Definition
Plain language
A benchmark that measures how often a research agent accidentally spills private facts through its public search queries.
As stated in the literature
A benchmark of chained multi-hop enterprise tasks where private answers must be carried into web queries, paired with a query-trail adversary scored at intent, answer, and full-information leakage levels.
Why it matters: It measures a subtle privacy risk that lives in an agent's search history, helping reveal leaks that single-step checks would miss.
For example, it checks whether an agent answering an internal question types a confidential figure into a public search box where an eavesdropper could read it.
Heard on the show
“It's called "MosaicLeaks: Privacy Risks in Querying-in-the-Open for Deep Research Agents," and the script you're listening to was written by Anthropic's Claude Opus 4.”Episode 104 — How Making a Research Agent Smarter Quietly Makes It Leak Your Secrets