●builderYou cannot rely on current open-source detectors to secure AI agents against tool-output injections without high false-positive rates.
●policyCurrent safety benchmarks may overstate the efficacy of prompt injection defenses in production agent environments.