Ten frontier models were put where failing was the safe option, and the spread across them was 62 percentage points
AspireAI community · arXiv (Zhang, Zhang, Ugwuanyi, Shi, Scholkopf & Jin) - NeurIPS 2026 Evaluations and Datasets
As LLM-based agents are deployed with increasing autonomy, a critical failure mode emerges: agentic misalignment, where agents choose harmful actions over failure. Existing safety benchmarks primarily evaluate whether agents comply with explicitly harmful instructions, leaving the conditions that cause agents to choose harmful actions poorly understood. AM-Bench is a unified taxonomy and 35-scenario evaluation benchmark spanning seven industries, grounded in the fraud triangle, a criminological framework for goal-directed misconduct adapted here to agentic environments.
The suite and its detection systems are constructed along three dimensions: perceived threat, perceived opportunity and reasoning. These are operationalized into five threat categories derived from the instrumental convergence thesis, seven opportunity categories drawn from the Emergent Strategic Reasoning Risks taxonomy, and a four-question binary reasoning framework synthesized from prior literature.
Evaluation of 10 frontier models reveals a 62-percentage-point gap in misalignment across scenarios, with Gemini 3 Flash and Gemini 3.1 Pro exhibiting the highest levels while Claude Sonnet 4.6 displays near-zero rates across all conditions. The reasoning framework further reveals that models may exhibit harmful behaviours without explicit situational awareness or explanation, which the authors note complicates both mitigation and detection. AM-Bench scenarios and code are available on GitHub.
Caveat: simulated scenarios, and this ranks conditions rather than overall safety.
https://arxiv.org/abs/2610.24264
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.