Work

Revocation Window in RAG: Quantifying Unauthorized Retrieval Exposure During Access-Control Propagation

Year: 2026
Type: article
Abstract: Retrieval-Augmented Generation (RAG) systems often store document chunks in a vector index together with a copy of their access-control metadata. When a user's permission is revoked in the authoritative permission store, this indexed copy may remain stale, allowing the vector database to keep returning protected content. This paper measures the duration of this revocation window and separates two outcomes: unauthorized index retrieval, in which the vector database returns a protected chunk, and LLM-context exposure, in which that chunk reaches the context assembled for the language model. We built a controlled synthetic testbed with PostgreSQL as the authoritative permission store and Qdrant as the vector index, and evaluated six enforcement configurations: periodic synchronization, event-driven synchronization, a no-worker control, query-time authorization, hybrid authorization, and query-time authorization with a 10-second decision cache. Each configuration retained 27 trials (5 for the control) within a 35-second observation window. The median context-level revocation propagation time was 9.92 s for 20-second periodic synchronization, 14 ms for event-driven synchronization, and 4.98 s for the cached query-time configuration. Under periodic synchronization, 87.7% of relevant queries placed the protected chunk in the context before revocation was enforced. Query-time and hybrid authorization blocked protected content at the first post-revocation probe (0 ms at probe resolution) even though the stale index kept returning it, while the no-worker control never achieved stable enforcement within the window. These findings describe the tested synthetic testbed and are not production guarantees.
Language: English
Open Access status: green
OpenAlex ID: W7236365743
FWCI:
Cites:
Cited by:
Related to: 10