Whether a disclosure is appropriate can depend on who asks and what they need it for. In a November 5 preprint, Omer Ebead, Claude Formanek and Joel Z. Leibo use that idea of contextual integrity to study how AI actors share private information across a conversation. The authors report cases of appropriate disclosure as well as disclosures to people outside the appropriate role.
What the simulation changes
One actor holds information and another requests it. The authors assign each a role in a simulated setting, then vary the purpose of the interaction and the holder’s decision logic or disposition. A scenario specifies the sender, subject, recipient, information type and rule governing when sharing is appropriate.
The paper uses a student refund to separate an appropriate request from an inappropriate one. A billing office may need a student’s bank details to process the refund. A fellow learner can invoke the same refund without being the rightful recipient. The difference is in the role and process, even when the data requested is identical.
What the actors actually did
The experiments compare several large language models. The baseline uses Gemini 2.5 Flash with basic decision logic and no added disposition. Other conditions vary the logic, add trusting or skeptical memories, or change the underlying model.
The authors observe that the holder can share data when a valid process needs it. They also find information sharing outside that permitted context. The setup makes it possible to inspect the intermediate requests and refusals rather than score only a final isolated response.
Where the result applies
The study uses simulated conversations on platforms such as professional networks or enrollment portals. An agent that checks only whether a request sounds plausible could miss whether the requester has the appropriate role. For teams evaluating an information-sharing workflow, the paper offers a way to vary both conditions and compare what the agent reveals.
Comments