Definition
Plain language
A training recipe that teaches a research agent to keep searching effectively while never typing private facts into its queries.
As stated in the literature
A method (PA-DR) that adds a learned leakage-penalty reward — taking the worse of direct and marginal mosaic leakage — alongside targeted situational task rewards, improving accuracy while cutting query-trail leakage.
Also called: PA-DR
Why it matters: It lets research agents stay both effective and private, addressing the risk that capable searching otherwise leaks sensitive facts.
For example, it trains an agent to find the answer to an internal question through web searches without ever typing the confidential answer into the search box.
Heard on the show
“They call it Privacy-Aware Deep Research.”Episode 104 — How Making a Research Agent Smarter Quietly Makes It Leak Your Secrets