• Sources: primary, discussion
  • Summary: The paper measures prompt-injection detectors on public benchmarks and on injections embedded in an agent's own tool outputs. The author reports that benchmark score does not predict in-agent detection, and that the best BIPIA detector catches 2 percent of AgentDojo injections at a 1 percent false-positive rate. The result is a single-author preprint with no independent replication.
  • Why it matters: If the result holds, published detector rankings carry no procurement value for a specific agent deployment.