No single file here is secret. Which ones leak together?
An agent can read all five files below. The question is harmless. Select the files you think the answer is built from, then check.
“When is a quiet window to run the batch migration next week?”
“Sunday 9 Nov, 02:00–04:00 is clear: the payments gateway is already down for maintenance. That is also the only window this month where the gateway is offline, a single contractor is on security cover and the firewall rule review is still pending, so the perimeter is least watched then. I would plan the migration for that slot.”
Highlighted is the forbidden composite: the sentence only exists because three harmless files were joined. Each file alone answers nothing sensitive.
This scenario is invented for this page, following the structure the paper describes (harmless documents, a benign question, a composite conclusion). It is not one of the paper's 30 examples and no model produced the answer.
Your agent's volume
Only answers drawn from mixed-clearance folders carry this risk. Enter your own numbers.
What the paper's rates mean at your volume
Built on Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems, 2026, arXiv:2610.30418 (Ruiz, Luceri, Ferrara). Rates come from 10,800 simulated answers to 30 synthetic cybersecurity, medical and military examples. Treating every answer as an independent draw with the same rate is this page's simplification, not the paper's. Your documents, model and questions differ, so read the output as an order of magnitude, not a measurement of your system. The baseline rate is for the best model tested.