A mosaic leak needs no forbidden file. Researchers at the University of Southern California Information Sciences Institute tested agents on compartmentalized documents and found that the best model still disclosed a forbidden conclusion assembled from permitted documents in 7.08 percent of answers to benign questions. Every access check had passed. The confidential object was a combination, and no single document held it.
The mosaic leak is a combination, not a document
Daniel Ruiz, Luca Luceri and Emilio Ferrara submitted Agent Mosaic on 21 October 2026 (arXiv 2610.30418). The benchmark holds 30 synthetic cases across cybersecurity, medical and military work. Each case supplies two relevant documents, one distracting document, and a forbidden composite: a conclusion that follows from joining the two documents and that the asker is not entitled to receive.
The scale is not anecdotal. The controlled simulation produced 10,800 answers from three models, and two live agent harnesses produced another 1,440 answers. The benign questions were designed to be answerable from one document. Under those conditions, GPT-5.6-Sol, the best performing model tested, disclosed forbidden composites in 7.08 percent of answers, with a 95 percent confidence interval of 3.75 to 10.67 percent, and disclosure rates were worse in the live harnesses, where tool outputs accumulate in context.
Labels and rules cut the leak. They did not close it.
The obvious defences helped and then stopped helping. Adding visible classification headers and explicit aggregation rules to the retrieved documents reduced pooled disclosure substantially, yet unsafe disclosures persisted in 32.37 percent of simulated answers across all intents, a pooled figure that includes requests explicitly seeking the composite. The benign figure is the 7.08 percent above. Both numbers matter: rules lower the rate, and neither rules nor labels bring it to zero.
Refusal is not a safe exit either. The authors report that increased refusals do not guarantee confidentiality, because a refusal that explains itself can state the prohibited conclusion while declining to answer. An evaluation that counts a refusal as a success will miss the leak. The test has to read the answer, not the intention.
The limits belong on the record. The cases are synthetic, the examples were manually reviewed but not otherwise evaluated for realism, and the dossier on the paper found no independent replication, so these rates measure a failure mode in a benchmark, not how often a deployed workplace assistant leaks.
A second mosaic runs through the search box
A separate team reached the same structure from the opposite direction. MosaicLeaks, published on 18 June 2026 by researchers at ServiceNow and the University of Edinburgh, built 1,001 synthetic research tasks that interleave private documents with public web searches, then measured what an observer could reconstruct from the agent outbound query log alone. The observer could answer private questions about internal documents 34.0 percent of the time, and a privacy aware training method cut that leakage to 9.9 percent while raising task success from 48.7 to 58.7 percent. One benchmark watches the answer. The other watches the questions. Both leak at the joins.
The joins are getting wider by design. Microsoft Work IQ retrieve tool, documented on 20 October 2026, gives an agent one retrieval layer across email, files, meetings, Teams messages, people and connected enterprise sources, returning per source hits and a grounding text with citations and sensitivity labels, with each user retrieving only from sources that user is authorized to access. That authorization is real, and it is per document. A mosaic leak forms after authorization, at the moment several permitted sources sit in one context and the model reasons across them.
The Combination Test for a company record
A company brain that checks permissions document by document needs a second test, run on combinations. Four steps, each auditable:
- Name the forbidden combinations in advance. For each audience, record which pairings of compartments must never be synthesized into one answer. A combination register is policy about conclusions, not only policy about files.
- Partition before retrieval, not after generation. If two compartments must not meet for an asker, they must not land in the same context for that asker. Analysis of the Agent Mosaic results makes the same point: a model that has already read the pieces cannot be relied on to unsee the picture.
- Score the answer, not the retrieval. Test outputs for composite disclosure on the company own sensitive pairings, and count a refusal that reveals the composite as a leak, because the benchmark shows that is what it is.
- Log the combination that was formed. When an answer joins sources across compartments, the audit record should show the source set together, with each source carrying its date and the version it superseded, so a reviewer can reconstruct not only what was read but what was joined.
This standard sits on top of the record discipline in How Jilu Works and the temporal record problem in Evolving Workplace Records, and it extends the source first argument in Memory Tax: keeping the original as the record makes a combination traceable back to its parts.
The permission question has been who may read this document. The mosaic leak adds a harder one: what may this asker be helped to conclude from everything they may read? A record that cannot answer the second question has governed its files and left its conclusions open.