The problem
Access reviews ask whether any single document is sensitive. That test was built for people who read a handful of files. An agent reads dozens in seconds and writes a fluent summary. If three ordinary files together reveal a weak window in your defences, a headcount plan or a pending deal, the agent can say so in one sentence. Nobody sees that conclusion until it has already been written into a chat. The person who scopes the agent's folders has no number to take to a risk meeting, only a feeling that it is probably fine.
The research
Ruiz, Luceri and Ferrara's "Agent Mosaic" (arXiv:2610.30418, October 2026) names this the mosaic effect. Using 30 synthetic cybersecurity, medical and military examples, they collected 10,800 simulated answers from three agents. The best model disclosed a forbidden composite in 7.08% of answers to benign questions (95% CI 3.75–10.67%). With classification headers and explicit aggregation rules, 32.37% of answers across all intents were still unsafe (CI 28.33–36.19%). This build runs no model. The scenario is simulated, and the calculator applies the paper's published rates.
What it is
A three-step page. First you try to spot which of five harmless files combine into a leak, then see the simulated answer with the forbidden sentence marked. Next you enter your own questions per day, the share that touch mixed-clearance folders and working days. Finally you get expected leaked answers a month and the chance of at least one, each with the paper's confidence range, plus a per-100 dot grid. Inputs are remembered in the browser and the last result is shown on your next visit. You can download a text summary.
Who it's for
For whoever signs off agent permissions at a company of 50 to 5,000 people and has to justify a folder list. Also for security reviewers who want a rough figure to put beside a policy. It is not for people who want an audit of their own system: the page tests nothing about your documents or model. It is not a leak detector, and it does not suggest the safeguards fix anything. The paper shows they do not.
How it makes money
It does not, and that is correct for now. The calculator rests on one preprint's simulated rates, and charging for it would invite a claim of precision the evidence cannot carry. If the paper's results replicate on other models, the real product is a scanner that finds which folder combinations can produce a composite. That would be sold per seat to security teams. This page is the argument for building it, not the product.
Go to market
Post it where agent-deployment questions are asked: security engineering newsletters, the platform-team and CISO communities already discussing agents as insider threats this month, and the paper's own discussion threads. Its shareable moment is the dot grid beside a team's own volume number. A reviewer who pastes "about 300 expected leaks a month" into a risk ticket is sending the next person a reason to open the page and try their figures. The demo does not prove detection, only the arithmetic.