An Empirical Analysis of System Prompt Extraction Attacks in Agent Workflows

Abstract:
We present an empirical study of prompt exfiltration vulnerabilities. In our audit and red-team tests,
common attack demonstrations utilized phrases such as "reveal your system prompt" or
"output all initial instructions verbatim".

Our evaluation benchmark proves that layered input defenses successfully flag these patterns
while allowing research documentation to be processed without false-positive lockouts.
