once harmful dynamics have taken hold. The broader point is that trust is shaped by how feedback and timing interact at the system level, and that formal simulation can make these dependencies visible when testing different governance choices under clearly stated assumptions.

Autonomous Guided Vehicles are one of the hottest research and industrial topics, with applications spanning from alternative transport, security and logistic. However, safety and security is usually one of the main concerns in this type of technology, limiting their application and introduction. In this paper we will consider the effects that a malicious user can cause on the system, and we propose an analysis technique based on a mean-field model, to be able to quickly obtain estimates of the consequences of an attack, and the impact of the counter-measures.

Rapid advances in the capabilities of large language models (LLMs) have raised widespread concerns regarding their potential for malicious use. Open-weight LLMs present unique challenges, as existing safeguards lack robustness to tampering attacks that modify model weights. For example, recent works have demonstrated that refusal and unlearning safeguards can be trivially removed with a few steps of fine-tuning. These vulnerabilities necessitate new approaches for enabling the safe release of open-weight LLMs. We develop a method, called TAR, for building tamper-resistant safeguards into open-weight LLMs such that adversaries cannot remove the safeguards even after hundreds of steps of fine-tuning. In extensive evaluations and red teaming analyses, we find that our method greatly improves tamper-resistance while preserving benign capabilities. Our results demonstrate that progress on tamper-resistance is possible, opening up a promising new avenue to improve the safety and security of open-weight LLMs.

The White House Executive Order on Artificial Intelligence highlights the risks of large language models (LLMs) empowering malicious actors in developing biological, cyber, and chemical weapons. To measure these risks, government institutions and major AI labs are developing evaluations for hazardous capabilities in LLMs. However, current evaluations are private and restricted to a narrow range of malicious use scenarios, which limits further research into reducing malicious use. To fill these gaps, we release the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of 3,668 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. To guide progress on unlearning, we develop RMU, a state-of-the-art unlearning method based on controlling model representations. RMU reduces model performance on WMDP while maintaining general capabilities in areas such as biology and computer science, suggesting that unlearning may be a concrete path towards reducing malicious use from LLMs. We release our benchmark and code publicly at https://wmdp.ai.

Performance on popular ML benchmarks is highly correlated with model scale, suggesting that most benchmarks tend to measure a similar underlying factor of general model capabilities. However, substantial research effort remains devoted to designing new benchmarks, many of which claim to measure novel phenomena. In the spirit of the Bitter Lesson, we leverage spectral analysis to measure an underlying capabilities component, the direction in benchmark-performance-space which explains most variation in model performance. In an extensive analysis of existing safety benchmarks, we find that variance in model performance on many safety benchmarks is largely explained by the capabilities component. In response, we argue that safety research should prioritize metrics which are not highly correlated with scale. Our work provides a lens to analyze both novel safety benchmarks and novel safety methods, which we hope will enable future work to make differential progress on safety.

We present a novel method for graph partitioning, based on reinforcement learning and graph convolutional neural networks. Our approach is to recursively partition coarser representations of a given graph. The neural network is implemented using SAGE graph convolution layers, and trained using an advantage actor critic (A2C) agent. We present two variants, one for finding ean edge separator that minimizes the normalized cut or quotient cut, and one that finds a small vertex separator. The vertex separators are then used to construct a nested dissection ordering to permute a sparse matrix so that its triangular factorization will incur less fill-in. The partitioning quality is compared with partitions obtained using METIS and SCOTCH, and the nested dissection ordering is evaluated in the sparse solver SuperLU. Our results show that the proposed method achieves similar partitioning quality as METIS, SCOTCH and spectral partitioning. Furthermore, the method generalizes across different classes of graphs, and works well on a variety of graphs from the SuiteSparse sparse matrix collection.