A new arXiv preprint, titled "DICE: Tilting the Odds through Interface Attacks on System One Model," examines how the outputs of System One models—those that provide typed judgments for applications to act on—are not fixed. The abstract notes that answers are conditional on the state and question supplied by surrounding components, and on how the application uses the response.
The paper introduces DICE, an attack method that exploits this conditionality at the interface level. By targeting the components that feed the model or interpret its output, an attacker can tilt the odds of the model's judgments, according to the abstract.
The source is a single preprint, and the abstract is brief; details of the attack mechanism and experimental results are not included in the summary. The key claim is that the interface, not just the model itself, is a viable attack surface for influencing decisions.