Skip to content
IAM Insights

Explainers · Glossary

Aggregation inference

aggregation inference · mosaic effect

Aggregation inference: the risk that combining individually authorized data accesses reveals information no single access would expose, so the combined result may itself be unauthorized; known from statistical disclosure control and classified-data handling.

reference →

In our briefings

What our briefings say about Aggregation inference itself, with the most direct claims first.

  1. Tallam singles out aggregation inference as the part that checking each access does not cover: every read in a workflow can be authorized while the result synthesized from them exposes something no single read would, the "mosaic effect" of classified information handling. She calls it "genuinely unsolved in the general case" and says agents give it new dimensions, because the synthesis is a neural network whose behavior is not formally specified and the resources it draws on may not be known in advance. Architecture, she says, can still make the contributing resources inspectable after the fact, limit which resource combinations are permitted, and track causal dependencies through the synthesis graph. Her requirement R4 asks for aggregation policies that can be expressed and are enforced before synthesized results are returned. The attack figure she cites, a 71% success rate across 14 enterprise scenarios, comes from another work her reference list credits to "Anonymous".

    An agent workflow checked only at its start may read data restricted mid-runBI-FIND-2026-2471

Cited in 2 briefings

Every briefing that discusses this term, newest first.

  1. An agent workflow checked only at its start may read data restricted mid-run

    Paper · · arXiv · Krti Tallam, Kamiwaza AI · May 6, 2026

  2. IPC passes a delegation chain between services; its signature proves only the last signer

    Draft · · IETF Datatracker · Ravi Kant Sharma, Ericsson · Jul 23, 2026