A new arXiv preprint (2610.15537) examines context poisoning in typed decision workflows, where models are limited to a predefined set of actions. The paper highlights a subtle gap: even when a model picks an action that fits the interface's schema, that choice can still lead to state changes that violate policy.
The authors argue that safety evaluations for such systems cannot stop at checking whether an output is schema-valid. Instead, they say, evaluating context poisoning requires connecting model decisions to execution outcomes—what actually happens when the action is taken in the broader workflow.
This points to a potential blind spot in current testing approaches. If researchers only validate that a model's output conforms to the allowed action types, they may miss attacks that exploit the downstream effects of those actions.