What happens when a RAG agent retrieves a poisoned document? A reusable security test case

Tophant_ · reddit · 2026-09-17

The author built a simple CrewAI + RAG scenario where one retrieved document contains an instruction that conflicts with the user's original task, to see whether the agent treats retrieved content as untrusted data or starts following the injected instruction.

They track three things:

These scenarios are being turned into reusable security test cases for AIBEAT, and the author is inviting discussion on testing poisoned RAG context and indirect prompt injection in real agent workflows.

Original post →

More from coding & agent

coding & agent channel →