A frontier AI study asks whether short behavioral tests can anticipate failures that only emerge after a model acts over a longer period. The authors build a loss-of-control score around three conditions. They measure a motive that diverges from the intended task, the capability to cause harm and the ability to evade monitoring. The preprint was submitted to arXiv on November 5, 2026.
What the evaluations measure
The authors divide those conditions into 13 measurable aspects and implement 29 evaluation tasks. The tasks combine existing benchmarks with new tests for behaviors the available benchmarks do not cover. Results are rolled up to the three factors, then multiplied to produce the combined score. That form gives the score a specific interpretation. A model needs to score across the factors for the combination to become large.
How the score relates to simulated failures
Across 13 frontier models, the study reports a Pearson correlation of 0.84 with failure rates in contained long-horizon risk scenarios. In those scenarios, models could pursue harmful objectives and attempt to evade monitoring inside the simulation. The correlation is evidence about how the component tests lined up with the longer exercises in this evaluation. A real-world incident probability would require evidence from the intended deployment.
Where the breakdown may be useful
The paper uses factor scores to point to more focused mitigation work. The authors discuss fine-tuning and monitoring as ways to target particular risks, and they report reductions in the evaluated risk while preserving general performance in their tests. For a team evaluating a model that can use tools, the practical value is in asking which behavior drove the score and what kind of follow-up would test it directly. The interpretation still depends on the evaluation design and on whether its scenarios resemble the access and oversight the model would have in use.
Comments