A new arXiv paper, DICE: Tilting the Odds through Interface Attacks on System One Model, highlights a vulnerability in how System One models are embedded in larger applications. These models offer typed judgments that applications can directly convert into decisions, but the paper notes that their answers are conditional on the state and question supplied by surrounding components, as well as on how the application uses the output.
The authors propose DICE, an attack method that targets this interface rather than the model itself. By manipulating the components that feed into the model or interpret its output, DICE can tilt the odds of the final judgment. The abstract does not provide specific attack details or results, but the framing suggests that the reliability of System One models depends heavily on the integrity of their surrounding pipeline.
This work underscores a broader concern: as AI models become decision-making components in larger systems, the interfaces between model and application become attack surfaces. The paper's contribution is to name and formalize this class of interface attacks, inviting further research on defending the full decision pipeline, not just the model weights.